Merge pull request #103 from modelscope/v1.4.1_dev

update v1.4.1

.github/workflows/publish.yml

@@ -0,0 +1,24 @@
+name: Publish to Comfy registry
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - main
+      - master
+    paths:
+      - "scepter/workflow/pyproject.toml"
+
+jobs:
+  publish-node:
+    name: Publish Custom Node to registry
+    runs-on: ubuntu-latest
+    # if this is a forked repository. Skipping the workflow.
+    if: github.event.repository.fork == false
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v4
+      - name: Publish Custom Node
+        uses: Comfy-Org/publish-node-action@main
+        with:
+          ## Add your own personal access token to your Github Repository secrets and reference it here.
+          personal_access_token: ${{ secrets.REGISTRY_ACCESS_TOKEN }}

.gitignore

@@ -9,12 +9,12 @@
 *.bin
 *.idea
 *.csv
+cache
 build
 dist
 dev
 scepter.egg-info
 .readthedocs.yml
-1.9
-#MANIFEST.in
 *resources
 *.ipynb_checkpoints*
+*.vscode

asset/workflow/sdxl_base.json

@@ -0,0 +1,163 @@
+{
+  "last_node_id": 4,
+  "last_link_id": 3,
+  "nodes": [
+    {
+      "id": 2,
+      "type": "PreviewImage",
+      "pos": {
+        "0": 937,
+        "1": 243
+      },
+      "size": {
+        "0": 284.03680419921875,
+        "1": 246
+      },
+      "flags": {
+        "collapsed": false
+      },
+      "order": 2,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 1
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "Node name for S&R": "PreviewImage"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 4,
+      "type": "ParameterNode",
+      "pos": {
+        "0": 0,
+        "1": 244
+      },
+      "size": {
+        "0": 311.9849853515625,
+        "1": 236.4765625
+      },
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "Result",
+          "type": "CONDITIONING",
+          "links": [
+            3
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "ParameterNode"
+      },
+      "widgets_values": [
+        "ddim",
+        50,
+        5,
+        0.5,
+        "trailing",
+        1024,
+        1024,
+        2024
+      ]
+    },
+    {
+      "id": 1,
+      "type": "ModelNode",
+      "pos": {
+        "0": 430,
+        "1": 243
+      },
+      "size": {
+        "0": 400,
+        "1": 234
+      },
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "parameters",
+          "type": "CONDITIONING",
+          "link": 3,
+          "shape": 7
+        },
+        {
+          "name": "mantras",
+          "type": "CONDITIONING",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "tuners",
+          "type": "CONDITIONING",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "controls",
+          "type": "CONDITIONING",
+          "link": null,
+          "shape": 7
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            1
+          ],
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "ModelNode"
+      },
+      "widgets_values": [
+        "SD_XL1.0",
+        "ModelScope",
+        "a cat",
+        ""
+      ]
+    }
+  ],
+  "links": [
+    [
+      1,
+      1,
+      0,
+      2,
+      0,
+      "IMAGE"
+    ],
+    [
+      3,
+      4,
+      0,
+      1,
+      0,
+      "CONDITIONING"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "ds": {
+      "scale": 0.9229599817706423,
+      "offset": [
+        128.91466506592707,
+        43.07074577975898
+      ]
+    }
+  },
+  "version": 0.4
+}

asset/workflow/sdxl_base_mantra.json

@@ -0,0 +1,202 @@
+{
+  "last_node_id": 5,
+  "last_link_id": 4,
+  "nodes": [
+    {
+      "id": 2,
+      "type": "PreviewImage",
+      "pos": {
+        "0": 937,
+        "1": 243
+      },
+      "size": {
+        "0": 284.03680419921875,
+        "1": 246
+      },
+      "flags": {
+        "collapsed": false
+      },
+      "order": 3,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 1
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "Node name for S&R": "PreviewImage"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 4,
+      "type": "ParameterNode",
+      "pos": {
+        "0": 0,
+        "1": 244
+      },
+      "size": {
+        "0": 311.9849853515625,
+        "1": 236.4765625
+      },
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "Result",
+          "type": "CONDITIONING",
+          "links": [
+            3
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "ParameterNode"
+      },
+      "widgets_values": [
+        "ddim",
+        50,
+        5,
+        0.5,
+        "trailing",
+        1024,
+        1024,
+        2024
+      ]
+    },
+    {
+      "id": 1,
+      "type": "ModelNode",
+      "pos": {
+        "0": 430,
+        "1": 243
+      },
+      "size": {
+        "0": 400,
+        "1": 234
+      },
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "parameters",
+          "type": "CONDITIONING",
+          "link": 3,
+          "shape": 7
+        },
+        {
+          "name": "mantras",
+          "type": "CONDITIONING",
+          "link": 4,
+          "shape": 7
+        },
+        {
+          "name": "tuners",
+          "type": "CONDITIONING",
+          "link": null,
+          "shape": 7
+        },
+        {
+          "name": "controls",
+          "type": "CONDITIONING",
+          "link": null,
+          "shape": 7
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            1
+          ],
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "ModelNode"
+      },
+      "widgets_values": [
+        "SD_XL1.0",
+        "ModelScope",
+        "a cat",
+        ""
+      ]
+    },
+    {
+      "id": 5,
+      "type": "MantrasNode",
+      "pos": {
+        "0": 3,
+        "1": 547
+      },
+      "size": {
+        "0": 315,
+        "1": 58
+      },
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "Result",
+          "type": "CONDITIONING",
+          "links": [
+            4
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "MantrasNode"
+      },
+      "widgets_values": [
+        "Flat 2D Art"
+      ]
+    }
+  ],
+  "links": [
+    [
+      1,
+      1,
+      0,
+      2,
+      0,
+      "IMAGE"
+    ],
+    [
+      3,
+      4,
+      0,
+      1,
+      0,
+      "CONDITIONING"
+    ],
+    [
+      4,
+      5,
+      0,
+      1,
+      1,
+      "CONDITIONING"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "ds": {
+      "scale": 1.0152559799477068,
+      "offset": [
+        58.35814123566077,
+        -65.12749162217047
+      ]
+    }
+  },
+  "version": 0.4
+}

asset/workflow/sdxl_base_mantra_tuner.json

@@ -0,0 +1,242 @@
+{
+  "last_node_id": 6,
+  "last_link_id": 5,
+  "nodes": [
+    {
+      "id": 2,
+      "type": "PreviewImage",
+      "pos": {
+        "0": 937,
+        "1": 243
+      },
+      "size": {
+        "0": 284.03680419921875,
+        "1": 246
+      },
+      "flags": {
+        "collapsed": false
+      },
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 1
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "Node name for S&R": "PreviewImage"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 4,
+      "type": "ParameterNode",
+      "pos": {
+        "0": 0,
+        "1": 244
+      },
+      "size": {
+        "0": 311.9849853515625,
+        "1": 236.4765625
+      },
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "Result",
+          "type": "CONDITIONING",
+          "links": [
+            3
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "ParameterNode"
+      },
+      "widgets_values": [
+        "ddim",
+        50,
+        5,
+        0.5,
+        "trailing",
+        1024,
+        1024,
+        2024
+      ]
+    },
+    {
+      "id": 5,
+      "type": "MantrasNode",
+      "pos": {
+        "0": 3,
+        "1": 547
+      },
+      "size": {
+        "0": 315,
+        "1": 58
+      },
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "Result",
+          "type": "CONDITIONING",
+          "links": [
+            4
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "MantrasNode"
+      },
+      "widgets_values": [
+        "Flat 2D Art"
+      ]
+    },
+    {
+      "id": 1,
+      "type": "ModelNode",
+      "pos": {
+        "0": 430,
+        "1": 243
+      },
+      "size": {
+        "0": 400,
+        "1": 234
+      },
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "parameters",
+          "type": "CONDITIONING",
+          "link": 3,
+          "shape": 7
+        },
+        {
+          "name": "mantras",
+          "type": "CONDITIONING",
+          "link": 4,
+          "shape": 7
+        },
+        {
+          "name": "tuners",
+          "type": "CONDITIONING",
+          "link": 5,
+          "shape": 7
+        },
+        {
+          "name": "controls",
+          "type": "CONDITIONING",
+          "link": null,
+          "shape": 7
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            1
+          ],
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "ModelNode"
+      },
+      "widgets_values": [
+        "SD_XL1.0",
+        "ModelScope",
+        "a cat",
+        ""
+      ]
+    },
+    {
+      "id": 6,
+      "type": "TunerNode",
+      "pos": {
+        "0": 8,
+        "1": 680
+      },
+      "size": {
+        "0": 315,
+        "1": 82
+      },
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "Result",
+          "type": "CONDITIONING",
+          "links": [
+            5
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "TunerNode"
+      },
+      "widgets_values": [
+        "SD_XL1.0_Pencil Sketch Drawing",
+        1
+      ]
+    }
+  ],
+  "links": [
+    [
+      1,
+      1,
+      0,
+      2,
+      0,
+      "IMAGE"
+    ],
+    [
+      3,
+      4,
+      0,
+      1,
+      0,
+      "CONDITIONING"
+    ],
+    [
+      4,
+      5,
+      0,
+      1,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      5,
+      6,
+      0,
+      1,
+      2,
+      "CONDITIONING"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "ds": {
+      "scale": 0.6934334949441366,
+      "offset": [
+        450.0923948318967,
+        30.783653239467935
+      ]
+    }
+  },
+  "version": 0.4
+}

asset/workflow/sdxl_base_mantra_tuner_control.json

@@ -0,0 +1,366 @@
+{
+  "last_node_id": 11,
+  "last_link_id": 9,
+  "nodes": [
+    {
+      "id": 1,
+      "type": "ModelNode",
+      "pos": {
+        "0": 439,
+        "1": 123
+      },
+      "size": {
+        "0": 400,
+        "1": 234
+      },
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "parameters",
+          "type": "CONDITIONING",
+          "link": 3,
+          "shape": 7
+        },
+        {
+          "name": "mantras",
+          "type": "CONDITIONING",
+          "link": 4,
+          "shape": 7
+        },
+        {
+          "name": "tuners",
+          "type": "CONDITIONING",
+          "link": 5,
+          "shape": 7
+        },
+        {
+          "name": "controls",
+          "type": "CONDITIONING",
+          "link": 6,
+          "shape": 7
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            1
+          ],
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "ModelNode"
+      },
+      "widgets_values": [
+        "SD_XL1.0",
+        "ModelScope",
+        "a cat",
+        ""
+      ]
+    },
+    {
+      "id": 5,
+      "type": "MantrasNode",
+      "pos": {
+        "0": 13,
+        "1": 336
+      },
+      "size": {
+        "0": 315,
+        "1": 58
+      },
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "Result",
+          "type": "CONDITIONING",
+          "links": [
+            4
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "MantrasNode"
+      },
+      "widgets_values": [
+        "Flat 2D Art"
+      ]
+    },
+    {
+      "id": 6,
+      "type": "TunerNode",
+      "pos": {
+        "0": 12,
+        "1": 455
+      },
+      "size": {
+        "0": 315,
+        "1": 82
+      },
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "Result",
+          "type": "CONDITIONING",
+          "links": [
+            5
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "TunerNode"
+      },
+      "widgets_values": [
+        "SD_XL1.0_Pencil Sketch Drawing",
+        1
+      ]
+    },
+    {
+      "id": 11,
+      "type": "NoteNode",
+      "pos": {
+        "0": 845,
+        "1": 506
+      },
+      "size": {
+        "0": 398.3059387207031,
+        "1": 210.83267211914062
+      },
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {
+        "Node name for S&R": "NoteNode"
+      },
+      "widgets_values": [
+        "This is a sample for quickly setting up ComfyUI using the Scepter open-source library:\n\n1) The first run involves downloading the models. By default, we will automatically pull models from ModelScope. The initial download may take some time, and you can also adjust the model source to change the model address.\n\n2) Currently, it supports various base models, basic settings for some inference hyperparameters, mantra settings, tuning model settings, and conditional generation node.\n\n3) In the future, we will gradually integrate interesting features to enrich the use cases.\n\n"
+      ]
+    },
+    {
+      "id": 8,
+      "type": "LoadImage",
+      "pos": {
+        "0": 471,
+        "1": 451
+      },
+      "size": {
+        "0": 315,
+        "1": 314
+      },
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            7
+          ]
+        },
+        {
+          "name": "MASK",
+          "type": "MASK",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LoadImage"
+      },
+      "widgets_values": [
+        "cat.jpg",
+        "image"
+      ]
+    },
+    {
+      "id": 7,
+      "type": "ControlNode",
+      "pos": {
+        "0": 15,
+        "1": 600
+      },
+      "size": {
+        "0": 330,
+        "1": 198
+      },
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source_image",
+          "type": "IMAGE",
+          "link": 7
+        }
+      ],
+      "outputs": [
+        {
+          "name": "Result",
+          "type": "CONDITIONING",
+          "links": [
+            6
+          ]
+        },
+        {
+          "name": "Control Image",
+          "type": "IMAGE",
+          "links": [],
+          "slot_index": 1
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "ControlNode"
+      },
+      "widgets_values": [
+        "SD_XL1.0_color",
+        "Color",
+        "CenterCrop",
+        1,
+        1024,
+        1024
+      ]
+    },
+    {
+      "id": 2,
+      "type": "PreviewImage",
+      "pos": {
+        "0": 959,
+        "1": 121
+      },
+      "size": {
+        "0": 284.03680419921875,
+        "1": 246
+      },
+      "flags": {
+        "collapsed": false
+      },
+      "order": 7,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 1
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "Node name for S&R": "PreviewImage"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 4,
+      "type": "ParameterNode",
+      "pos": {
+        "0": 13,
+        "1": 43
+      },
+      "size": {
+        "0": 311.9849853515625,
+        "1": 236.4765625
+      },
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "Result",
+          "type": "CONDITIONING",
+          "links": [
+            3
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "ParameterNode"
+      },
+      "widgets_values": [
+        "ddim",
+        50,
+        5,
+        0.5,
+        "trailing",
+        1024,
+        1024,
+        2024
+      ]
+    }
+  ],
+  "links": [
+    [
+      1,
+      1,
+      0,
+      2,
+      0,
+      "IMAGE"
+    ],
+    [
+      3,
+      4,
+      0,
+      1,
+      0,
+      "CONDITIONING"
+    ],
+    [
+      4,
+      5,
+      0,
+      1,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      5,
+      6,
+      0,
+      1,
+      2,
+      "CONDITIONING"
+    ],
+    [
+      6,
+      7,
+      0,
+      1,
+      3,
+      "CONDITIONING"
+    ],
+    [
+      7,
+      8,
+      0,
+      7,
+      0,
+      "IMAGE"
+    ]
+  ],
+  "groups": [],
+  "config": {},
+  "extra": {
+    "ds": {
+      "scale": 0.7627768444385667,
+      "offset": [
+        254.62860058494985,
+        87.05734194144193
+      ]
+    }
+  },
+  "version": 0.4
+}

docs/en/tasks/ace.md

@@ -0,0 +1,184 @@
+<p align="center">
+
+  <h2 align="center"><img src="https://raw.githubusercontent.com/ali-vilab/ACE/refs/heads/main/assets/figures/icon.png" height=16> : All-round Creator and Editor Following <br> Instructions via Diffusion Transformer</h2>
+
+  <p align="center">
+    <a href="https://arxiv.org/abs/2410.00086"><img src='https://img.shields.io/badge/arXiv-ACE-red' alt='Paper PDF'></a>
+    <a href='https://ali-vilab.github.io/ace-page'><img src='https://img.shields.io/badge/Project_Page-ACE-blue' alt='Project Page'></a>
+    <a href='https://github.com/modelscope/scepter'><img src='https://img.shields.io/badge/Scepter-ACE-green'></a>
+    <a href='https://huggingface.co/spaces/scepter-studio/ACE-Chat'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Space-orange'></a>
+    <a href='https://huggingface.co/scepter-studio/ACE-0.6B-512px'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model-orange'></a>
+    <a href='https://www.modelscope.cn/models/iic/ACE-0.6B-512px'><img src='https://img.shields.io/badge/ModelScope-Model-purple'></a>
+    <br>
+    <strong>Zhen Han*</strong>
+    ·
+    <strong>Zeyinzi Jiang*</strong>
+    ·
+    <strong>Yulin Pan*</strong>
+    ·
+    <strong>Jingfeng Zhang*</strong>
+    ·
+    <strong>Chaojie Mao*</strong>
+    <br>
+    <strong>Chenwei Xie</strong>
+    ·
+    <strong>Yu Liu</strong>
+    ·
+    <strong>Jingren Zhou</strong>
+    <br>
+    Tongyi Lab, Alibaba Group
+  </p>
+  <table align="center">
+    <tr>
+    <td>
+      <img src="https://raw.githubusercontent.com/ali-vilab/ACE/refs/heads/main/assets/figures/teaser.png">
+    </td>
+    </tr>
+  </table>
+
+
+## 🚀 Installation
+Install the necessary packages with `pip`: 
+```bash
+pip install -r requirements.txt
+```
+
+##  🔥 ACE Models
+|    **Model**     |                                                                                                                                                                                                            **Status**                                                                                                                                                                                                             | 
+|:----------------:|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------:|
+|  ACE-0.6B-512px  |          [![Demo link](https://img.shields.io/badge/Demo-ACE_Chat-purple)](https://huggingface.co/spaces/scepter-studio/ACE-Chat)<br>[![ModelScope link](https://img.shields.io/badge/ModelScope-Model-blue)](https://www.modelscope.cn/models/iic/ACE-0.6B-512px)  [![HuggingFace link](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model-yellow)](https://huggingface.co/scepter-studio/ACE-0.6B-512px)          |
+| ACE-0.6B-1024px  | [![Demo link](https://img.shields.io/badge/Demo-ACE_Refiner_Chat-purple)](https://huggingface.co/spaces/scepter-studio/ACE-Refiner-Chat)<br>[![ModelScope link](https://img.shields.io/badge/ModelScope-Model-blue)](https://www.modelscope.cn/models/iic/ACE-0.6B-1024px)  [![HuggingFace link](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model-yellow)](https://huggingface.co/scepter-studio/ACE-0.6B-1024px) |             |
+## 🖼 Model Performance Visualization
+
+The current model's parameters scale of ACE is 0.6B, which imposes certain limitations on the quality of image generation. [FLUX.1-Dev](https://huggingface.co/black-forest-labs/FLUX.1-dev), on the other hand, 
+has a significant advantage in text-to-image generation quality. By using SDEdit, we can effectively leverage the generative capabilities of FLUX to further enhance the image results generated by ACE. Based on the above considerations, we have designed the ACE-Refiner pipeline, as shown in the diagram below.
+
+![ACE_REFINER](https://raw.githubusercontent.com/ali-vilab/ACE/refs/heads/main/assets/ace_method/ace_refiner_process.webp)
+
+As shown in the figure below, when the strength 
+σ of the generated image is high, the generated image will suffer from fidelity loss compared to the original image. Conversely, lower 
+σ does not significantly improve the image quality. Therefore, users can make a trade-off between fidelity to the generated result and the image quality based on their own needs. 
+Users can set the value of "REFINER_SCALE" in the configuration file `config/inference_config/models/ace_0.6b_1024_refiner.yaml`. 
+We recommend that users use the advance options in the [webui-demo](#-chat-bot-) for effect verification.
+
+![ACE_REFINER_EXAMPLE](https://raw.githubusercontent.com/ali-vilab/ACE/refs/heads/main/assets/ace_method/ace_refiner.webp)
+
+
+We compared the generation and editing performance of different models on several tasks, as shown as following. 
+![Samples](https://raw.githubusercontent.com/ali-vilab/ACE/refs/heads/main/assets/ace_method/samples_compare.webp)
+
+
+## 🔥 Training
+
+We offer a demonstration training YAML that enables the end-to-end training of ACE using a toy dataset. For a comprehensive overview of the hyperparameter configurations, please consult `config/ace_0.6b_512_train.yaml`.
+
+### Prepare datasets
+
+Please find the dataset class located in `modules/data/dataset/dataset.py`, 
+designed to facilitate end-to-end training using an open-source toy dataset. 
+Download a dataset zip file from [modelscope](https://www.modelscope.cn/models/iic/scepter/resolve/master/datasets/hed_pair.zip), and then extract its contents into the `cache/datasets/` directory.
+
+Should you wish to prepare your own datasets, we recommend consulting `modules/data/dataset/dataset.py` for detailed guidance on the required data format.
+
+### Prepare initial weight
+The ACE checkpoint has been uploaded to both ModelScope and HuggingFace platforms:
+* [ModelScope](https://www.modelscope.cn/models/iic/ACE-0.6B-512px)
+* [HuggingFace](https://huggingface.co/scepter-studio/ACE-0.6B-512px)
+
+In the provided training YAML configuration, we have designated the Modelscope URL as the default checkpoint URL. Should you wish to transition to Hugging Face, you can effortlessly achieve this by modifying the PRETRAINED_MODEL value within the YAML file (replace the prefix "ms://iic" to "hf://scepter-studio").
+
+
+### Start training
+
+You can easily start training procedure by executing the following command:
+```bash
+# ACE-0.6B-512px
+PYTHONPATH=. python tools/run_train.py --cfg config/ace_0.6b_512_train.yaml
+# ACE-0.6B-1024px
+PYTHONPATH=. python tools/run_train.py --cfg config/ace_0.6b_1024_train.yaml
+```
+
+## 🚀 Inference
+
+We provide a simple inference demo that allows users to generate images from text descriptions.
+```bash
+ PYTHONPATH=. python tools/run_inference.py --cfg config/inference_config/models/ace_0.6b_512.yaml --instruction "make the boy cry, his eyes filled with tears" --seed 199999 --input_image examples/input_images/example0.webp
+```
+We recommend runing the examples for quick testing. Running the following command will run the example inference and the results will be saved in `examples/output_images/`.
+```bash
+PYTHONPATH=. python tools/run_inference.py --cfg config/inference_config/models/ace_0.6b_512.yaml
+```
+
+## 💬 Chat Bot 
+We have developed an chatbot UI utilizing Gradio, designed to transform user input in natural language into visually stunning images that align semantically with the provided instructions. Users can effortlessly initiate the chatbot app by executing the following command:
+```bash
+python chatbot/run_gradio.py --cfg chatbot/config/chatbot_ui.yaml --server_port 2024
+```
+
+<table align="center">
+  <tr>
+  <td>
+    <img src="https://raw.githubusercontent.com/ali-vilab/ACE/refs/heads/main/assets/videos/demo_chat.gif">
+  </td>
+  </tr>
+</table>
+
+## ⚙️️ ComfyUI Workflow
+
+![Workflow](https://raw.githubusercontent.com/ali-vilab/ACE/refs/heads/main/assets/comfyui/ace_example.jpg)
+
+We support the use of ACE in the ComfyUI Workflow through the following methods:
+
+1) Automatic installation directly via the ComfyUI Manager by searching for the **ComfyUI-Scepter** node.
+2) Manually install by moving custom_nodes from Scepter to ComfyUI.
+```shell
+git clone https://github.com/modelscope/scepter.git
+cd path/to/scepter
+pip install -e .
+cp -r path/to/scepter/workflow/ path/to/ComfyUI/custom_nodes/ComfyUI-Scepter
+cd path/to/ComfyUI
+python main.py
+```
+
+**Note**: You can use the nodes by dragging the sample images below into ComfyUI. Additionally, our nodes can automatically pull models from ModelScope or HuggingFace by selecting the *model_source* field, or you can place the already downloaded models in a local path.
+
+<table><tbody>
+  <tr>
+    <th align="center" colspan="4">ACE Workflow Examples</th>
+  </tr>
+  <tr>
+    <th align="center" colspan="1">Control</th>
+    <th align="center" colspan="1">Semantic</th>
+    <th align="center" colspan="1">Element</th>
+  </tr>
+  <tr>
+    <td>
+      <a href="https://raw.githubusercontent.com/ali-vilab/ACE/refs/heads/main/assets/comfyui/ace_control.png" target="_blank">
+        <img src="https://raw.githubusercontent.com/ali-vilab/ACE/refs/heads/main/assets/comfyui/ace_control.png" width="200">
+      </a>
+    </td>
+    <td>
+      <a href="https://raw.githubusercontent.com/ali-vilab/ACE/refs/heads/main/assets/comfyui/ace_semantic.png" target="_blank">
+        <img src="https://raw.githubusercontent.com/ali-vilab/ACE/refs/heads/main/assets/comfyui/ace_semantic.png" width="200">
+      </a>
+    </td>
+    <td>
+      <a href="https://raw.githubusercontent.com/ali-vilab/ACE/refs/heads/main/assets/comfyui/ace_element.png" target="_blank">
+        <img src="https://raw.githubusercontent.com/ali-vilab/ACE/refs/heads/main/assets/comfyui/ace_element.png" width="200">
+      </a>
+    </td>
+  </tr>
+</tbody>
+</table>
+
+
+## 📝 Citation
+
+```bibtex
+@article{han2024ace,
+  title={ACE: All-round Creator and Editor Following Instructions via Diffusion Transformer},
+  author={Han, Zhen and Jiang, Zeyinzi and Pan, Yulin and Zhang, Jingfeng and Mao, Chaojie and Xie, Chenwei and Liu, Yu and Zhou, Jingren},
+  journal={arXiv preprint arXiv:2410.00086},
+  year={2024}
+}
+```

environment.yaml

@@ -1,10 +0,0 @@
-name: scepter
-channels:
-  - defaults
-dependencies:
-  - python==3.8
-  - pip>=20.3
-  - numpy>=1.23.1
-  - pip:
-      - -r requirements/recommended.txt
-      - -r requirements.txt

install.py

@@ -0,0 +1,22 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import os
+import shutil
+import subprocess
+import sys
+
+if sys.argv[0] == 'install.py':
+    sys.path.append('.')   # for portable version
+
+source_folder = os.path.join(os.path.dirname(__file__), "scepter/workflow")
+current_dir = os.path.dirname(__file__)
+destination_folder = os.path.join(os.path.dirname(current_dir), "ComfyUI-Scepter")
+
+if not os.path.exists(destination_folder):
+    shutil.copytree(source_folder, destination_folder)
+    print(f"{os.path.abspath(source_folder)} copy to {os.path.abspath(destination_folder)} success!")
+else:
+    print(f"{os.path.abspath(destination_folder)} exist.")
+
+# pip install scepter
+subprocess.check_call([sys.executable, '-m', 'pip', 'install', 'scepter'])

readme.md

@@ -14,11 +14,15 @@ SCEPTER integrates popular community-driven implementations as well as proprieta
 SCEPTER offers 3 core components:
 - [Generative training and inference framework](#tutorials)
 - [Easy implementation of popular approaches](#currently-supported-approaches)
-- [Interactive user interface: SCEPTER Studio](#launch)
+- [Interactive user interface: SCEPTER Studio & Comfy UI](#launch)
 
 
 ## 🎉 News
-- [2024.09]: We introduce **ACE**, an **A**ll-round **C**reator and **E**ditor adept at executing a diverse array of image editing tasks tailored to your specifications. Built upon the cutting-edge Diffusion Transformer architecture, ACE has been extensively trained on a comprehensive dataset to seamlessly interpret and execute any natural language instruction. For further information, please consult the [project page]().
+- [🔥🔥🔥 2025.01]: We report ACE++, an instruction-based diffusion framework that tackles various image generation and editing tasks. The code and paper is available on [ACE++](https://ali-vilab.github.io/ACE_plus_page/).
+- [2024.11]: Supports video files, video annotation, caption translation in data management, and inference & training of the [CogVideoX](https://arxiv.org/abs/2408.06072).
+- [2024.10]: We are pleased to announce the release of the code for [ACE](https://arxiv.org/abs/2410.00086), supporting Customized Training / Comfy UI Workflow / gradio-based ChatBot Interface. 
+- [2024.10]: Support for inference and tuning with [FLUX](https://huggingface.co/black-forest-labs/FLUX.1-dev), as well as for building [ComfyUI](https://github.com/comfyanonymous/ComfyUI) workflows using this framework.
+- [2024.09]: We introduce **ACE**, an **A**ll-round **C**reator and **E**ditor adept at executing a diverse array of image editing tasks tailored to your specifications. Built upon the cutting-edge Diffusion Transformer architecture, ACE has been extensively trained on a comprehensive dataset to seamlessly interpret and execute any natural language instruction. For further information, please consult the [project page](https://ali-vilab.github.io/ace-page/).
 - [2024.07]: Support the inference and training of open-source generative models based on the [DiT](https://arxiv.org/abs/2212.09748) architecture, such as [SD3](https://arxiv.org/pdf/2403.03206) and [PixArt](https://arxiv.org/abs/2310.00426).
 - [2024.05]: Introducing SCEPTER v1, supporting customized image edit tasks! Simply provide 10 image pairs, SCEPTER will tune an edit tuner for your own Image-to-Image tasks, like `Clay Style`, `De-Text`, `Segmentation`, etc.
 - [2024.04]: New [StyleBooth](https://ali-vilab.github.io/stylebooth-page/) demo on SCEPTER Studio for`Text-Based Style Editing`.
@@ -29,29 +33,108 @@ SCEPTER offers 3 core components:
 - [2023.12]: We propose [SCEdit](https://arxiv.org/abs/2312.11392), an efficient and controllable generation framework.
 - [2023.12]: We release [🪄SCEPTER](https://github.com/modelscope/scepter/) library.
 
+[//]: # (## 🖼 Gallery for Recent Works)
 
-## 🖼 Gallery for Recent Works
+[//]: # ()
+[//]: # (### FLUX Tuners)
 
-### <img src="asset/images/ace/logo.png" height=20> <img src="asset/images/ace/text.png" height=20>
+[//]: # ()
+[//]: # (<table><tbody>)
 
-ACE is a unified foundational model framework that supports a wide range of visual generation tasks. By defining CU for unifying multi-modal inputs across different tasks and incorporating long-
-context CU, we introduce historical contextual information into visual generation tasks, paving
-the way for ChatGPT-like dialog systems in visual generation.
+[//]: # (  <tr>)
 
-<a href="https://ali-vilab.github.io/ace-page/">
-  <img src="asset/images/ace/teaser_dy.gif" width=1024>
-</a>
+[//]: # (    <th align="center" colspan="3">Yarn Style</th>)
+
+[//]: # (    <th align="center" colspan="3">Soft Watercolor Style</th>)
+
+[//]: # (  </tr>)
+
+[//]: # (  <tr>)
+
+[//]: # (    <td><img src="asset/images/flux_tuner/flux_tuner_2_1.webp" width="200"></td>)
+
+[//]: # (    <td><img src="asset/images/flux_tuner/flux_tuner_2_2.webp" width="200"></td>)
+
+[//]: # (    <td><img src="asset/images/flux_tuner/flux_tuner_2_3.webp" width="200"></td>)
+
+[//]: # (    <td><img src="asset/images/flux_tuner/flux_tuner_1_1.webp" width="200"></td>)
+
+[//]: # (    <td><img src="asset/images/flux_tuner/flux_tuner_1_2.webp" width="200"></td>)
+
+[//]: # (    <td><img src="asset/images/flux_tuner/flux_tuner_1_3.webp" width="200"></td>)
+
+[//]: # (  </tr>)
+
+[//]: # (  <tr>)
+
+[//]: # (    <th align="center" colspan="3">Travel Style</th>)
+
+[//]: # (    <th align="center" colspan="3">WuKong Style</th>)
+
+[//]: # (  </tr>)
+
+[//]: # (  <tr>)
+
+[//]: # (    <td><img src="asset/images/flux_tuner/flux_tuner_3_1.webp" width="200"></td>)
+
+[//]: # (    <td><img src="asset/images/flux_tuner/flux_tuner_3_2.webp" width="200"></td>)
+
+[//]: # (    <td><img src="asset/images/flux_tuner/flux_tuner_3_3.webp" width="200"></td>)
+
+[//]: # (    <td><img src="asset/images/flux_tuner/flux_tuner_4_1.webp" width="200"></td>)
+
+[//]: # (    <td><img src="asset/images/flux_tuner/flux_tuner_4_2.webp" width="200"></td>)
+
+[//]: # (    <td><img src="asset/images/flux_tuner/flux_tuner_4_3.webp" width="200"></td>)
+
+[//]: # (  </tr>)
+
+[//]: # (</tbody>)
+
+[//]: # (</table>)
+
+### ComfyUI Workflow
+
+![Workflow](asset/workflow/workflow.jpg)
+
+<table><tbody>
+  <tr>
+    <th align="center" colspan="4">Example Workflow Case</th>
+  </tr>
+  <tr>
+    <th align="center" colspan="1">Base</th>
+    <th align="center" colspan="1">+Mantra</th>
+    <th align="center" colspan="1">+Tuner</th>
+    <th align="center" colspan="1">+Control</th>
+  </tr>
+  <tr>
+    <td>
+      <a href="asset/workflow/sdxl_base.json" target="_blank">
+        <img src="asset/workflow/sdxl_base.jpg" width="200">
+      </a>
+    </td>
+    <td>
+      <a href="asset/workflow/sdxl_base_mantra.json" target="_blank">
+        <img src="asset/workflow/sdxl_base_mantra.jpg" width="200">
+      </a>
+    </td>
+    <td>
+      <a href="asset/workflow/sdxl_base_mantra_tuner.json" target="_blank">
+        <img src="asset/workflow/sdxl_base_mantra_tuner.jpg" width="200">
+      </a>
+    </td>
+    <td>
+      <a href="asset/workflow/sdxl_base_mantra_tuner_control.json" target="_blank">
+        <img src="asset/workflow/sdxl_base_mantra_tuner_control.jpg" width="200">
+      </a>
+    </td>
+  </tr>
+</tbody>
+</table>
 
 
 ## 🛠️ Installation
 
-- Create new environment with `conda` command:
-
-```shell
-conda env create -f environment.yaml
-conda activate scepter
-```
-
 - Install with `pip` command:
 
 We recommend installing the specific version of PyTorch and accelerate toolbox [xFormers](https://pypi.org/project/xformers/). You can install these recommended version by pip:
@@ -75,16 +158,19 @@ pip install scepter
 
 ### Currently supported approaches
 
-|            Tasks             |                   Methods                    | Links                                                                                                                                                                                                                                                       |
-|:----------------------------:|:--------------------------------------------:|:------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-|   Text-to-image generation   |                   SD v1.5                    | [![Hugging Face Repo](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Repo-blue)](https://huggingface.co/runwayml/stable-diffusion-v1-5)                                                                                                         |
-|   Text-to-image generation   |                   SD v2.1                    | [![Hugging Face Repo](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Repo-blue)](https://huggingface.co/runwayml/stable-diffusion-v1-5)                                                                                                         |
-|   Text-to-image generation   |                    SD-XL                     | [![Hugging Face Repo](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Repo-blue)](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)                                                                                               |
-|       Efficient Tuning       |                     LoRA                     | [![Arxiv   link](https://img.shields.io/static/v1?label=arXiv&message=LoRA&color=red&logo=arxiv)](https://arxiv.org/abs/2106.09685)                                                                                                                         |
-|       Efficient Tuning       |            Res-Tuning(NeurIPS23)             | [![Arxiv   link](https://img.shields.io/static/v1?label=arXiv&message=Res-Tuing&color=red&logo=arxiv)](https://arxiv.org/abs/2310.19859) [![Page link](https://img.shields.io/badge/Page-ResTuning-Gree)](https://res-tuning.github.io/)                    |
-| Controllable image synthesis | [🌟SCEdit(CVPR24)](docs/en/tasks/scedit.md)  | [![Arxiv   link](https://img.shields.io/static/v1?label=arXiv&message=SCEdit&color=red&logo=arxiv)](https://arxiv.org/abs/2312.11392)   [![Page link](https://img.shields.io/badge/Page-SCEdit-Gree)](https://scedit.github.io/)                            |
-|        Image editing         |     [🌟LAR-Gen](docs/en/tasks/largen.md)     | [![Arxiv   link](https://img.shields.io/static/v1?label=arXiv&message=LARGen&color=red&logo=arxiv)](https://arxiv.org/abs/2403.19534)   [![Page link](https://img.shields.io/badge/Page-LARGen-Gree)](https://ali-vilab.github.io/largen-page/)             |
-|        Image editing         |  [🌟StyleBooth](docs/en/tasks/stylebooth.md) | [![Arxiv   link](https://img.shields.io/static/v1?label=arXiv&message=StyleBooth&color=red&logo=arxiv)](https://arxiv.org/abs/2404.12154)   [![Page link](https://img.shields.io/badge/Page-StyleBooth-Gree)](https://ali-vilab.github.io/stylebooth-page/) |
+|            Tasks             |                     Methods                      | Links                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             |
+|:----------------------------:|:------------------------------------------------:|:------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+|   Text-to-image Generation   |                     SD v1.5                      | [![Hugging Face Repo](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Repo-blue)](https://huggingface.co/runwayml/stable-diffusion-v1-5)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               |
+|   Text-to-image Generation   |                     SD v2.1                      | [![Hugging Face Repo](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Repo-blue)](https://huggingface.co/runwayml/stable-diffusion-v1-5)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               |
+|   Text-to-image Generation   |                      SD-XL                       | [![Hugging Face Repo](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Repo-blue)](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                     |
+|   Text-to-image Generation   |                       FLUX                       | [![Hugging Face Repo](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Repo-blue)](https://huggingface.co/black-forest-labs/FLUX.1-dev)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 |
+|       Efficient Tuning       |                       LoRA                       | [![Arxiv   link](https://img.shields.io/static/v1?label=arXiv&message=LoRA&color=red&logo=arxiv)](https://arxiv.org/abs/2106.09685)                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                               |
+|       Efficient Tuning       |              Res-Tuning(NeurIPS23)               | [![Arxiv   link](https://img.shields.io/static/v1?label=arXiv&message=Res-Tuing&color=red&logo=arxiv)](https://arxiv.org/abs/2310.19859) [![Page link](https://img.shields.io/badge/Page-ResTuning-Gree)](https://res-tuning.github.io/)                                                                                                                                                                                                                                                                                                                                                                                                                          |
+| Controllable Image Synthesis |   [🌟SCEdit(CVPR24)](docs/en/tasks/scedit.md)    | [![Arxiv   link](https://img.shields.io/static/v1?label=arXiv&message=SCEdit&color=red&logo=arxiv)](https://arxiv.org/abs/2312.11392)   [![Page link](https://img.shields.io/badge/Page-SCEdit-Gree)](https://scedit.github.io/)                                                                                                                                                                                                                                                                                                                                                                                                                                  |
+|        Image Editing         |       [🌟LAR-Gen](docs/en/tasks/largen.md)       | [![Arxiv   link](https://img.shields.io/static/v1?label=arXiv&message=LARGen&color=red&logo=arxiv)](https://arxiv.org/abs/2403.19534)   [![Page link](https://img.shields.io/badge/Page-LARGen-Gree)](https://ali-vilab.github.io/largen-page/)                                                                                                                                                                                                                                                                                                                                                                                                                   |
+|        Image Editing         |   [🌟StyleBooth](docs/en/tasks/stylebooth.md)    | [![Arxiv   link](https://img.shields.io/static/v1?label=arXiv&message=StyleBooth&color=red&logo=arxiv)](https://arxiv.org/abs/2404.12154)   [![Page link](https://img.shields.io/badge/Page-StyleBooth-Gree)](https://ali-vilab.github.io/stylebooth-page/)                                                                                                                                                                                                                                                                                                                                                                                                       |
+| Image Generation and Editing |  [🌟ACE](https://ali-vilab.github.io/ace-page/)  | [![Arxiv   link](https://img.shields.io/static/v1?label=arXiv&message=ACE&color=red&logo=arxiv)](https://arxiv.org/abs/2410.00086)   [![Page link](https://img.shields.io/badge/Page-ACE-Gree)](https://ali-vilab.github.io/ace-page/) [![Demo link](https://img.shields.io/badge/Demo-ACE-purple)](https://huggingface.co/spaces/scepter-studio/ACE-Chat) <br> [![ModelScope link](https://img.shields.io/badge/ModelScope-Model-blue)](https://www.modelscope.cn/models/iic/ACE-0.6B-512px)  [![HuggingFace link](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model-yellow)](https://huggingface.co/scepter-studio/ACE-0.6B-512px)               |
+| Image Generation and Editing | [🌟ACE++](https://ali-vilab.github.io/ACE_plus_page/) | [![Arxiv   link](https://img.shields.io/static/v1?label=arXiv&message=ACEPlus&color=red&logo=arxiv)](https://arxiv.org/abs/2501.02487)   [![Page link](https://img.shields.io/badge/Page-ACE++-Gree)](https://ali-vilab.github.io/ACE_plus_page/) [![Demo link](https://img.shields.io/badge/Demo-ACE++-purple)](https://huggingface.co/spaces/scepter-studio/ACE-Plus) <br> [![ModelScope link](https://img.shields.io/badge/ModelScope-Model-blue)](https://www.modelscope.cn/models/iic/ACE_Plus/summary)  [![HuggingFace link](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model-yellow)](https://huggingface.co/ali-vilab/ACE_Plus/tree/main) |
 
 
 ## 🖥️ SCEPTER Studio
@@ -118,6 +204,24 @@ Therefore, subsequent startups will become much faster (about one minute) as dow
 We deploy a work studio on Modelscope that includes only the inference tab, please refer to [ms_scepter_studio](https://www.modelscope.cn/studios/iic/scepter_studio/summary) and [hf_scepter_studio](https://huggingface.co/spaces/modelscope/scepter_studio)
 
 
+
+## ⚙️️ ComfyUI Workflow
+
+We support the use of all models in the ComfyUI Workflow through the following methods:
+
+1) Automatic installation directly via the ComfyUI Manager by searching for the **ComfyUI-Scepter** node.
+2) Manually install by moving custom_nodes from Scepter to ComfyUI.
+```shell
+git clone https://github.com/modelscope/scepter.git
+cd path/to/scepter
+pip install -e .
+cp -r path/to/scepter/workflow/ path/to/ComfyUI/custom_nodes/ComfyUI-Scepter
+cd path/to/ComfyUI
+python main.py
+```
+
+**Note**: You can use the nodes by dragging the sample images into ComfyUI. Additionally, our nodes can automatically pull models from ModelScope or HuggingFace by selecting the *model_source* field, or you can place the already downloaded models in a local path.
+
 ## 🔍 Learn More
 
 - [Alibaba TongYi Vision Intelligence Lab](https://github.com/ali-vilab)
@@ -150,4 +254,4 @@ This project is licensed under the [Apache License (Version 2.0)](https://github
 
 
 ## Acknowledgement
-Thanks to [Stability-AI](https://github.com/Stability-AI), [SWIFT library](https://github.com/modelscope/swift/) and [Fooocus](https://github.com/lllyasviel/Fooocus) for their awesome work.
+Thanks to [Stability-AI](https://github.com/Stability-AI), [SWIFT library](https://github.com/modelscope/swift/), [Fooocus](https://github.com/lllyasviel/Fooocus) and [ComfyUI](https://github.com/comfyanonymous/ComfyUI) for their awesome work.

requirements/framework.txt

@@ -1,9 +1,8 @@
 albumentations
 beautifulsoup4
-bezier
 einops
-modelscope==1.14.0
-ms-swift>=2.0.1
+modelscope[framework]
+ms-swift
 numpy
 open_clip_torch
 opencv-python
@@ -12,5 +11,7 @@ oss2>=2.15.0
 pycocotools
 pyyaml>=5.3.1
 scikit-image
+scikit-learn
+sentencepiece
 torchsde
 transformers

requirements/recommended.txt

@@ -1,4 +1,5 @@
 git+https://github.com/cocodataset/panopticapi.git
-torch==2.0.1
-torchvision==0.15.2
-xformers==0.0.21
+torch==2.4.1
+torchvision==0.19.1
+flash-attn==2.5.8
+xformers==0.0.28

requirements/scepter_studio.txt

@@ -1,5 +1,6 @@
 bitsandbytes
 gradio
+gradio_imageslider
 imagehash
 psutil
 tiktoken

scepter/__init__.py

@@ -1,18 +1,28 @@
 # -*- coding: utf-8 -*-
 # Copyright (c) Alibaba, Inc. and its affiliates.
-import os
+from typing import TYPE_CHECKING
+from scepter.modules.utils.import_utils import LazyImportModule
 
-import scepter
-from scepter.modules import data, model, opt, solver, transform, utils
-from scepter.tools.helper import get_module_list as module_list
-from scepter.tools.helper import \
-    get_module_object_config as configures_by_objects
-from scepter.tools.helper import get_module_objects as objects_by_module
-from scepter.version import __version__, version_info
 
-dirname = os.path.dirname(scepter.__file__)
+if TYPE_CHECKING:
+    from scepter.modules import data, model, opt, solver, transform, utils
+    from scepter.tools.helper import get_module_list as module_list
+    from scepter.tools.helper import \
+        get_module_object_config as configures_by_objects
+    from scepter.tools.helper import get_module_objects as objects_by_module
+    from scepter.version import __version__, version_info
+else:
+    _import_structure = {
+        'modules': ['data', 'model', 'opt', 'solver', 'transform', 'utils'],
+        'helper': ['get_module_list', 'get_module_object_config', 'get_module_objects'],
+        'version': ['__version__', 'version_info']
+    }
 
-__all__ = [
-    utils, transform, data, model, solver, version_info, opt, '__version__',
-    'dirname'
-]
+    import sys
+    sys.modules[__name__] = LazyImportModule(
+        __name__,
+        globals()['__file__'],
+        _import_structure,
+        module_spec=__spec__,
+        extra_objects={},
+    )

scepter/methods/edit/dit_ace_0.6b_1024.yaml

@@ -0,0 +1,161 @@
+ENV:
+  BACKEND: nccl
+  SEED: 2024
+#
+SOLVER:
+  NAME: ACESolver
+  RESUME_FROM:
+  LOAD_MODEL_ONLY: True
+  USE_FSDP: False
+  SHARDING_STRATEGY:
+  USE_AMP: True
+  DTYPE: float16
+  CHANNELS_LAST: True
+  MAX_STEPS: 500
+  MAX_EPOCHS: -1
+  NUM_FOLDS: 1
+  ACCU_STEP: 1
+  EVAL_INTERVAL: 50
+  RESCALE_LR: False
+  #
+  WORK_DIR: ./cache/save_data/ace_0.6b_1024
+  LOG_FILE: std_log.txt
+  #
+  FILE_SYSTEM:
+    - NAME: "HuggingfaceFs"
+      TEMP_DIR: ./cache/cache_data
+    - NAME: "LocalFs"
+      TEMP_DIR: ./cache/cache_data
+    - NAME: "ModelscopeFs"
+      TEMP_DIR: ./cache/cache_data
+
+  #
+  MODEL:
+    NAME: LatentDiffusionACE
+    PRETRAINED_MODEL:
+    IGNORE_KEYS: [ ]
+    SCALE_FACTOR: 0.18215
+    SIZE_FACTOR: 8
+    DECODER_BIAS: 0.5
+    DEFAULT_N_PROMPT:
+    USE_EMA: True
+    EVAL_EMA: False
+    TEXT_IDENTIFIER: [ '{image}', '{image1}', '{image2}', '{image3}', '{image4}', '{image5}', '{image6}', '{image7}', '{image8}', '{image9}' ]
+    USE_TEXT_POS_EMBEDDINGS: True
+    #
+    DIFFUSION:
+      NAME: BaseDiffusion
+      PREDICTION_TYPE: eps
+      MIN_SNR_GAMMA:
+      NOISE_SCHEDULER:
+        NAME: LinearScheduler
+        NUM_TIMESTEPS: 1000
+        BETA_MIN: 0.0001
+        BETA_MAX: 0.02
+    #
+    DIFFUSION_MODEL:
+      NAME: ACE
+      PRETRAINED_MODEL: ms://iic/ACE-0.6B-1024px@models/dit/ace_0.6b_1024px.pth
+      IGNORE_KEYS: [ ]
+      PATCH_SIZE: 2
+      IN_CHANNELS: 4
+      HIDDEN_SIZE: 1152
+      DEPTH: 28
+      NUM_HEADS: 16
+      MLP_RATIO: 4.0
+      PRED_SIGMA: True
+      DROP_PATH: 0.0
+      WINDOW_DIZE: 0
+      Y_CHANNELS: 4096
+      MAX_SEQ_LEN: 4096
+      QK_NORM: True
+      USE_GRAD_CHECKPOINT: True
+      ATTENTION_BACKEND: flash_attn
+    #
+    FIRST_STAGE_MODEL:
+      NAME: AutoencoderKL
+      EMBED_DIM: 4
+      PRETRAINED_MODEL: ms://iic/ACE-0.6B-1024px@models/vae/vae.bin
+      IGNORE_KEYS: []
+      #
+      ENCODER:
+        NAME: Encoder
+        CH: 128
+        OUT_CH: 3
+        NUM_RES_BLOCKS: 2
+        IN_CHANNELS: 3
+        ATTN_RESOLUTIONS: [ ]
+        CH_MULT: [ 1, 2, 4, 4 ]
+        Z_CHANNELS: 4
+        DOUBLE_Z: True
+        DROPOUT: 0.0
+        RESAMP_WITH_CONV: True
+      #
+      DECODER:
+        NAME: Decoder
+        CH: 128
+        OUT_CH: 3
+        NUM_RES_BLOCKS: 2
+        IN_CHANNELS: 3
+        ATTN_RESOLUTIONS: [ ]
+        CH_MULT: [ 1, 2, 4, 4 ]
+        Z_CHANNELS: 4
+        DROPOUT: 0.0
+        RESAMP_WITH_CONV: True
+        GIVE_PRE_END: False
+        TANH_OUT: False
+    #
+    COND_STAGE_MODEL:
+      NAME: T5EmbedderHF
+      PRETRAINED_MODEL: ms://iic/ACE-0.6B-1024px@models/text_encoder/t5-v1_1-xxl/
+      TOKENIZER_PATH: ms://iic/ACE-0.6B-1024px@models/tokenizer/t5-v1_1-xxl
+      LENGTH: 120
+      T5_DTYPE: bfloat16
+      ADDED_IDENTIFIER: [ '{image}', '{caption}', '{mask}', '{ref_image}', '{image1}', '{image2}', '{image3}', '{image4}', '{image5}', '{image6}', '{image7}', '{image8}', '{image9}' ]
+      CLEAN: whitespace
+      USE_GRAD: False
+    LOSS:
+      NAME: ReconstructLoss
+      LOSS_TYPE: l2
+  #
+  SAMPLE_ARGS:
+    SAMPLER: ddim
+    SAMPLE_STEPS: 20
+    GUIDE_SCALE: 4.5
+    GUIDE_RESCALE: 0.5
+  #
+  OPTIMIZER:
+    NAME: AdamW
+    LEARNING_RATE: 1e-7
+    EPS: 1e-10
+    WEIGHT_DECAY: 5e-4
+  #
+  TRAIN_DATA:
+    NAME: ImageTextPairMSDatasetForACE
+    MODE: train
+    MS_DATASET_NAME: cache/datasets/hed_pair
+    MS_DATASET_NAMESPACE: ""
+    MS_DATASET_SPLIT: "train"
+    MS_DATASET_SUBNAME: ""
+    PROMPT_PREFIX: ""
+    REPLACE_STYLE: False
+    MAX_SEQ_LEN: 4096
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    NUM_WORKERS: 1
+    SAMPLER:
+      NAME: LoopSampler
+  #
+  TRAIN_HOOKS:
+    -
+      NAME: BackwardHook
+      PRIORITY: 0
+    -
+      NAME: LogHook
+      LOG_INTERVAL: 50
+    -
+      NAME: CheckpointHook
+      INTERVAL: 100
+    -
+      NAME: ProbeDataHook
+      PROB_INTERVAL: 100

scepter/methods/edit/dit_ace_0.6b_512.yaml

@@ -0,0 +1,161 @@
+ENV:
+  BACKEND: nccl
+  SEED: 2024
+#
+SOLVER:
+  NAME: ACESolver
+  RESUME_FROM:
+  LOAD_MODEL_ONLY: True
+  USE_FSDP: False
+  SHARDING_STRATEGY:
+  USE_AMP: True
+  DTYPE: float16
+  CHANNELS_LAST: True
+  MAX_STEPS: 500
+  MAX_EPOCHS: -1
+  NUM_FOLDS: 1
+  ACCU_STEP: 1
+  EVAL_INTERVAL: 50
+  RESCALE_LR: False
+  #
+  WORK_DIR: ./cache/save_data/ace_0.6b_512
+  LOG_FILE: std_log.txt
+  #
+  FILE_SYSTEM:
+    - NAME: "HuggingfaceFs"
+      TEMP_DIR: ./cache/cache_data
+    - NAME: "LocalFs"
+      TEMP_DIR: ./cache/cache_data
+    - NAME: "ModelscopeFs"
+      TEMP_DIR: ./cache/cache_data
+
+  #
+  MODEL:
+    NAME: LatentDiffusionACE
+    PRETRAINED_MODEL:
+    IGNORE_KEYS: [ ]
+    SCALE_FACTOR: 0.18215
+    SIZE_FACTOR: 8
+    DECODER_BIAS: 0.5
+    DEFAULT_N_PROMPT:
+    USE_EMA: True
+    EVAL_EMA: False
+    TEXT_IDENTIFIER: [ '{image}', '{image1}', '{image2}', '{image3}', '{image4}', '{image5}', '{image6}', '{image7}', '{image8}', '{image9}' ]
+    USE_TEXT_POS_EMBEDDINGS: True
+    #
+    DIFFUSION:
+      NAME: BaseDiffusion
+      PREDICTION_TYPE: eps
+      MIN_SNR_GAMMA:
+      NOISE_SCHEDULER:
+        NAME: LinearScheduler
+        NUM_TIMESTEPS: 1000
+        BETA_MIN: 0.0001
+        BETA_MAX: 0.02
+    #
+    DIFFUSION_MODEL:
+      NAME: ACE
+      PRETRAINED_MODEL: ms://iic/ACE-0.6B-512px@models/dit/ace_0.6b_512px.pth
+      IGNORE_KEYS: [ ]
+      PATCH_SIZE: 2
+      IN_CHANNELS: 4
+      HIDDEN_SIZE: 1152
+      DEPTH: 28
+      NUM_HEADS: 16
+      MLP_RATIO: 4.0
+      PRED_SIGMA: True
+      DROP_PATH: 0.0
+      WINDOW_DIZE: 0
+      Y_CHANNELS: 4096
+      MAX_SEQ_LEN: 1024
+      QK_NORM: True
+      USE_GRAD_CHECKPOINT: True
+      ATTENTION_BACKEND: flash_attn
+    #
+    FIRST_STAGE_MODEL:
+      NAME: AutoencoderKL
+      EMBED_DIM: 4
+      PRETRAINED_MODEL: ms://iic/ACE-0.6B-512px@models/vae/vae.bin
+      IGNORE_KEYS: []
+      #
+      ENCODER:
+        NAME: Encoder
+        CH: 128
+        OUT_CH: 3
+        NUM_RES_BLOCKS: 2
+        IN_CHANNELS: 3
+        ATTN_RESOLUTIONS: [ ]
+        CH_MULT: [ 1, 2, 4, 4 ]
+        Z_CHANNELS: 4
+        DOUBLE_Z: True
+        DROPOUT: 0.0
+        RESAMP_WITH_CONV: True
+      #
+      DECODER:
+        NAME: Decoder
+        CH: 128
+        OUT_CH: 3
+        NUM_RES_BLOCKS: 2
+        IN_CHANNELS: 3
+        ATTN_RESOLUTIONS: [ ]
+        CH_MULT: [ 1, 2, 4, 4 ]
+        Z_CHANNELS: 4
+        DROPOUT: 0.0
+        RESAMP_WITH_CONV: True
+        GIVE_PRE_END: False
+        TANH_OUT: False
+    #
+    COND_STAGE_MODEL:
+      NAME: T5EmbedderHF
+      PRETRAINED_MODEL: ms://iic/ACE-0.6B-512px@models/text_encoder/t5-v1_1-xxl/
+      TOKENIZER_PATH: ms://iic/ACE-0.6B-512px@models/tokenizer/t5-v1_1-xxl
+      LENGTH: 120
+      T5_DTYPE: bfloat16
+      ADDED_IDENTIFIER: [ '{image}', '{caption}', '{mask}', '{ref_image}', '{image1}', '{image2}', '{image3}', '{image4}', '{image5}', '{image6}', '{image7}', '{image8}', '{image9}' ]
+      CLEAN: whitespace
+      USE_GRAD: False
+    LOSS:
+      NAME: ReconstructLoss
+      LOSS_TYPE: l2
+  #
+  SAMPLE_ARGS:
+    SAMPLER: ddim
+    SAMPLE_STEPS: 20
+    GUIDE_SCALE: 4.5
+    GUIDE_RESCALE: 0.5
+  #
+  OPTIMIZER:
+    NAME: AdamW
+    LEARNING_RATE: 1e-7
+    EPS: 1e-10
+    WEIGHT_DECAY: 5e-4
+  #
+  TRAIN_DATA:
+    NAME: ImageTextPairMSDatasetForACE
+    MODE: train
+    MS_DATASET_NAME: cache/datasets/hed_pair
+    MS_DATASET_NAMESPACE: ""
+    MS_DATASET_SPLIT: "train"
+    MS_DATASET_SUBNAME: ""
+    PROMPT_PREFIX: ""
+    REPLACE_STYLE: False
+    MAX_SEQ_LEN: 1024
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    NUM_WORKERS: 1
+    SAMPLER:
+      NAME: LoopSampler
+  #
+  TRAIN_HOOKS:
+    -
+      NAME: BackwardHook
+      PRIORITY: 0
+    -
+      NAME: LogHook
+      LOG_INTERVAL: 50
+    -
+      NAME: CheckpointHook
+      INTERVAL: 100
+    -
+      NAME: ProbeDataHook
+      PROB_INTERVAL: 100

scepter/methods/examples/generation/dit_cogvideox1.5_5b_i2v_lora.yaml

@@ -0,0 +1,277 @@
+ENV:
+  BACKEND: nccl
+  SEED: 42
+  TENSOR_PARALLEL_SIZE: 1
+  PIPELINE_PARALLEL_SIZE: 1
+  SYS_ENVS:
+    TORCH_CUDNN_V8_API_ENABLED: '1'
+    TOKENIZERS_PARALLELISM: 'false'
+    TF_CPP_MIN_LOG_LEVEL: '3'
+    PYTORCH_CUDA_ALLOC_CONF: 'expandable_segments:True'
+#
+SOLVER:
+  NAME: LatentDiffusionVideoSolver
+  MAX_STEPS: 2000
+  USE_AMP: True
+  DTYPE: bfloat16
+  USE_FAIRSCALE: False
+  USE_FSDP: True
+  LOAD_MODEL_ONLY: False
+  ENABLE_GRADSCALER: False
+  USE_SCALER: False
+  RESUME_FROM:
+  WORK_DIR: ./cache/save_data/dit_cogvideox1.5_5b_i2v_lora
+  LOG_FILE: std_log.txt
+  EVAL_INTERVAL: 100
+  LOG_TRAIN_NUM: 4
+  FPS: 16
+  SHARDING_STRATEGY: full_shard
+  FSDP_REDUCE_DTYPE: float32
+  FSDP_BUFFER_DTYPE: float32
+  FSDP_SHARD_MODULES: [ 'model', 'cond_stage_model.model']
+  SAVE_MODULES: [ 'model', 'cond_stage_model.model']
+  TRAIN_MODULES: ['model']
+  #
+  FILE_SYSTEM:
+    NAME: "ModelscopeFs"
+    TEMP_DIR: "./cache/cache_data"
+  #
+  TUNER:
+    - NAME: SwiftLoRA
+      R: 64
+      LORA_ALPHA: 64
+      LORA_DROPOUT: 0.0
+      BIAS: "none"
+      TARGET_MODULES: "model.*(.to_k|.to_q|.to_v|.to_out.0)$"
+  #
+  MODEL:
+    NAME: LatentDiffusionCogVideoX
+    PRETRAINED_MODEL:
+    PARAMETERIZATION: v
+    TIMESTEPS: 1000
+    MIN_SNR_GAMMA: 3.0
+    ZERO_TERMINAL_SNR: True
+    SCALE_FACTOR_SPATIAL: 8
+    SCALE_FACTOR_TEMPORAL: 4
+    SCALING_FACTOR_IMAGE: 0.7  
+    NOISED_IMAGE_DROPOUT: 0.05
+    INVERT_SCALE_LATENTS: True
+    IGNORE_KEYS: [ ]
+    DEFAULT_N_PROMPT:
+    USE_EMA: False
+    EVAL_EMA: False
+    DIFFUSION:
+      NAME: BaseDiffusion
+      PREDICTION_TYPE: v
+      USE_DYNAMIC_CFG: False
+      NOISE_SCHEDULER:
+        NAME: ScaledLinearScheduler
+        BETA_MIN: 0.00085
+        BETA_MAX: 0.012
+        SNR_SHIFT_SCALE: 1.0  
+        RESCALE_BETAS_ZERO_SNR: True
+      DIFFUSION_SAMPLERS:
+        NAME: DDIMSampler
+        DISCRETIZATION_TYPE: trailing
+        ETA: 0.0
+    #
+    DIFFUSION_MODEL:
+      NAME: CogVideoXTransformer3DModel
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: # 5b-I2V diff
+        - ms://ZhipuAI/CogVideoX1.5-5B-I2V@transformer/diffusion_pytorch_model-00001-of-00003.safetensors
+        - ms://ZhipuAI/CogVideoX1.5-5B-I2V@transformer/diffusion_pytorch_model-00002-of-00003.safetensors
+        - ms://ZhipuAI/CogVideoX1.5-5B-I2V@transformer/diffusion_pytorch_model-00003-of-00003.safetensors
+      NUM_ATTENTION_HEADS: 48  
+      ATTENTION_HEAD_DIM: 64
+      IN_CHANNELS: 32
+      LATENT_CHANNELS: 16
+      OUT_CHANNELS: 16
+      FLIP_SIN_TO_COS: True
+      FREQ_SHIFT: 0
+      TIME_EMBED_DIM: 512
+      TEXT_EMBED_DIM: 4096
+      OFS_EMBED_DIM: 512 # v1.5 diff
+      NUM_LAYERS: 42  
+      DROPOUT: 0.0
+      ATTENTION_BIAS: True
+      SAMPLE_WIDTH: 300
+      SAMPLE_HEIGHT: 300
+      SAMPLE_FRAMES: 81
+      PATCH_SIZE: 2
+      PATCH_SIZE_T: 2  # v1.5 diff
+      PATCH_BIAS: False  # v1.5 diff
+      TEMPORAL_COMPRESSION_RATIO: 4
+      MAX_TEXT_SEQ_LENGTH: 224
+      ACTIVATION_FN: "gelu-approximate"
+      TIMESTEP_ACTIVATION_FN: "silu"
+      NORM_ELEMENTWISE_AFFINE: True
+      NORM_EPS: 1e-5
+      SPATIAL_INTERPOLATION_SCALE: 1.875
+      TEMPORAL_INTERPOLATION_SCALE: 1.0
+      USE_ROTARY_POSITIONAL_EMBEDDINGS: True 
+      USE_LEARNED_POSITIONAL_EMBEDDINGS: False
+      GRADIENT_CHECKPOINTING: True
+    #
+    FIRST_STAGE_MODEL:
+      NAME: AutoencoderKLCogVideoX
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: ms://ZhipuAI/CogVideoX1.5-5B-I2V@vae/diffusion_pytorch_model.safetensors 
+      SAMPLE_HEIGHT: 768
+      SAMPLE_WIDTH: 1360
+      USE_QUANT_CONV: False
+      USE_POST_QUANT_CONV: False
+      USE_SLICING: True
+      USE_TILING: True
+      GRADIENT_CHECKPOINTING: True
+      ENCODER:
+        NAME: CogVideoXEncoder3D
+        IN_CHANNELS: 3
+        OUT_CHANNELS: 16
+        UP_BLOCK_TYPES: [ "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+      DECODER:
+        NAME: CogVideoXDecoder3D
+        IN_CHANNELS: 16
+        OUT_CHANNELS: 3
+        UP_BLOCK_TYPES: [ "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+    #
+    COND_STAGE_MODEL:
+      NAME: T5EmbedderHF
+      PRETRAINED_MODEL: ms://AI-ModelScope/t5-v1_1-xxl
+      TOKENIZER_PATH: ms://AI-ModelScope/t5-v1_1-xxl
+      LENGTH: 224
+      CLEAN:
+      USE_GRAD: False
+      T5_DTYPE: bfloat16
+    #
+    LOSS:
+      NAME: ReconstructLoss
+      LOSS_TYPE: l2
+  #
+  SAMPLE_ARGS:
+    SAMPLER: ddim
+    SAMPLE_STEPS: 50
+    SEED: 42
+    GUIDE_SCALE: 6.0
+    GUIDE_RESCALE: 0.0
+    NUM_FRAMES: 81
+    IMAGE_SIZE: [768, 1360]
+  #
+  OPTIMIZER:
+    NAME: Adam
+    LEARNING_RATE: 1e-3
+    BETAS: [ 0.9, 0.95 ]
+    EPS: 1e-8
+    WEIGHT_DECAY: 0.0
+    AMSGRAD: False
+  #
+#  LR_SCHEDULER:
+#    NAME: StepAnnealingLR
+#    WARMUP_STEPS: 200
+#    TOTAL_STEPS: 2000
+#    DECAY_MODE: 'cosine'
+  #
+  TRAIN_DATA:
+    NAME: VideoGenDataset
+    MODE: train
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    NUM_WORKERS: 0
+    NUM_FRAMES: 85
+    FPS: 16
+    HEIGHT: 768
+    WIDTH: 1360
+    PROMPT_PREFIX: 'DISNEY '
+    DATA_TYPE: 'i2v'
+    SAMPLER:
+      NAME: MixtureOfSamplers
+      SUB_SAMPLERS:
+        - NAME: MultiLevelBatchSampler
+          PROB: 1.0
+          FIELDS: [ "video_path", "prompt" ]
+          DELIMITER: '#;#'
+          PATH_PREFIX: cache/datasets/Disney-VideoGeneration-Dataset/
+          INDEX_FILE: cache/datasets/Disney-VideoGeneration-Dataset/index.jsonl
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ "video", "image", "prompt" ]
+        META_KEYS: [ ]
+  #
+#  EVAL_DATA:
+#    NAME: Text2ImageDataset
+#    MODE: eval
+#    PROMPT_FILE:
+#    PROMPT_DATA: [ "A cat running.#;#asset/images/edit_tuner/cat_512.jpg" ]
+#    FIELDS: [ "prompt", "img_path" ]
+#    DELIMITER: '#;#'
+#    PROMPT_PREFIX: ''
+#    PIN_MEMORY: True
+#    BATCH_SIZE: 1
+#    USE_NUM: 8
+#    NUM_WORKERS: 0
+#    IMAGE_SIZE: [768, 1360]
+#    TRANSFORMS:
+#      - NAME: LoadImageFromFileList
+#        FILE_KEYS: [ 'img_path' ]
+#        RGB_ORDER: RGB
+#        BACKEND: pillow
+#      - NAME: FlexibleResize
+#        INTERPOLATION: bilinear
+#        SIZE: [768, 1360]
+#        INPUT_KEY: [ 'img' ]
+#        OUTPUT_KEY: [ 'img' ]
+#        BACKEND: pillow
+#      - NAME: FlexibleCenterCrop
+#        SIZE: [768, 1360]
+#        INPUT_KEY: [ 'img' ]
+#        OUTPUT_KEY: [ 'img' ]
+#        BACKEND: pillow
+#      - NAME: ImageToTensor
+#        INPUT_KEY: [ 'img' ]
+#        OUTPUT_KEY: [ 'img' ]
+#        BACKEND: pillow
+#      - NAME: Normalize
+#        MEAN: [ 0.5,  0.5,  0.5 ]
+#        STD: [ 0.5,  0.5,  0.5 ]
+#        INPUT_KEY: [ 'img' ]
+#        OUTPUT_KEY: [ 'image' ]
+#        BACKEND: torchvision
+#      - NAME: Select
+#        KEYS: [ 'image', 'prompt' ]
+#        META_KEYS: [ 'image_size' ]
+  #
+  TRAIN_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0
+    - NAME: BackwardHook
+      PRIORITY: 10
+    - NAME: LogHook
+      LOG_INTERVAL: 10
+      PRIORITY: 20
+    - NAME: CheckpointHook
+      INTERVAL: 1000
+      PRIORITY: 40
+  #
+#  EVAL_HOOKS:
+#    - NAME: ProbeDataHook
+#      PROB_INTERVAL: 100
+#      PRIORITY: 0

scepter/methods/examples/generation/dit_cogvideox1.5_5b_lora.yaml

@@ -0,0 +1,248 @@
+ENV:
+  BACKEND: nccl
+  SEED: 42
+  TENSOR_PARALLEL_SIZE: 1
+  PIPELINE_PARALLEL_SIZE: 1
+  SYS_ENVS:
+    TORCH_CUDNN_V8_API_ENABLED: '1'
+    TOKENIZERS_PARALLELISM: 'false'
+    TF_CPP_MIN_LOG_LEVEL: '3'
+    PYTORCH_CUDA_ALLOC_CONF: 'expandable_segments:True'
+#
+SOLVER:
+  NAME: LatentDiffusionVideoSolver
+  MAX_STEPS: 2000
+  USE_AMP: True
+  DTYPE: bfloat16
+  USE_FAIRSCALE: False
+  USE_FSDP: True
+  LOAD_MODEL_ONLY: False
+  ENABLE_GRADSCALER: False
+  USE_SCALER: False
+  RESUME_FROM:
+  WORK_DIR: ./cache/save_data/dit_cogvideox1.5_5b_lora
+  LOG_FILE: std_log.txt
+  EVAL_INTERVAL: 100
+  LOG_TRAIN_NUM: 4
+  FPS: 16
+  SHARDING_STRATEGY: full_shard
+  FSDP_REDUCE_DTYPE: float32
+  FSDP_BUFFER_DTYPE: float32
+  FSDP_SHARD_MODULES: [ 'model', 'cond_stage_model.model']
+  SAVE_MODULES: [ 'model', 'cond_stage_model.model']
+  TRAIN_MODULES: ['model']
+  #
+  FILE_SYSTEM:
+    NAME: "ModelscopeFs"
+    TEMP_DIR: "./cache/cache_data"
+  #
+  TUNER:
+    - NAME: SwiftLoRA
+      R: 64
+      LORA_ALPHA: 64
+      LORA_DROPOUT: 0.0
+      BIAS: "none"
+      TARGET_MODULES: "model.*(.to_k|.to_q|.to_v|.to_out.0)$"
+  #
+  MODEL:
+    NAME: LatentDiffusionCogVideoX
+    PRETRAINED_MODEL:
+    PARAMETERIZATION: v
+    TIMESTEPS: 1000
+    MIN_SNR_GAMMA: 3.0
+    ZERO_TERMINAL_SNR: True
+    SCALE_FACTOR_SPATIAL: 8
+    SCALE_FACTOR_TEMPORAL: 4
+    SCALING_FACTOR_IMAGE: 0.7
+    INVERT_SCALE_LATENTS: True
+    IGNORE_KEYS: [ ]
+    DEFAULT_N_PROMPT:
+    USE_EMA: False
+    EVAL_EMA: False
+    DIFFUSION:
+      NAME: BaseDiffusion
+      PREDICTION_TYPE: v
+      USE_DYNAMIC_CFG: False
+      NOISE_SCHEDULER:
+        NAME: ScaledLinearScheduler
+        BETA_MIN: 0.00085
+        BETA_MAX: 0.012
+        SNR_SHIFT_SCALE: 1.0  
+        RESCALE_BETAS_ZERO_SNR: True
+      DIFFUSION_SAMPLERS:
+        NAME: DDIMSampler
+        DISCRETIZATION_TYPE: trailing
+        ETA: 0.0
+    #
+    DIFFUSION_MODEL:
+      NAME: CogVideoXTransformer3DModel
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: 
+        - ms://ZhipuAI/CogVideoX1.5-5B@transformer/diffusion_pytorch_model-00001-of-00003.safetensors
+        - ms://ZhipuAI/CogVideoX1.5-5B@transformer/diffusion_pytorch_model-00002-of-00003.safetensors
+        - ms://ZhipuAI/CogVideoX1.5-5B@transformer/diffusion_pytorch_model-00003-of-00003.safetensors
+      NUM_ATTENTION_HEADS: 48  
+      ATTENTION_HEAD_DIM: 64
+      IN_CHANNELS: 16
+      OUT_CHANNELS: 16
+      FLIP_SIN_TO_COS: True
+      FREQ_SHIFT: 0
+      TIME_EMBED_DIM: 512
+      TEXT_EMBED_DIM: 4096
+      NUM_LAYERS: 42  
+      DROPOUT: 0.0
+      ATTENTION_BIAS: True
+      SAMPLE_WIDTH: 300
+      SAMPLE_HEIGHT: 300
+      SAMPLE_FRAMES: 81
+      PATCH_SIZE: 2
+      PATCH_SIZE_T: 2  # v1.5 diff
+      PATCH_BIAS: False  # v1.5 diff
+      TEMPORAL_COMPRESSION_RATIO: 4
+      MAX_TEXT_SEQ_LENGTH: 224
+      ACTIVATION_FN: "gelu-approximate"
+      TIMESTEP_ACTIVATION_FN: "silu"
+      NORM_ELEMENTWISE_AFFINE: True
+      NORM_EPS: 1e-5
+      SPATIAL_INTERPOLATION_SCALE: 1.875
+      TEMPORAL_INTERPOLATION_SCALE: 1.0
+      USE_ROTARY_POSITIONAL_EMBEDDINGS: True 
+      USE_LEARNED_POSITIONAL_EMBEDDINGS: False
+      GRADIENT_CHECKPOINTING: True
+    #
+    FIRST_STAGE_MODEL:
+      NAME: AutoencoderKLCogVideoX
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: ms://ZhipuAI/CogVideoX1.5-5B@vae/diffusion_pytorch_model.safetensors 
+      SAMPLE_HEIGHT: 768
+      SAMPLE_WIDTH: 1360
+      USE_QUANT_CONV: False
+      USE_POST_QUANT_CONV: False
+      USE_SLICING: True
+      USE_TILING: True
+      GRADIENT_CHECKPOINTING: True
+      ENCODER:
+        NAME: CogVideoXEncoder3D
+        IN_CHANNELS: 3
+        OUT_CHANNELS: 16
+        UP_BLOCK_TYPES: [ "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+      DECODER:
+        NAME: CogVideoXDecoder3D
+        IN_CHANNELS: 16
+        OUT_CHANNELS: 3
+        UP_BLOCK_TYPES: [ "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+    #
+    COND_STAGE_MODEL:
+      NAME: T5EmbedderHF
+      PRETRAINED_MODEL: ms://AI-ModelScope/t5-v1_1-xxl
+      TOKENIZER_PATH: ms://AI-ModelScope/t5-v1_1-xxl
+      LENGTH: 224
+      CLEAN:
+      USE_GRAD: False
+      T5_DTYPE: bfloat16
+    #
+    LOSS:
+      NAME: ReconstructLoss
+      LOSS_TYPE: l2
+  #
+  SAMPLE_ARGS:
+    SAMPLER: ddim
+    SAMPLE_STEPS: 50
+    SEED: 42
+    GUIDE_SCALE: 6.0
+    GUIDE_RESCALE: 0.0
+    NUM_FRAMES: 81
+    IMAGE_SIZE: [768, 1360]
+  #
+  OPTIMIZER:
+    NAME: Adam
+    LEARNING_RATE: 1e-3
+    BETAS: [ 0.9, 0.95 ]
+    EPS: 1e-8
+    WEIGHT_DECAY: 0.0
+    AMSGRAD: False
+  #
+#  LR_SCHEDULER:
+#    NAME: StepAnnealingLR
+#    WARMUP_STEPS: 200
+#    TOTAL_STEPS: 2000
+#    DECAY_MODE: 'cosine'
+  #
+  TRAIN_DATA:
+    NAME: VideoGenDataset
+    MODE: train
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    NUM_WORKERS: 0
+    NUM_FRAMES: 85
+    FPS: 16
+    HEIGHT: 768
+    WIDTH: 1360
+    PROMPT_PREFIX: 'DISNEY '
+    SAMPLER:
+      NAME: MixtureOfSamplers
+      SUB_SAMPLERS:
+        - NAME: MultiLevelBatchSampler
+          PROB: 1.0
+          FIELDS: [ "video_path", "prompt" ]
+          DELIMITER: '#;#'
+          PATH_PREFIX: cache/datasets/Disney-VideoGeneration-Dataset/
+          INDEX_FILE: cache/datasets/Disney-VideoGeneration-Dataset/index.jsonl
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ 'video', "prompt" ]
+        META_KEYS: [ ]
+  #
+  EVAL_DATA:
+    NAME: Text2ImageDataset
+    MODE: eval
+    PROMPT_FILE:
+    PROMPT_DATA: [ "A girl riding a bike." ]
+    IMAGE_SIZE: [ 768, 1360 ]
+    FIELDS: [ "prompt" ]
+    DELIMITER: '#;#'
+    PROMPT_PREFIX: 'DISNEY ' # ''
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    USE_NUM: 8
+    NUM_WORKERS: 0
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ 'index', 'prompt' ]
+        META_KEYS: [ 'image_size' ]
+  #
+  TRAIN_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0
+    - NAME: BackwardHook
+      PRIORITY: 10
+    - NAME: LogHook
+      LOG_INTERVAL: 10
+      PRIORITY: 20
+    - NAME: CheckpointHook
+      INTERVAL: 1000
+      PRIORITY: 40
+  #
+  EVAL_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0

scepter/methods/examples/generation/dit_cogvideox_2b_lora.yaml

@@ -0,0 +1,239 @@
+ENV:
+  BACKEND: nccl
+  SEED: 42
+  TENSOR_PARALLEL_SIZE: 1
+  PIPELINE_PARALLEL_SIZE: 1
+  SYS_ENVS:
+    TORCH_CUDNN_V8_API_ENABLED: '1'
+    TOKENIZERS_PARALLELISM: 'false'
+    TF_CPP_MIN_LOG_LEVEL: '3'
+    PYTORCH_CUDA_ALLOC_CONF: 'expandable_segments:True'
+#
+SOLVER:
+  NAME: LatentDiffusionVideoSolver
+  MAX_STEPS: 2000
+  USE_AMP: True
+  DTYPE: bfloat16
+  USE_FAIRSCALE: False
+  USE_FSDP: True
+  LOAD_MODEL_ONLY: False
+  RESUME_FROM:
+  WORK_DIR: ./cache/save_data/dit_cogvideox_2b_lora
+  LOG_FILE: std_log.txt
+  EVAL_INTERVAL: 100
+  LOG_TRAIN_NUM: 4
+  ENABLE_GRADSCALER: False
+  USE_SCALER: False
+  FPS: 8
+  SHARDING_STRATEGY: full_shard
+  FSDP_REDUCE_DTYPE: float32
+  FSDP_BUFFER_DTYPE: float32
+  FSDP_SHARD_MODULES: [ 'model', 'cond_stage_model.model']
+  SAVE_MODULES: [ 'model', 'cond_stage_model.model']
+  TRAIN_MODULES: ['model']
+  #
+  FILE_SYSTEM:
+    NAME: "ModelscopeFs"
+    TEMP_DIR: "./cache/cache_data"
+  #
+  TUNER:
+    - NAME: SwiftLoRA
+      R: 64
+      LORA_ALPHA: 64
+      LORA_DROPOUT: 0.0
+      BIAS: "none"
+      TARGET_MODULES: "model.*(.to_k|.to_q|.to_v|.to_out.0)$"
+  #
+  MODEL:
+    NAME: LatentDiffusionCogVideoX
+    PRETRAINED_MODEL:
+    PARAMETERIZATION: v
+    TIMESTEPS: 1000
+    MIN_SNR_GAMMA: 3.0
+    ZERO_TERMINAL_SNR: True
+    SCALE_FACTOR_SPATIAL: 8
+    SCALE_FACTOR_TEMPORAL: 4
+    SCALING_FACTOR_IMAGE: 1.15258426
+    IGNORE_KEYS: [ ]
+    DEFAULT_N_PROMPT:
+    USE_EMA: False
+    EVAL_EMA: False
+    DIFFUSION:
+      NAME: BaseDiffusion
+      PREDICTION_TYPE: v
+      NOISE_SCHEDULER:
+        NAME: ScaledLinearScheduler
+        BETA_MIN: 0.00085
+        BETA_MAX: 0.012
+        SNR_SHIFT_SCALE: 3.0
+        RESCALE_BETAS_ZERO_SNR: True
+      DIFFUSION_SAMPLERS:
+        NAME: DDIMSampler
+        DISCRETIZATION_TYPE: trailing
+        ETA: 0.0
+    #
+    DIFFUSION_MODEL:
+      NAME: CogVideoXTransformer3DModel
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: ms://AI-ModelScope/CogVideoX-2b@transformer/diffusion_pytorch_model.safetensors
+      NUM_ATTENTION_HEADS: 30
+      ATTENTION_HEAD_DIM: 64
+      IN_CHANNELS: 16
+      OUT_CHANNELS: 16
+      FLIP_SIN_TO_COS: True
+      FREQ_SHIFT: 0
+      TIME_EMBED_DIM: 512
+      TEXT_EMBED_DIM: 4096
+      NUM_LAYERS: 30
+      DROPOUT: 0.0
+      ATTENTION_BIAS: True
+      SAMPLE_WIDTH: 90
+      SAMPLE_HEIGHT: 60
+      SAMPLE_FRAMES: 49
+      PATCH_SIZE: 2
+      TEMPORAL_COMPRESSION_RATIO: 4
+      MAX_TEXT_SEQ_LENGTH: 226
+      ACTIVATION_FN: "gelu-approximate"
+      TIMESTEP_ACTIVATION_FN: "silu"
+      NORM_ELEMENTWISE_AFFINE: True
+      NORM_EPS: 1e-5
+      SPATIAL_INTERPOLATION_SCALE: 1.875
+      TEMPORAL_INTERPOLATION_SCALE: 1.0
+      USE_ROTARY_POSITIONAL_EMBEDDINGS: False
+      USE_LEARNED_POSITIONAL_EMBEDDINGS: False
+      GRADIENT_CHECKPOINTING: False
+    #
+    FIRST_STAGE_MODEL:
+      NAME: AutoencoderKLCogVideoX
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: ms://AI-ModelScope/CogVideoX-2b@vae/diffusion_pytorch_model.safetensors
+      SAMPLE_HEIGHT: 480
+      SAMPLE_WIDTH: 720
+      USE_QUANT_CONV: False
+      USE_POST_QUANT_CONV: False
+      USE_SLICING: True
+      USE_TILING: True
+      GRADIENT_CHECKPOINTING: False
+      ENCODER:
+        NAME: CogVideoXEncoder3D
+        IN_CHANNELS: 3
+        OUT_CHANNELS: 16
+        UP_BLOCK_TYPES: [ "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: False
+      DECODER:
+        NAME: CogVideoXDecoder3D
+        IN_CHANNELS: 16
+        OUT_CHANNELS: 3
+        UP_BLOCK_TYPES: [ "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: False
+    #
+    COND_STAGE_MODEL:
+      NAME: T5EmbedderHF
+      PRETRAINED_MODEL: ms://AI-ModelScope/t5-v1_1-xxl
+      TOKENIZER_PATH: ms://AI-ModelScope/t5-v1_1-xxl
+      LENGTH: 226
+      CLEAN:
+      USE_GRAD: False
+    #
+    LOSS:
+      NAME: ReconstructLoss
+      LOSS_TYPE: l2
+  #
+  SAMPLE_ARGS:
+    SAMPLER: ddim
+    SAMPLE_STEPS: 50
+    SEED: 42
+    GUIDE_SCALE: 6.0
+    GUIDE_RESCALE: 0.0
+    NUM_FRAMES: 49
+  #
+  OPTIMIZER:
+    NAME: Adam
+    LEARNING_RATE: 1e-3
+    BETAS: [ 0.9, 0.95 ]
+    EPS: 1e-8
+    WEIGHT_DECAY: 0.0
+    AMSGRAD: False
+  #
+#  LR_SCHEDULER:
+#    NAME: StepAnnealingLR
+#    WARMUP_STEPS: 200
+#    TOTAL_STEPS: 2000
+#    DECAY_MODE: 'cosine'
+  #
+  TRAIN_DATA:
+    NAME: VideoGenDataset
+    MODE: train
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    NUM_WORKERS: 4
+    NUM_FRAMES: 49
+    FPS: 8
+    HEIGHT: 480
+    WIDTH: 720
+    PROMPT_PREFIX: 'DISNEY '
+    SAMPLER:
+      NAME: MixtureOfSamplers
+      SUB_SAMPLERS:
+        - NAME: MultiLevelBatchSampler
+          PROB: 1.0
+          FIELDS: [ "video_path", "prompt" ]
+          DELIMITER: '#;#'
+          PATH_PREFIX: cache/datasets/Disney-VideoGeneration-Dataset/
+          INDEX_FILE: cache/datasets/Disney-VideoGeneration-Dataset/index.jsonl
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ 'video', "prompt" ]
+        META_KEYS: [ ]
+  #
+  EVAL_DATA:
+    NAME: Text2ImageDataset
+    MODE: eval
+    PROMPT_FILE:
+    PROMPT_DATA: [ "A girl riding a bike.", "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical atmosphere of this unique musical performance." ]
+    IMAGE_SIZE: [ 480, 720 ]
+    FIELDS: [ "prompt" ]
+    DELIMITER: '#;#'
+    PROMPT_PREFIX: 'DISNEY '
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    USE_NUM: 8
+    NUM_WORKERS: 4
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ 'index', 'prompt' ]
+        META_KEYS: [ 'image_size' ]
+  #
+  TRAIN_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0
+    - NAME: BackwardHook
+      PRIORITY: 10
+    - NAME: LogHook
+      LOG_INTERVAL: 10
+      PRIORITY: 20
+    - NAME: CheckpointHook
+      INTERVAL: 1000
+      PRIORITY: 40
+  #
+  EVAL_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0

scepter/methods/examples/generation/dit_cogvideox_5b_i2v_lora.yaml

@@ -0,0 +1,270 @@
+ENV:
+  BACKEND: nccl
+  SEED: 42
+  TENSOR_PARALLEL_SIZE: 1
+  PIPELINE_PARALLEL_SIZE: 1
+  SYS_ENVS:
+    TORCH_CUDNN_V8_API_ENABLED: '1'
+    TOKENIZERS_PARALLELISM: 'false'
+    TF_CPP_MIN_LOG_LEVEL: '3'
+    PYTORCH_CUDA_ALLOC_CONF: 'expandable_segments:True'
+#
+SOLVER:
+  NAME: LatentDiffusionVideoSolver
+  MAX_STEPS: 2000
+  USE_AMP: True
+  DTYPE: bfloat16
+  USE_FAIRSCALE: False
+  USE_FSDP: True
+  LOAD_MODEL_ONLY: False
+  ENABLE_GRADSCALER: False
+  USE_SCALER: False
+  RESUME_FROM:
+  WORK_DIR: ./cache/save_data/dit_cogvideox_5b_i2v_lora
+  LOG_FILE: std_log.txt
+  EVAL_INTERVAL: 100
+  LOG_TRAIN_NUM: 4
+  FPS: 8
+  SHARDING_STRATEGY: full_shard
+  FSDP_REDUCE_DTYPE: float32
+  FSDP_BUFFER_DTYPE: float32
+  FSDP_SHARD_MODULES: [ 'model', 'cond_stage_model.model']
+  SAVE_MODULES: [ 'model', 'cond_stage_model.model']
+  TRAIN_MODULES: ['model']
+  #
+  FILE_SYSTEM:
+    NAME: "ModelscopeFs"
+    TEMP_DIR: "./cache/cache_data"
+  #
+  TUNER:
+    - NAME: SwiftLoRA
+      R: 64
+      LORA_ALPHA: 64
+      LORA_DROPOUT: 0.0
+      BIAS: "none"
+      TARGET_MODULES: "model.*(.to_k|.to_q|.to_v|.to_out.0)$"
+  #
+  MODEL:
+    NAME: LatentDiffusionCogVideoX
+    PRETRAINED_MODEL:
+    PARAMETERIZATION: v
+    TIMESTEPS: 1000
+    MIN_SNR_GAMMA: 3.0
+    ZERO_TERMINAL_SNR: True
+    SCALE_FACTOR_SPATIAL: 8
+    SCALE_FACTOR_TEMPORAL: 4
+    SCALING_FACTOR_IMAGE: 0.7  # 5b diff
+    NOISED_IMAGE_DROPOUT: 0.05
+    IGNORE_KEYS: [ ]
+    DEFAULT_N_PROMPT:
+    USE_EMA: False
+    EVAL_EMA: False
+    DIFFUSION:
+      NAME: BaseDiffusion
+      PREDICTION_TYPE: v
+      NOISE_SCHEDULER:
+        NAME: ScaledLinearScheduler
+        BETA_MIN: 0.00085
+        BETA_MAX: 0.012
+        SNR_SHIFT_SCALE: 1.0  # 5b diff
+        RESCALE_BETAS_ZERO_SNR: True
+      DIFFUSION_SAMPLERS:
+        NAME: DDIMSampler
+        DISCRETIZATION_TYPE: trailing
+        ETA: 0.0
+    #
+    DIFFUSION_MODEL:
+      NAME: CogVideoXTransformer3DModel
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: # 5b-I2V diff
+        - ms://AI-ModelScope/CogVideoX-5b-I2V@transformer/diffusion_pytorch_model-00001-of-00003.safetensors
+        - ms://AI-ModelScope/CogVideoX-5b-I2V@transformer/diffusion_pytorch_model-00002-of-00003.safetensors
+        - ms://AI-ModelScope/CogVideoX-5b-I2V@transformer/diffusion_pytorch_model-00003-of-00003.safetensors
+      NUM_ATTENTION_HEADS: 48  # 5b diff
+      ATTENTION_HEAD_DIM: 64
+      IN_CHANNELS: 32  # 5b-I2V diff
+      LATENT_CHANNELS: 16
+      OUT_CHANNELS: 16
+      FLIP_SIN_TO_COS: True
+      FREQ_SHIFT: 0
+      TIME_EMBED_DIM: 512
+      TEXT_EMBED_DIM: 4096
+      NUM_LAYERS: 42  # 5b diff
+      DROPOUT: 0.0
+      ATTENTION_BIAS: True
+      SAMPLE_WIDTH: 90
+      SAMPLE_HEIGHT: 60
+      SAMPLE_FRAMES: 49
+      PATCH_SIZE: 2
+      TEMPORAL_COMPRESSION_RATIO: 4
+      MAX_TEXT_SEQ_LENGTH: 226
+      ACTIVATION_FN: "gelu-approximate"
+      TIMESTEP_ACTIVATION_FN: "silu"
+      NORM_ELEMENTWISE_AFFINE: True
+      NORM_EPS: 1e-5
+      SPATIAL_INTERPOLATION_SCALE: 1.875
+      TEMPORAL_INTERPOLATION_SCALE: 1.0
+      USE_ROTARY_POSITIONAL_EMBEDDINGS: True # 5b diff
+      USE_LEARNED_POSITIONAL_EMBEDDINGS: True  # 5b-I2V diff
+      GRADIENT_CHECKPOINTING: True
+    #
+    FIRST_STAGE_MODEL:
+      NAME: AutoencoderKLCogVideoX
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: ms://AI-ModelScope/CogVideoX-5b-I2V@vae/diffusion_pytorch_model.safetensors # 5b diff
+      SAMPLE_HEIGHT: 480
+      SAMPLE_WIDTH: 720
+      USE_QUANT_CONV: False
+      USE_POST_QUANT_CONV: False
+      USE_SLICING: True
+      USE_TILING: True
+      GRADIENT_CHECKPOINTING: True
+      ENCODER:
+        NAME: CogVideoXEncoder3D
+        IN_CHANNELS: 3
+        OUT_CHANNELS: 16
+        UP_BLOCK_TYPES: [ "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+      DECODER:
+        NAME: CogVideoXDecoder3D
+        IN_CHANNELS: 16
+        OUT_CHANNELS: 3
+        UP_BLOCK_TYPES: [ "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+    #
+    COND_STAGE_MODEL:
+      NAME: T5EmbedderHF
+      PRETRAINED_MODEL: ms://AI-ModelScope/t5-v1_1-xxl
+      TOKENIZER_PATH: ms://AI-ModelScope/t5-v1_1-xxl
+      LENGTH: 226
+      CLEAN:
+      USE_GRAD: False
+    #
+    LOSS:
+      NAME: ReconstructLoss
+      LOSS_TYPE: l2
+  #
+  SAMPLE_ARGS:
+    SAMPLER: ddim
+    SAMPLE_STEPS: 50
+    SEED: 42
+    GUIDE_SCALE: 6.0
+    GUIDE_RESCALE: 0.0
+    NUM_FRAMES: 49
+  #
+  OPTIMIZER:
+    NAME: Adam
+    LEARNING_RATE: 1e-3
+    BETAS: [ 0.9, 0.95 ]
+    EPS: 1e-8
+    WEIGHT_DECAY: 0.0
+    AMSGRAD: False
+  #
+#  LR_SCHEDULER:
+#    NAME: StepAnnealingLR
+#    WARMUP_STEPS: 200
+#    TOTAL_STEPS: 2000
+#    DECAY_MODE: 'cosine'
+  #
+  TRAIN_DATA:
+    NAME: VideoGenDataset
+    MODE: train
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    NUM_WORKERS: 0
+    NUM_FRAMES: 49
+    FPS: 8
+    HEIGHT: 480
+    WIDTH: 720
+    PROMPT_PREFIX: 'DISNEY '
+    DATA_TYPE: 'i2v'
+    SAMPLER:
+      NAME: MixtureOfSamplers
+      SUB_SAMPLERS:
+        - NAME: MultiLevelBatchSampler
+          PROB: 1.0
+          FIELDS: [ "video_path", "prompt" ]
+          DELIMITER: '#;#'
+          PATH_PREFIX: cache/datasets/Disney-VideoGeneration-Dataset/
+          INDEX_FILE: cache/datasets/Disney-VideoGeneration-Dataset/index.jsonl
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ "video", "image", "prompt" ]
+        META_KEYS: [ ]
+  #
+#  EVAL_DATA:
+#    NAME: Text2ImageDataset
+#    MODE: eval
+#    PROMPT_FILE:
+#    PROMPT_DATA: [ "A cat running.#;#asset/images/edit_tuner/cat_512.jpg" ]
+#    FIELDS: [ "prompt", "img_path" ]
+#    DELIMITER: '#;#'
+#    PROMPT_PREFIX: ''
+#    PIN_MEMORY: True
+#    BATCH_SIZE: 1
+#    USE_NUM: 8
+#    NUM_WORKERS: 0
+#    IMAGE_SIZE: [ 480, 720 ]
+#    TRANSFORMS:
+#      - NAME: LoadImageFromFileList
+#        FILE_KEYS: [ 'img_path' ]
+#        RGB_ORDER: RGB
+#        BACKEND: pillow
+#      - NAME: FlexibleResize
+#        INTERPOLATION: bilinear
+#        SIZE: [ 480, 720 ]
+#        INPUT_KEY: [ 'img' ]
+#        OUTPUT_KEY: [ 'img' ]
+#        BACKEND: pillow
+#      - NAME: FlexibleCenterCrop
+#        SIZE: [ 480, 720 ]
+#        INPUT_KEY: [ 'img' ]
+#        OUTPUT_KEY: [ 'img' ]
+#        BACKEND: pillow
+#      - NAME: ImageToTensor
+#        INPUT_KEY: [ 'img' ]
+#        OUTPUT_KEY: [ 'img' ]
+#        BACKEND: pillow
+#      - NAME: Normalize
+#        MEAN: [ 0.5,  0.5,  0.5 ]
+#        STD: [ 0.5,  0.5,  0.5 ]
+#        INPUT_KEY: [ 'img' ]
+#        OUTPUT_KEY: [ 'image' ]
+#        BACKEND: torchvision
+#      - NAME: Select
+#        KEYS: [ 'image', 'prompt' ]
+#        META_KEYS: [ 'image_size' ]
+  #
+  TRAIN_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0
+    - NAME: BackwardHook
+      PRIORITY: 10
+    - NAME: LogHook
+      LOG_INTERVAL: 10
+      PRIORITY: 20
+    - NAME: CheckpointHook
+      INTERVAL: 1000
+      PRIORITY: 40
+  #
+#  EVAL_HOOKS:
+#    - NAME: ProbeDataHook
+#      PROB_INTERVAL: 100
+#      PRIORITY: 0

scepter/methods/examples/generation/dit_cogvideox_5b_lora.yaml

@@ -0,0 +1,277 @@
+ENV:
+  BACKEND: nccl
+  SEED: 42
+  TENSOR_PARALLEL_SIZE: 1
+  PIPELINE_PARALLEL_SIZE: 1
+  SYS_ENVS:
+    TORCH_CUDNN_V8_API_ENABLED: '1'
+    TOKENIZERS_PARALLELISM: 'false'
+    TF_CPP_MIN_LOG_LEVEL: '3'
+    PYTORCH_CUDA_ALLOC_CONF: 'expandable_segments:True'
+#
+SOLVER:
+  NAME: LatentDiffusionVideoSolver
+  MAX_STEPS: 2000
+  USE_AMP: True
+  DTYPE: bfloat16
+  USE_FAIRSCALE: False
+  USE_FSDP: True
+  LOAD_MODEL_ONLY: False
+  ENABLE_GRADSCALER: False
+  USE_SCALER: False
+  RESUME_FROM:
+  WORK_DIR: ./cache/save_data/dit_cogvideox_5b_lora
+  LOG_FILE: std_log.txt
+  EVAL_INTERVAL: 100
+  LOG_TRAIN_NUM: 4
+  FPS: 8
+  SHARDING_STRATEGY: full_shard
+  FSDP_REDUCE_DTYPE: float32
+  FSDP_BUFFER_DTYPE: float32
+  FSDP_SHARD_MODULES: [ 'model', 'cond_stage_model.model']
+  SAVE_MODULES: [ 'model', 'cond_stage_model.model']
+  TRAIN_MODULES: ['model']
+  #
+  FILE_SYSTEM:
+    NAME: "ModelscopeFs"
+    TEMP_DIR: "./cache/cache_data"
+  #
+  TUNER:
+    - NAME: SwiftLoRA
+      R: 64
+      LORA_ALPHA: 64
+      LORA_DROPOUT: 0.0
+      BIAS: "none"
+      TARGET_MODULES: "model.*(.to_k|.to_q|.to_v|.to_out.0)$"
+  #
+  MODEL:
+    NAME: LatentDiffusionCogVideoX
+    PRETRAINED_MODEL:
+    PARAMETERIZATION: v
+    TIMESTEPS: 1000
+    MIN_SNR_GAMMA: 3.0
+    ZERO_TERMINAL_SNR: True
+    SCALE_FACTOR_SPATIAL: 8
+    SCALE_FACTOR_TEMPORAL: 4
+    SCALING_FACTOR_IMAGE: 0.7  # 5b diff
+    IGNORE_KEYS: [ ]
+    DEFAULT_N_PROMPT:
+    USE_EMA: False
+    EVAL_EMA: False
+    DIFFUSION:
+      NAME: BaseDiffusion
+      PREDICTION_TYPE: v
+      NOISE_SCHEDULER:
+        NAME: ScaledLinearScheduler
+        BETA_MIN: 0.00085
+        BETA_MAX: 0.012
+        SNR_SHIFT_SCALE: 1.0  # 5b diff
+        RESCALE_BETAS_ZERO_SNR: True
+      DIFFUSION_SAMPLERS:
+        NAME: DDIMSampler
+        DISCRETIZATION_TYPE: trailing
+        ETA: 0.0
+    #
+    DIFFUSION_MODEL:
+      NAME: CogVideoXTransformer3DModel
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: # 5b diff
+        - ms://AI-ModelScope/CogVideoX-5b@transformer/diffusion_pytorch_model-00001-of-00002.safetensors
+        - ms://AI-ModelScope/CogVideoX-5b@transformer/diffusion_pytorch_model-00002-of-00002.safetensors
+      NUM_ATTENTION_HEADS: 48  # 5b diff
+      ATTENTION_HEAD_DIM: 64
+      IN_CHANNELS: 16
+      OUT_CHANNELS: 16
+      FLIP_SIN_TO_COS: True
+      FREQ_SHIFT: 0
+      TIME_EMBED_DIM: 512
+      TEXT_EMBED_DIM: 4096
+      NUM_LAYERS: 42  # 5b diff
+      DROPOUT: 0.0
+      ATTENTION_BIAS: True
+      SAMPLE_WIDTH: 90
+      SAMPLE_HEIGHT: 60
+      SAMPLE_FRAMES: 49
+      PATCH_SIZE: 2
+      TEMPORAL_COMPRESSION_RATIO: 4
+      MAX_TEXT_SEQ_LENGTH: 226
+      ACTIVATION_FN: "gelu-approximate"
+      TIMESTEP_ACTIVATION_FN: "silu"
+      NORM_ELEMENTWISE_AFFINE: True
+      NORM_EPS: 1e-5
+      SPATIAL_INTERPOLATION_SCALE: 1.875
+      TEMPORAL_INTERPOLATION_SCALE: 1.0
+      USE_ROTARY_POSITIONAL_EMBEDDINGS: True # 5b diff
+      USE_LEARNED_POSITIONAL_EMBEDDINGS: False
+      GRADIENT_CHECKPOINTING: True
+    #
+    FIRST_STAGE_MODEL:
+      NAME: AutoencoderKLCogVideoX
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: ms://AI-ModelScope/CogVideoX-5b@vae/diffusion_pytorch_model.safetensors # 5b diff
+      SAMPLE_HEIGHT: 480
+      SAMPLE_WIDTH: 720
+      USE_QUANT_CONV: False
+      USE_POST_QUANT_CONV: False
+      USE_SLICING: True
+      USE_TILING: True
+      GRADIENT_CHECKPOINTING: True
+      ENCODER:
+        NAME: CogVideoXEncoder3D
+        IN_CHANNELS: 3
+        OUT_CHANNELS: 16
+        UP_BLOCK_TYPES: [ "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+      DECODER:
+        NAME: CogVideoXDecoder3D
+        IN_CHANNELS: 16
+        OUT_CHANNELS: 3
+        UP_BLOCK_TYPES: [ "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+    #
+    COND_STAGE_MODEL:
+      NAME: T5EmbedderHF
+      PRETRAINED_MODEL: ms://AI-ModelScope/t5-v1_1-xxl
+      TOKENIZER_PATH: ms://AI-ModelScope/t5-v1_1-xxl
+      LENGTH: 226
+      CLEAN:
+      USE_GRAD: False
+    #
+    LOSS:
+      NAME: ReconstructLoss
+      LOSS_TYPE: l2
+  #
+  SAMPLE_ARGS:
+    SAMPLER: ddim
+    SAMPLE_STEPS: 50
+    SEED: 42
+    GUIDE_SCALE: 6.0
+    GUIDE_RESCALE: 0.0
+    NUM_FRAMES: 49
+  #
+  OPTIMIZER:
+    NAME: Adam
+    LEARNING_RATE: 1e-3
+    BETAS: [ 0.9, 0.95 ]
+    EPS: 1e-8
+    WEIGHT_DECAY: 0.0
+    AMSGRAD: False
+  #
+#  LR_SCHEDULER:
+#    NAME: StepAnnealingLR
+#    WARMUP_STEPS: 200
+#    TOTAL_STEPS: 2000
+#    DECAY_MODE: 'cosine'
+  #
+  TRAIN_DATA:
+    NAME: VideoGenDatasetOTF
+    MODE: train
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    NUM_WORKERS: 4
+    NUM_FRAMES: 49
+    FPS: 8
+    HEIGHT: 480
+    WIDTH: 720
+    PROMPT_PREFIX: 'DISNEY '
+    DELIMITER: '#;#'
+    FIELDS: [ 'video_path', 'prompt' ]
+    PATH_PREFIX: cache/datasets/Disney-VideoGeneration-Dataset/
+    DATA_FILE: cache/datasets/Disney-VideoGeneration-Dataset/index.jsonl
+    SAMPLER:
+      NAME: LoopSampler
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ 'video', 'video_latent', "prompt" ]
+        META_KEYS: [ ]
+    MODEL:
+      NAME: AutoencoderKLCogVideoX
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: ms://AI-ModelScope/CogVideoX-5b@vae/diffusion_pytorch_model.safetensors
+      SAMPLE_HEIGHT: 480
+      SAMPLE_WIDTH: 720
+      USE_QUANT_CONV: False
+      USE_POST_QUANT_CONV: False
+      USE_SLICING: True
+      USE_TILING: True
+      GRADIENT_CHECKPOINTING: True
+      ENCODER:
+        NAME: CogVideoXEncoder3D
+        IN_CHANNELS: 3
+        OUT_CHANNELS: 16
+        UP_BLOCK_TYPES: [ "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+      DECODER:
+        NAME: CogVideoXDecoder3D
+        IN_CHANNELS: 16
+        OUT_CHANNELS: 3
+        UP_BLOCK_TYPES: [ "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+  #
+  EVAL_DATA:
+    NAME: Text2ImageDataset
+    MODE: eval
+    PROMPT_FILE:
+    PROMPT_DATA: [ "A girl riding a bike.", "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical atmosphere of this unique musical performance." ]
+    IMAGE_SIZE: [ 480, 720 ]
+    FIELDS: [ "prompt" ]
+    DELIMITER: '#;#'
+    PROMPT_PREFIX: 'DISNEY '
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    USE_NUM: 8
+    NUM_WORKERS: 4
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ 'index', 'prompt' ]
+        META_KEYS: [ 'image_size' ]
+  #
+  TRAIN_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0
+    - NAME: BackwardHook
+      PRIORITY: 10
+    - NAME: LogHook
+      LOG_INTERVAL: 10
+      PRIORITY: 20
+    - NAME: CheckpointHook
+      INTERVAL: 1000
+      PRIORITY: 40
+  #
+  EVAL_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0

scepter/methods/examples/generation/dit_flux1.0_dev_1024_lora.yaml

@@ -0,0 +1,246 @@
+ENV:
+  BACKEND: nccl
+  SEED: 166666
+SOLVER:
+  NAME: LatentDiffusionSolver
+  MAX_STEPS: 100000
+  USE_AMP: True
+  DTYPE: bfloat16
+  USE_FAIRSCALE: False
+  USE_FSDP: True
+  LOAD_MODEL_ONLY: False
+  ENABLE_GRADSCALER: False
+  USE_SCALER: False
+  RESUME_FROM:
+  WORK_DIR: ./cache/save_data/dit_flux_dev_1024_lora
+  LOG_FILE: std_log.txt
+  EVAL_INTERVAL: 100
+  LOG_TRAIN_NUM: 16
+  FSDP_REDUCE_DTYPE: float32
+  FSDP_BUFFER_DTYPE: float32
+  FSDP_SHARD_MODULES: [ 'model', 'cond_stage_model.t5_model' ] #
+  SAVE_MODULES: [ 'model'] #
+  TRAIN_MODULES: ['model']
+  #
+  FILE_SYSTEM:
+    NAME: "ModelscopeFs"
+    TEMP_DIR: "./cache/cache_data"
+  #
+  FREEZE:
+  TUNER:
+    - NAME: SwiftLoRA
+      R: 4
+      LORA_ALPHA: 4
+      LORA_DROPOUT: 0.0
+      BIAS: "none"
+      TARGET_MODULES: "(model.double_blocks.*(.qkv|.proj|.img_mod.lin|.txt_mod.lin))|(model.single_blocks.*(.linear1|.linear2|.modulation.lin))$"
+  ##
+  MODEL:
+    NAME: LatentDiffusionFlux
+    PARAMETERIZATION: rf
+    TIMESTEPS: 1000
+    MIN_SNR_GAMMA:
+    ZERO_TERMINAL_SNR: False
+    PRETRAINED_MODEL:
+    IGNORE_KEYS: [ ]
+    DEFAULT_N_PROMPT:
+    USE_EMA: False
+    EVAL_EMA: False
+    DIFFUSION:
+      NAME: DiffusionFluxRF
+      PREDICTION_TYPE: raw
+      NOISE_SCHEDULER:
+        NAME: FlowMatchSigmaScheduler
+        WEIGHTING_SCHEME: logit_normal
+        SHIFT: 3.0
+        LOGIT_MEAN: 0.0
+        LOGIT_STD: 1.0
+        MODE_SCALE: 1.29
+      SAMPLER_SCHEDULER:
+        NAME: FlowMatchFluxShiftScheduler
+        SHIFT: False
+        SIGMOID_SCALE: 1
+        BASE_SHIFT: 0.5
+        MAX_SHIFT: 1.15
+        #
+    DIFFUSION_MODEL:
+      NAME: Flux
+      PRETRAINED_MODEL: ms://AI-ModelScope/FLUX.1-dev@flux1-dev.safetensors
+      IN_CHANNELS: 64
+      HIDDEN_SIZE: 3072
+      NUM_HEADS: 24
+      AXES_DIM: [ 16, 56, 56 ]
+      THETA: 10000
+      VEC_IN_DIM: 768
+      GUIDANCE_EMBED: True
+      CONTEXT_IN_DIM: 4096
+      MLP_RATIO: 4.0
+      QKV_BIAS: True
+      DEPTH: 19
+      DEPTH_SINGLE_BLOCKS: 38
+      USE_GRAD_CHECKPOINT: True
+
+    #
+    FIRST_STAGE_MODEL:
+      NAME: AutoencoderKLFlux
+      EMBED_DIM: 16
+      PRETRAINED_MODEL: ms://AI-ModelScope/FLUX.1-dev@ae.safetensors
+      IGNORE_KEYS: [ ]
+      BATCH_SIZE: 8
+      USE_CONV: False
+      SCALE_FACTOR: 0.3611
+      SHIFT_FACTOR: 0.1159
+      #
+      ENCODER:
+        NAME: Encoder
+        USE_CHECKPOINT: True
+        CH: 128
+        OUT_CH: 3
+        NUM_RES_BLOCKS: 2
+        IN_CHANNELS: 3
+        ATTN_RESOLUTIONS: [ ]
+        CH_MULT: [ 1, 2, 4, 4 ]
+        Z_CHANNELS: 16
+        DOUBLE_Z: True
+        DROPOUT: 0.0
+        RESAMP_WITH_CONV: True
+      #
+      DECODER:
+        NAME: Decoder
+        USE_CHECKPOINT: True
+        CH: 128
+        OUT_CH: 3
+        NUM_RES_BLOCKS: 2
+        IN_CHANNELS: 3
+        ATTN_RESOLUTIONS: [ ]
+        CH_MULT: [ 1, 2, 4, 4 ]
+        Z_CHANNELS: 16
+        DROPOUT: 0.0
+        RESAMP_WITH_CONV: True
+        GIVE_PRE_END: False
+        TANH_OUT: False
+    #
+    COND_STAGE_MODEL:
+      NAME: T5PlusClipFluxEmbedder
+      T5_MODEL:
+        NAME: HFEmbedder
+        HF_MODEL_CLS: T5EncoderModel
+        MODEL_PATH: ms://AI-ModelScope/FLUX.1-dev@text_encoder_2/
+        HF_TOKENIZER_CLS: T5Tokenizer
+        TOKENIZER_PATH: ms://AI-ModelScope/FLUX.1-dev@tokenizer_2/
+        MAX_LENGTH: 512
+        OUTPUT_KEY: last_hidden_state
+        D_TYPE: bfloat16
+        BATCH_INFER: False
+        CLEAN: whitespace
+      CLIP_MODEL:
+        NAME: HFEmbedder
+        HF_MODEL_CLS: CLIPTextModel
+        MODEL_PATH: ms://AI-ModelScope/FLUX.1-dev@text_encoder/
+        HF_TOKENIZER_CLS: CLIPTokenizer
+        TOKENIZER_PATH: ms://AI-ModelScope/FLUX.1-dev@tokenizer/
+        MAX_LENGTH: 77
+        OUTPUT_KEY: pooler_output
+        D_TYPE: bfloat16
+        BATCH_INFER: True
+        CLEAN: whitespace
+        USE_GRAD_CHECKPOINT: True
+  #
+  SAMPLE_ARGS:
+    SAMPLE_STEPS: 50
+    SAMPLER: flow_euler
+    SEED: 2024
+    IMAGE_SIZE: [ 1024, 1024 ]
+    GUIDE_SCALE: 3.5
+  #
+  OPTIMIZER:
+    NAME: AdamW
+    LEARNING_RATE: 4e-4
+    BETAS: [ 0.9, 0.999 ]
+    EPS: 1e-8
+    WEIGHT_DECAY: 1e-2
+    AMSGRAD: False
+  #
+  TRAIN_DATA:
+    NAME: ImageTextPairMSDataset
+    MODE: train
+    MS_DATASET_NAME: style_custom_dataset
+    MS_DATASET_NAMESPACE: damo
+    MS_DATASET_SUBNAME: 3D
+    PROMPT_PREFIX: ""
+    MS_DATASET_SPLIT: train
+    MS_REMAP_KEYS: { 'Image:FILE': 'Target:FILE' }
+    REPLACE_STYLE: False
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    NUM_WORKERS: 4
+    SAMPLER:
+      NAME: LoopSampler
+    TRANSFORMS:
+      - NAME: LoadImageFromFile
+        RGB_ORDER: RGB
+        BACKEND: pillow
+      - NAME: FlexibleResize
+        INTERPOLATION: bilinear
+        SIZE: [ 1024, 1024 ]
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'img' ]
+        BACKEND: pillow
+      - NAME: FlexibleCenterCrop
+        SIZE: [ 1024, 1024 ]
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'img' ]
+        BACKEND: pillow
+      - NAME: ImageToTensor
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'img' ]
+        BACKEND: pillow
+      - NAME: Normalize
+        MEAN: [ 0.5,  0.5,  0.5 ]
+        STD: [ 0.5,  0.5,  0.5 ]
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'image' ]
+        BACKEND: torchvision
+      - NAME: Select
+        KEYS: [ 'image', 'prompt' ]
+        META_KEYS: [ 'data_key' ]
+  #
+  EVAL_DATA:
+    NAME: Text2ImageDataset
+    MODE: eval
+    PROMPT_FILE:
+    PROMPT_DATA: [ "a boy wearing a jacket", "a dog running on the lawn" ]
+    IMAGE_SIZE: [ 1024, 1024 ]
+    FIELDS: [ "prompt" ]
+    DELIMITER: '#;#'
+    PROMPT_PREFIX: ''
+    PIN_MEMORY: True
+    BATCH_SIZE: 2
+    NUM_WORKERS: 4
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ 'index', 'prompt' ]
+        META_KEYS: [ 'image_size' ]
+  #
+  TRAIN_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0
+    - NAME: BackwardHook
+#      GRADIENT_CLIP: 1.0
+      PRIORITY: 10
+    - NAME: LogHook
+      LOG_INTERVAL: 10
+    -
+      NAME: TensorboardLogHook
+    -
+      NAME: CheckpointHook
+      INTERVAL: 10000
+      PRIORITY: 200
+      SAVE_LAST: True
+      SAVE_NAME_PREFIX: 'step'
+      DISABLE_SNAPSHOT: True
+  EVAL_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0

scepter/methods/examples/generation/dit_flux1.0_schnell_1024_lora.yaml

@@ -0,0 +1,267 @@
+ENV:
+  BACKEND: nccl
+  SEED: 166666
+SOLVER:
+  NAME: LatentDiffusionSolver
+  MAX_STEPS: 100000
+  USE_AMP: True
+  DTYPE: bfloat16
+  USE_FAIRSCALE: False
+  USE_FSDP: True
+  LOAD_MODEL_ONLY: False
+  ENABLE_GRADSCALER: False
+  USE_SCALER: False
+  RESUME_FROM:
+  WORK_DIR: ./cache/save_data/dit_flux_schnell_1024_lora
+  LOG_FILE: std_log.txt
+  EVAL_INTERVAL: 100
+  LOG_TRAIN_NUM: 16
+  FSDP_REDUCE_DTYPE: float32
+  FSDP_BUFFER_DTYPE: float32
+  FSDP_SHARD_MODULES: [ 'model', 'cond_stage_model.t5_model' ] #
+  SAVE_MODULES: [ 'model']
+  TRAIN_MODULES: ['model']
+  #
+  FILE_SYSTEM:
+    NAME: "ModelscopeFs"
+    TEMP_DIR: "./cache/cache_data"
+  #
+  FREEZE:
+  TUNER:
+    - NAME: SwiftLoRA
+      R: 4
+      LORA_ALPHA: 4
+      LORA_DROPOUT: 0.0
+      BIAS: "none"
+      TARGET_MODULES: "(model.double_blocks.*(.qkv|.proj|.img_mod.lin|.txt_mod.lin))|(model.single_blocks.*(.linear1|.linear2|.modulation.lin))$"
+  ##
+  MODEL:
+    NAME: LatentDiffusionFlux
+    PARAMETERIZATION: rf
+    TIMESTEPS: 1000
+    MIN_SNR_GAMMA:
+    ZERO_TERMINAL_SNR: False
+    PRETRAINED_MODEL:
+    IGNORE_KEYS: [ ]
+    DEFAULT_N_PROMPT:
+    USE_EMA: False
+    EVAL_EMA: False
+    DIFFUSION:
+      NAME: DiffusionFluxRF
+      PREDICTION_TYPE: raw
+      NOISE_SCHEDULER:
+        NAME: FlowMatchSigmaScheduler
+        # WEIGHTING_SCHEME DESCRIPTION: The weighting scheme for sampling timesteps, choose from ['sigma_sqrt', 'logit_normal', 'mode', 'cosmap', 'none']. TYPE: str default: 'logit_normal'
+        WEIGHTING_SCHEME: logit_normal
+        SHIFT: 3.0
+        # LOGIT_MEAN DESCRIPTION: The mean of the logit distribution for sampling timesteps. TYPE: float default: 0.0
+        LOGIT_MEAN: 0.0
+        # LOGIT_STD DESCRIPTION: The standard deviation of the logit distribution for sampling timesteps. TYPE: float default: 1.0
+        LOGIT_STD: 1.0
+        # MODE_SCALE DESCRIPTION: The scale factor for the mode of the logit distribution for sampling timesteps. TYPE: float default: 1.29
+        MODE_SCALE: 1.29
+      SAMPLER_SCHEDULER:
+        # NAME DESCRIPTION:  TYPE:  default: 'FlowMatchFluxShiftScheduler'
+        NAME: FlowMatchFluxShiftScheduler
+        # SHIFT DESCRIPTION: Use timestamp shift or not, default is True. TYPE: bool default: True
+        SHIFT: False
+        # SIGMOID_SCALE DESCRIPTION: The scale of sigmoid function for sampling timesteps. TYPE: int default: 1
+        SIGMOID_SCALE: 1
+        # BASE_SHIFT DESCRIPTION: The base shift factor for the timestamp. TYPE: float default: 0.5
+        BASE_SHIFT: 0.5
+        # MAX_SHIFT DESCRIPTION: The max shift factor for the timestamp. TYPE: float default: 1.15
+        MAX_SHIFT: 1.15
+        #
+    DIFFUSION_MODEL:
+      # NAME DESCRIPTION:  TYPE:  default: 'Flux'
+      NAME: Flux
+      PRETRAINED_MODEL: ms://AI-ModelScope/FLUX.1-schnell@flux1-schnell.safetensors
+      # IN_CHANNELS DESCRIPTION: model's input channels. TYPE: int default: 64
+      IN_CHANNELS: 64
+      # HIDDEN_SIZE DESCRIPTION: model's hidden size. TYPE: int default: 1024
+      HIDDEN_SIZE: 3072
+      # NUM_HEADS DESCRIPTION: number of heads in the transformer. TYPE: int default: 16
+      NUM_HEADS: 24
+      # AXES_DIM DESCRIPTION: dimensions of the axes of the positional encoding. TYPE: list default: [16, 56, 56]
+      AXES_DIM: [ 16, 56, 56 ]
+      # THETA DESCRIPTION: theta for positional encoding. TYPE: int default: 10000
+      THETA: 10000
+      # VEC_IN_DIM DESCRIPTION: dimension of the vector input. TYPE: int default: 768
+      VEC_IN_DIM: 768
+      # GUIDANCE_EMBED DESCRIPTION: whether to use guidance embedding. TYPE: bool default: False
+      GUIDANCE_EMBED: False
+      # CONTEXT_IN_DIM DESCRIPTION: dimension of the context input. TYPE: int default: 4096
+      CONTEXT_IN_DIM: 4096
+      # MLP_RATIO DESCRIPTION: ratio of mlp hidden size to hidden size. TYPE: float default: 4.0
+      MLP_RATIO: 4.0
+      # QKV_BIAS DESCRIPTION: whether to use bias in qkv projection. TYPE: bool default: True
+      QKV_BIAS: True
+      # DEPTH DESCRIPTION: number of transformer blocks. TYPE: int default: 19
+      DEPTH: 19
+      # DEPTH_SINGLE_BLOCKS DESCRIPTION: number of transformer blocks in the single stream block. TYPE: int default: 38
+      DEPTH_SINGLE_BLOCKS: 38
+      USE_GRAD_CHECKPOINT: True
+
+    #
+    FIRST_STAGE_MODEL:
+      NAME: AutoencoderKLFlux
+      EMBED_DIM: 16
+      PRETRAINED_MODEL: ms://AI-ModelScope/FLUX.1-schnell@ae.safetensors
+      IGNORE_KEYS: [ ]
+      BATCH_SIZE: 8
+      USE_CONV: False
+      SCALE_FACTOR: 0.3611
+      SHIFT_FACTOR: 0.1159
+      #
+      ENCODER:
+        NAME: Encoder
+        USE_CHECKPOINT: True
+        CH: 128
+        OUT_CH: 3
+        NUM_RES_BLOCKS: 2
+        IN_CHANNELS: 3
+        ATTN_RESOLUTIONS: [ ]
+        CH_MULT: [ 1, 2, 4, 4 ]
+        Z_CHANNELS: 16
+        DOUBLE_Z: True
+        DROPOUT: 0.0
+        RESAMP_WITH_CONV: True
+      #
+      DECODER:
+        NAME: Decoder
+        USE_CHECKPOINT: True
+        CH: 128
+        OUT_CH: 3
+        NUM_RES_BLOCKS: 2
+        IN_CHANNELS: 3
+        ATTN_RESOLUTIONS: [ ]
+        CH_MULT: [ 1, 2, 4, 4 ]
+        Z_CHANNELS: 16
+        DROPOUT: 0.0
+        RESAMP_WITH_CONV: True
+        GIVE_PRE_END: False
+        TANH_OUT: False
+    #
+    COND_STAGE_MODEL:
+      NAME: T5PlusClipFluxEmbedder
+      T5_MODEL:
+        NAME: HFEmbedder
+        HF_MODEL_CLS: T5EncoderModel
+        MODEL_PATH: ms://AI-ModelScope/FLUX.1-schnell@text_encoder_2/
+        HF_TOKENIZER_CLS: T5Tokenizer
+        TOKENIZER_PATH: ms://AI-ModelScope/FLUX.1-schnell@tokenizer_2/
+        MAX_LENGTH: 256
+        OUTPUT_KEY: last_hidden_state
+        D_TYPE: bfloat16
+        BATCH_INFER: False
+        CLEAN: whitespace
+      CLIP_MODEL:
+        NAME: HFEmbedder
+        HF_MODEL_CLS: CLIPTextModel
+        MODEL_PATH: ms://AI-ModelScope/FLUX.1-schnell@text_encoder/
+        HF_TOKENIZER_CLS: CLIPTokenizer
+        TOKENIZER_PATH: ms://AI-ModelScope/FLUX.1-schnell@tokenizer/
+        MAX_LENGTH: 77
+        OUTPUT_KEY: pooler_output
+        D_TYPE: bfloat16
+        BATCH_INFER: True
+        CLEAN: whitespace
+  #
+  SAMPLE_ARGS:
+    SAMPLE_STEPS: 4
+    SAMPLER: flow_euler
+    SEED: 2024
+    IMAGE_SIZE: [ 1024, 1024 ]
+    GUIDE_SCALE: 3.5
+  #
+  OPTIMIZER:
+    NAME: AdamW
+    LEARNING_RATE: 4e-4
+    BETAS: [ 0.9, 0.999 ]
+    EPS: 1e-8
+    WEIGHT_DECAY: 1e-2
+    AMSGRAD: False
+  #
+  TRAIN_DATA:
+    NAME: ImageTextPairMSDataset
+    MODE: train
+    MS_DATASET_NAME: style_custom_dataset
+    MS_DATASET_NAMESPACE: damo
+    MS_DATASET_SUBNAME: 3D
+    PROMPT_PREFIX: ""
+    MS_DATASET_SPLIT: train
+    MS_REMAP_KEYS: { 'Image:FILE': 'Target:FILE' }
+    REPLACE_STYLE: False
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    NUM_WORKERS: 4
+    SAMPLER:
+      NAME: LoopSampler
+    TRANSFORMS:
+      - NAME: LoadImageFromFile
+        RGB_ORDER: RGB
+        BACKEND: pillow
+      - NAME: FlexibleResize
+        INTERPOLATION: bilinear
+        SIZE: [ 1024, 1024 ]
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'img' ]
+        BACKEND: pillow
+      - NAME: FlexibleCenterCrop
+        SIZE: [ 1024, 1024 ]
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'img' ]
+        BACKEND: pillow
+      - NAME: ImageToTensor
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'img' ]
+        BACKEND: pillow
+      - NAME: Normalize
+        MEAN: [ 0.5,  0.5,  0.5 ]
+        STD: [ 0.5,  0.5,  0.5 ]
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'image' ]
+        BACKEND: torchvision
+      - NAME: Select
+        KEYS: [ 'image', 'prompt' ]
+        META_KEYS: [ 'data_key' ]
+  #
+  EVAL_DATA:
+    NAME: Text2ImageDataset
+    MODE: eval
+    PROMPT_FILE:
+    PROMPT_DATA: [ "a boy wearing a jacket", "a dog running on the lawn" ]
+    IMAGE_SIZE: [ 1024, 1024 ]
+    FIELDS: [ "prompt" ]
+    DELIMITER: '#;#'
+    PROMPT_PREFIX: ''
+    PIN_MEMORY: True
+    BATCH_SIZE: 2
+    NUM_WORKERS: 4
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ 'index', 'prompt' ]
+        META_KEYS: [ 'image_size' ]
+  #
+  TRAIN_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0
+    - NAME: BackwardHook
+#      GRADIENT_CLIP: 1.0
+      PRIORITY: 10
+    - NAME: LogHook
+      LOG_INTERVAL: 10
+    -
+      NAME: TensorboardLogHook
+    -
+      NAME: CheckpointHook
+      INTERVAL: 10000
+      PRIORITY: 200
+      SAVE_LAST: True
+      SAVE_NAME_PREFIX: 'step'
+      DISABLE_SNAPSHOT: True
+  EVAL_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0

scepter/methods/scedit/ctr/sd15_512_sce_ctr_hed.yaml

@@ -168,13 +168,13 @@ SOLVER:
         RGB_ORDER: RGB
         BACKEND: pillow
       - NAME: Resize
-        SIZE: 768
+        SIZE: 512
         INTERPOLATION: bilinear
         INPUT_KEY: [ 'img' ]
         OUTPUT_KEY: [ 'img' ]
         BACKEND: pillow
       - NAME: CenterCrop
-        SIZE: 768
+        SIZE: 512
         INPUT_KEY: [ 'img' ]
         OUTPUT_KEY: [ 'img' ]
         BACKEND: pillow
@@ -216,13 +216,13 @@ SOLVER:
         RGB_ORDER: RGB
         BACKEND: pillow
       - NAME: Resize
-        SIZE: 768
+        SIZE: 512
         INTERPOLATION: bilinear
         INPUT_KEY: [ 'img' ]
         OUTPUT_KEY: [ 'img' ]
         BACKEND: pillow
       - NAME: CenterCrop
-        SIZE: 768
+        SIZE: 512
         INPUT_KEY: [ 'img' ]
         OUTPUT_KEY: [ 'img' ]
         BACKEND: pillow

scepter/methods/studio/chatbot/chatbot.yaml

@@ -0,0 +1,25 @@
+WORK_DIR: chatbot
+FILE_SYSTEM:
+  - NAME: LocalFs
+    TEMP_DIR: ./cache/cache_data
+  - NAME: ModelscopeFs
+    TEMP_DIR: ./cache/cache_data
+  - NAME: HuggingfaceFs
+    TEMP_DIR: ./cache/cache_data
+#
+ENABLE_I2V: False
+SKIP_EXAMPLES: True
+#
+MODEL:
+  EDIT_MODEL:
+    MODEL_CFG_DIR: scepter/methods/studio/chatbot/models/
+  I2V:
+    MODEL_NAME: CogVideoX-5b-I2V
+    MODEL_DIR: ms://ZhipuAI/CogVideoX-5b-I2V/
+  CAPTIONER:
+    MODEL_NAME: InternVL2-2B
+    MODEL_DIR: ms://OpenGVLab/InternVL2-2B/
+    PROMPT: '<image>\nThis image is the first frame of a video. Based on this image, please imagine what changes may occur in the next few seconds of the video. Please output brief description, such as "a dog running" or "a person turns to left". No more than 30 words.'
+  ENHANCER:
+    MODEL_NAME: Meta-Llama-3.1-8B-Instruct
+    MODEL_DIR: ms://LLM-Research/Meta-Llama-3.1-8B-Instruct/

scepter/methods/studio/chatbot/models/ace_0.6b_1024.yaml

@@ -0,0 +1,128 @@
+NAME: ACE_0.6B_1024
+IS_DEFAULT: False
+USE_DYNAMIC_MODEL: True
+DEFAULT_PARAS:
+  PARAS:
+  #
+  INPUT:
+    INPUT_IMAGE:
+    INPUT_MASK:
+    TASK:
+    PROMPT: ""
+    NEGATIVE_PROMPT: ""
+    OUTPUT_HEIGHT: 1024
+    OUTPUT_WIDTH: 1024
+    SAMPLER: ddim
+    SAMPLE_STEPS: 50
+    GUIDE_SCALE: 4.5
+    GUIDE_RESCALE: 0.5
+    SEED: -1
+    TAR_INDEX: 0
+  OUTPUT:
+    LATENT:
+    IMAGES:
+    SEED:
+  MODULES_PARAS:
+    FIRST_STAGE_MODEL:
+      FUNCTION:
+        - NAME: encode
+          DTYPE: float16
+          INPUT: ["IMAGE"]
+        - NAME: decode
+          DTYPE: float16
+          INPUT: ["LATENT"]
+    #
+    DIFFUSION_MODEL:
+      FUNCTION:
+        - NAME: forward
+          DTYPE: float16
+          INPUT: ["SAMPLE_STEPS", "SAMPLE", "GUIDE_SCALE"]
+    #
+    COND_STAGE_MODEL:
+      FUNCTION:
+        - NAME: encode_list_of_list
+          DTYPE: bfloat16
+          INPUT: ["PROMPT"]
+#
+MODEL:
+  NAME: LatentDiffusionACE
+  PRETRAINED_MODEL:
+  IGNORE_KEYS: [ ]
+  SCALE_FACTOR: 0.18215
+  SIZE_FACTOR: 8
+  DECODER_BIAS: 0.5
+  DEFAULT_N_PROMPT: ""
+  TEXT_IDENTIFIER: [ '{image}', '{image1}', '{image2}', '{image3}', '{image4}', '{image5}', '{image6}', '{image7}', '{image8}', '{image9}' ]
+  USE_TEXT_POS_EMBEDDINGS: True
+  #
+  DIFFUSION:
+    NAME: BaseDiffusion
+    PREDICTION_TYPE: eps
+    MIN_SNR_GAMMA:
+    NOISE_SCHEDULER:
+      NAME: LinearScheduler
+      NUM_TIMESTEPS: 1000
+      BETA_MIN: 0.0001
+      BETA_MAX: 0.02
+  #
+  DIFFUSION_MODEL:
+    NAME: ACE
+    PRETRAINED_MODEL: ms://iic/ACE-0.6B-1024px@models/dit/ace_0.6b_1024px.pth
+    IGNORE_KEYS: [ ]
+    PATCH_SIZE: 2
+    IN_CHANNELS: 4
+    HIDDEN_SIZE: 1152
+    DEPTH: 28
+    NUM_HEADS: 16
+    MLP_RATIO: 4.0
+    PRED_SIGMA: True
+    DROP_PATH: 0.0
+    WINDOW_DIZE: 0
+    Y_CHANNELS: 4096
+    MAX_SEQ_LEN: 4096
+    QK_NORM: True
+    USE_GRAD_CHECKPOINT: True
+    ATTENTION_BACKEND: flash_attn
+  #
+  FIRST_STAGE_MODEL:
+    NAME: AutoencoderKL
+    EMBED_DIM: 4
+    PRETRAINED_MODEL: ms://iic/ACE-0.6B-1024px@models/vae/vae.bin
+    IGNORE_KEYS: []
+    #
+    ENCODER:
+      NAME: Encoder
+      CH: 128
+      OUT_CH: 3
+      NUM_RES_BLOCKS: 2
+      IN_CHANNELS: 3
+      ATTN_RESOLUTIONS: [ ]
+      CH_MULT: [ 1, 2, 4, 4 ]
+      Z_CHANNELS: 4
+      DOUBLE_Z: True
+      DROPOUT: 0.0
+      RESAMP_WITH_CONV: True
+    #
+    DECODER:
+      NAME: Decoder
+      CH: 128
+      OUT_CH: 3
+      NUM_RES_BLOCKS: 2
+      IN_CHANNELS: 3
+      ATTN_RESOLUTIONS: [ ]
+      CH_MULT: [ 1, 2, 4, 4 ]
+      Z_CHANNELS: 4
+      DROPOUT: 0.0
+      RESAMP_WITH_CONV: True
+      GIVE_PRE_END: False
+      TANH_OUT: False
+  #
+  COND_STAGE_MODEL:
+    NAME: T5EmbedderHF
+    PRETRAINED_MODEL: ms://iic/ACE-0.6B-1024px@models/text_encoder/t5-v1_1-xxl/
+    TOKENIZER_PATH: ms://iic/ACE-0.6B-1024px@models/tokenizer/t5-v1_1-xxl
+    LENGTH: 120
+    T5_DTYPE: bfloat16
+    ADDED_IDENTIFIER: [ '{image}', '{caption}', '{mask}', '{ref_image}', '{image1}', '{image2}', '{image3}', '{image4}', '{image5}', '{image6}', '{image7}', '{image8}', '{image9}' ]
+    CLEAN: whitespace
+    USE_GRAD: False

scepter/methods/studio/chatbot/models/ace_0.6b_1024_refiner.yaml

@@ -0,0 +1,284 @@
+NAME: ACE_0.6B_1024_REFINER
+IS_DEFAULT: False
+USE_DYNAMIC_MODEL: True
+DEFAULT_PARAS:
+  PARAS:
+  #
+  INPUT:
+    INPUT_IMAGE:
+    INPUT_MASK:
+    TASK:
+    PROMPT: ""
+    NEGATIVE_PROMPT: ""
+    OUTPUT_HEIGHT: 1024
+    OUTPUT_WIDTH: 1024
+    SAMPLER: ddim
+    SAMPLE_STEPS: 50
+    GUIDE_SCALE: 4.5
+    GUIDE_RESCALE: 0.5
+    SEED: -1
+    TAR_INDEX: 0
+    REFINER_SCALE: 0.2
+    USE_ACE: True
+    #REFINER_PROMPT: "High Resolution, Sharpness, Clarity, Detail Enhancement, Noise Reduction, HD, 4k, Image Restoration, HDR"
+    REFINER_PROMPT: "High Resolution, Sharpness, Clarity, Detail Enhancement, Noise Reduction, HD, 4k, Image Restoration, HDR"
+  OUTPUT:
+    LATENT:
+    IMAGES:
+    SEED:
+  MODULES_PARAS:
+    FIRST_STAGE_MODEL:
+      FUNCTION:
+        - NAME: encode
+          DTYPE: float16
+          INPUT: ["IMAGE"]
+        - NAME: decode
+          DTYPE: float16
+          INPUT: ["LATENT"]
+    #
+    DIFFUSION_MODEL:
+      FUNCTION:
+        - NAME: forward
+          DTYPE: float16
+          INPUT: ["SAMPLE_STEPS", "SAMPLE", "GUIDE_SCALE"]
+    #
+    COND_STAGE_MODEL:
+      FUNCTION:
+        - NAME: encode_list_of_list
+          DTYPE: bfloat16
+          INPUT: ["PROMPT"]
+#
+MODEL:
+  NAME: LatentDiffusionACE
+  PRETRAINED_MODEL:
+  IGNORE_KEYS: [ ]
+  SCALE_FACTOR: 0.18215
+  SIZE_FACTOR: 8
+  DECODER_BIAS: 0.5
+  DEFAULT_N_PROMPT: ""
+  TEXT_IDENTIFIER: [ '{image}', '{image1}', '{image2}', '{image3}', '{image4}', '{image5}', '{image6}', '{image7}', '{image8}', '{image9}' ]
+  USE_TEXT_POS_EMBEDDINGS: True
+  #
+  DIFFUSION:
+    NAME: BaseDiffusion
+    PREDICTION_TYPE: eps
+    MIN_SNR_GAMMA:
+    NOISE_SCHEDULER:
+      NAME: LinearScheduler
+      NUM_TIMESTEPS: 1000
+      BETA_MIN: 0.0001
+      BETA_MAX: 0.02
+  #
+  DIFFUSION_MODEL:
+    NAME: ACE
+    PRETRAINED_MODEL: ms://iic/ACE-0.6B-1024px@models/dit/ace_0.6b_1024px.pth
+    IGNORE_KEYS: [ ]
+    PATCH_SIZE: 2
+    IN_CHANNELS: 4
+    HIDDEN_SIZE: 1152
+    DEPTH: 28
+    NUM_HEADS: 16
+    MLP_RATIO: 4.0
+    PRED_SIGMA: True
+    DROP_PATH: 0.0
+    WINDOW_DIZE: 0
+    Y_CHANNELS: 4096
+    MAX_SEQ_LEN: 4096
+    QK_NORM: True
+    USE_GRAD_CHECKPOINT: True
+    ATTENTION_BACKEND: flash_attn
+  #
+  FIRST_STAGE_MODEL:
+    NAME: AutoencoderKL
+    EMBED_DIM: 4
+    PRETRAINED_MODEL: ms://iic/ACE-0.6B-1024px@models/vae/vae.bin
+    IGNORE_KEYS: []
+    #
+    ENCODER:
+      NAME: Encoder
+      CH: 128
+      OUT_CH: 3
+      NUM_RES_BLOCKS: 2
+      IN_CHANNELS: 3
+      ATTN_RESOLUTIONS: [ ]
+      CH_MULT: [ 1, 2, 4, 4 ]
+      Z_CHANNELS: 4
+      DOUBLE_Z: True
+      DROPOUT: 0.0
+      RESAMP_WITH_CONV: True
+    #
+    DECODER:
+      NAME: Decoder
+      CH: 128
+      OUT_CH: 3
+      NUM_RES_BLOCKS: 2
+      IN_CHANNELS: 3
+      ATTN_RESOLUTIONS: [ ]
+      CH_MULT: [ 1, 2, 4, 4 ]
+      Z_CHANNELS: 4
+      DROPOUT: 0.0
+      RESAMP_WITH_CONV: True
+      GIVE_PRE_END: False
+      TANH_OUT: False
+  #
+  COND_STAGE_MODEL:
+    NAME: T5EmbedderHF
+    PRETRAINED_MODEL: ms://iic/ACE-0.6B-1024px@models/text_encoder/t5-v1_1-xxl/
+    TOKENIZER_PATH: ms://iic/ACE-0.6B-1024px@models/tokenizer/t5-v1_1-xxl
+    LENGTH: 120
+    T5_DTYPE: bfloat16
+    ADDED_IDENTIFIER: [ '{image}', '{caption}', '{mask}', '{ref_image}', '{image1}', '{image2}', '{image3}', '{image4}', '{image5}', '{image6}', '{image7}', '{image8}', '{image9}' ]
+    CLEAN: whitespace
+    USE_GRAD: False
+
+ACE_PROMPT: [
+  "A cute cartoon rabbit holding a whiteboard that says 'ACE Refiner', standing in a sunny meadow filled with flowers, with a big smile and bright colors.",
+  "A beautiful young woman with long flowing hair, wearing a summer dress, holding a whiteboard that reads 'ACE Refiner' while sitting on a park bench surrounded by cherry blossoms.",
+  "An adorable cartoon cat wearing oversized glasses, holding a whiteboard that says 'ACE Refiner', perched on a stack of colorful books in a cozy library setting.",
+  "A charming girl with pigtails, wearing a cute school uniform, enthusiastically holding a whiteboard that has 'ACE Refiner' written on it, in a bright and cheerful classroom full of educational posters.",
+  "A friendly cartoon dog with floppy ears, sitting in front of a doghouse, proudly holding a whiteboard that says 'ACE Refiner', with a playful expression and a blue sky in the background.",
+  "A cute anime girl with big expressive eyes, dressed in a colorful outfit, holding a whiteboard that reads 'ACE Refiner' in a fantastical landscape filled with mythical creatures.",
+  "A vibrant cartoon fox holding a whiteboard that says 'ACE Refiner', standing on a rock by a sparkling stream, surrounded by lush greenery and butterflies.",
+  "A stylish young woman in a business outfit, smiling as she holds a whiteboard written with 'ACE Refiner', in a modern office filled with plants and natural light.",
+  "A cute cartoon unicorn holding a sparkling whiteboard that says 'ACE Refiner', frolicking in a magical forest, with rainbows and stars in the background.",
+  "A happy family, consisting of a cute little girl and her playful puppy, holding a whiteboard that says 'ACE Refiner', together in their backyard on a sunny day."
+]
+REFINER_MODEL:
+  NAME: ""
+  IS_DEFAULT: False
+  DEFAULT_PARAS:
+    PARAS:
+      RESOLUTIONS: [ [ 1024, 1024 ] ]
+    INPUT:
+      INPUT_IMAGE:
+      INPUT_MASK:
+      TASK:
+      PROMPT: ""
+      NEGATIVE_PROMPT: ""
+      OUTPUT_HEIGHT: 1024
+      OUTPUT_WIDTH: 1024
+      SAMPLER: flow_euler
+      SAMPLE_STEPS: 30
+      GUIDE_SCALE: 3.5
+      GUIDE_RESCALE:
+    OUTPUT:
+      LATENT:
+      IMAGES:
+      SEED:
+    MODULES_PARAS:
+      FIRST_STAGE_MODEL:
+        FUNCTION:
+          - NAME: encode
+            DTYPE: bfloat16
+            INPUT: [ "IMAGE" ]
+          - NAME: decode
+            DTYPE: bfloat16
+            INPUT: [ "LATENT" ]
+        PARAS:
+          SCALE_FACTOR: 1.5305
+          SHIFT_FACTOR: 0.0609
+          SIZE_FACTOR: 8
+      DIFFUSION_MODEL:
+        FUNCTION:
+          - NAME: forward
+            DTYPE: bfloat16
+            INPUT: [ "SAMPLE_STEPS", "SAMPLE", "GUIDE_SCALE" ]
+      COND_STAGE_MODEL:
+        FUNCTION:
+          - NAME: encode
+            DTYPE: bfloat16
+            INPUT: [ "PROMPT" ]
+
+  MODEL:
+    DIFFUSION:
+      NAME: DiffusionFluxRF
+      PREDICTION_TYPE: raw
+      NOISE_SCHEDULER:
+        NAME: FlowMatchSigmaScheduler
+        WEIGHTING_SCHEME: logit_normal
+        SHIFT: 3.0
+        LOGIT_MEAN: 0.0
+        LOGIT_STD: 1.0
+        MODE_SCALE: 1.29
+    DIFFUSION_MODEL:
+      NAME: FluxMR
+      PRETRAINED_MODEL: ms://AI-ModelScope/FLUX.1-dev@flux1-dev.safetensors
+      IN_CHANNELS: 64
+      OUT_CHANNELS: 64
+      HIDDEN_SIZE: 3072
+      NUM_HEADS: 24
+      AXES_DIM: [ 16, 56, 56 ]
+      THETA: 10000
+      VEC_IN_DIM: 768
+      GUIDANCE_EMBED: True
+      CONTEXT_IN_DIM: 4096
+      MLP_RATIO: 4.0
+      QKV_BIAS: True
+      DEPTH: 19
+      DEPTH_SINGLE_BLOCKS: 38
+      USE_GRAD_CHECKPOINT: True
+      ATTN_BACKEND: flash_attn
+    #
+    FIRST_STAGE_MODEL:
+      NAME: AutoencoderKLFlux
+      EMBED_DIM: 16
+      PRETRAINED_MODEL:  ms://AI-ModelScope/FLUX.1-dev@ae.safetensors
+      IGNORE_KEYS: [ ]
+      BATCH_SIZE: 8
+      USE_CONV: False
+      SCALE_FACTOR: 0.3611
+      SHIFT_FACTOR: 0.1159
+      #
+      ENCODER:
+        NAME: Encoder
+        USE_CHECKPOINT: False
+        CH: 128
+        OUT_CH: 3
+        NUM_RES_BLOCKS: 2
+        IN_CHANNELS: 3
+        ATTN_RESOLUTIONS: [ ]
+        CH_MULT: [ 1, 2, 4, 4 ]
+        Z_CHANNELS: 16
+        DOUBLE_Z: True
+        DROPOUT: 0.0
+        RESAMP_WITH_CONV: True
+      #
+      DECODER:
+        NAME: Decoder
+        USE_CHECKPOINT: False
+        CH: 128
+        OUT_CH: 3
+        NUM_RES_BLOCKS: 2
+        IN_CHANNELS: 3
+        ATTN_RESOLUTIONS: [ ]
+        CH_MULT: [ 1, 2, 4, 4 ]
+        Z_CHANNELS: 16
+        DROPOUT: 0.0
+        RESAMP_WITH_CONV: True
+        GIVE_PRE_END: False
+        TANH_OUT: False
+    #
+    COND_STAGE_MODEL:
+      NAME: T5PlusClipFluxEmbedder
+      T5_MODEL:
+        NAME: HFEmbedder
+        HF_MODEL_CLS: T5EncoderModel
+        MODEL_PATH: ms://AI-ModelScope/FLUX.1-dev@text_encoder_2/
+        HF_TOKENIZER_CLS: T5Tokenizer
+        TOKENIZER_PATH: ms://AI-ModelScope/FLUX.1-dev@tokenizer_2/
+        MAX_LENGTH: 512
+        OUTPUT_KEY: last_hidden_state
+        D_TYPE: bfloat16
+        BATCH_INFER: False
+        CLEAN: whitespace
+      CLIP_MODEL:
+        NAME: HFEmbedder
+        HF_MODEL_CLS: CLIPTextModel
+        MODEL_PATH: ms://AI-ModelScope/FLUX.1-dev@text_encoder/
+        HF_TOKENIZER_CLS: CLIPTokenizer
+        TOKENIZER_PATH: ms://AI-ModelScope/FLUX.1-dev@tokenizer/
+        MAX_LENGTH: 77
+        OUTPUT_KEY: pooler_output
+        D_TYPE: bfloat16
+        BATCH_INFER: True
+        CLEAN: whitespace

scepter/methods/studio/chatbot/models/ace_0.6b_512.yaml

@@ -0,0 +1,128 @@
+NAME: ACE_0.6B_512
+IS_DEFAULT: True
+USE_DYNAMIC_MODEL: True
+DEFAULT_PARAS:
+  PARAS:
+  #
+  INPUT:
+    INPUT_IMAGE:
+    INPUT_MASK:
+    TASK:
+    PROMPT: ""
+    NEGATIVE_PROMPT: ""
+    OUTPUT_HEIGHT: 512
+    OUTPUT_WIDTH: 512
+    SAMPLER: ddim
+    SAMPLE_STEPS: 20
+    GUIDE_SCALE: 4.5
+    GUIDE_RESCALE: 0.5
+    SEED: -1
+    TAR_INDEX: 0
+  OUTPUT:
+    LATENT:
+    IMAGES:
+    SEED:
+  MODULES_PARAS:
+    FIRST_STAGE_MODEL:
+      FUNCTION:
+        - NAME: encode
+          DTYPE: float16
+          INPUT: ["IMAGE"]
+        - NAME: decode
+          DTYPE: float16
+          INPUT: ["LATENT"]
+    #
+    DIFFUSION_MODEL:
+      FUNCTION:
+        - NAME: forward
+          DTYPE: float16
+          INPUT: ["SAMPLE_STEPS", "SAMPLE", "GUIDE_SCALE"]
+    #
+    COND_STAGE_MODEL:
+      FUNCTION:
+        - NAME: encode_list_of_list
+          DTYPE: bfloat16
+          INPUT: ["PROMPT"]
+#
+MODEL:
+  NAME: LatentDiffusionACE
+  PRETRAINED_MODEL:
+  IGNORE_KEYS: [ ]
+  SCALE_FACTOR: 0.18215
+  SIZE_FACTOR: 8
+  DECODER_BIAS: 0.5
+  DEFAULT_N_PROMPT: ""
+  TEXT_IDENTIFIER: [ '{image}', '{image1}', '{image2}', '{image3}', '{image4}', '{image5}', '{image6}', '{image7}', '{image8}', '{image9}' ]
+  USE_TEXT_POS_EMBEDDINGS: True
+  #
+  DIFFUSION:
+    NAME: BaseDiffusion
+    PREDICTION_TYPE: eps
+    MIN_SNR_GAMMA:
+    NOISE_SCHEDULER:
+      NAME: LinearScheduler
+      NUM_TIMESTEPS: 1000
+      BETA_MIN: 0.0001
+      BETA_MAX: 0.02
+  #
+  DIFFUSION_MODEL:
+    NAME: ACE
+    PRETRAINED_MODEL: ms://iic/ACE-0.6B-512px@models/dit/ace_0.6b_512px.pth
+    IGNORE_KEYS: [ ]
+    PATCH_SIZE: 2
+    IN_CHANNELS: 4
+    HIDDEN_SIZE: 1152
+    DEPTH: 28
+    NUM_HEADS: 16
+    MLP_RATIO: 4.0
+    PRED_SIGMA: True
+    DROP_PATH: 0.0
+    WINDOW_DIZE: 0
+    Y_CHANNELS: 4096
+    MAX_SEQ_LEN: 1024
+    QK_NORM: True
+    USE_GRAD_CHECKPOINT: True
+    ATTENTION_BACKEND: flash_attn
+  #
+  FIRST_STAGE_MODEL:
+    NAME: AutoencoderKL
+    EMBED_DIM: 4
+    PRETRAINED_MODEL: ms://iic/ACE-0.6B-512px@models/vae/vae.bin
+    IGNORE_KEYS: []
+    #
+    ENCODER:
+      NAME: Encoder
+      CH: 128
+      OUT_CH: 3
+      NUM_RES_BLOCKS: 2
+      IN_CHANNELS: 3
+      ATTN_RESOLUTIONS: [ ]
+      CH_MULT: [ 1, 2, 4, 4 ]
+      Z_CHANNELS: 4
+      DOUBLE_Z: True
+      DROPOUT: 0.0
+      RESAMP_WITH_CONV: True
+    #
+    DECODER:
+      NAME: Decoder
+      CH: 128
+      OUT_CH: 3
+      NUM_RES_BLOCKS: 2
+      IN_CHANNELS: 3
+      ATTN_RESOLUTIONS: [ ]
+      CH_MULT: [ 1, 2, 4, 4 ]
+      Z_CHANNELS: 4
+      DROPOUT: 0.0
+      RESAMP_WITH_CONV: True
+      GIVE_PRE_END: False
+      TANH_OUT: False
+  #
+  COND_STAGE_MODEL:
+    NAME: T5EmbedderHF
+    PRETRAINED_MODEL: ms://iic/ACE-0.6B-512px@models/text_encoder/t5-v1_1-xxl/
+    TOKENIZER_PATH: ms://iic/ACE-0.6B-512px@models/tokenizer/t5-v1_1-xxl
+    LENGTH: 120
+    T5_DTYPE: bfloat16
+    ADDED_IDENTIFIER: [ '{image}', '{caption}', '{mask}', '{ref_image}', '{image1}', '{image2}', '{image3}', '{image4}', '{image5}', '{image6}', '{image7}', '{image8}', '{image9}' ]
+    CLEAN: whitespace
+    USE_GRAD: False

scepter/methods/studio/inference/dit/cogvideox_2b_pro.yaml

@@ -0,0 +1,152 @@
+NAME: COGVIDEOX_2B
+IS_DEFAULT: False
+DEFAULT_PARAS:
+  PARAS:
+    RESOLUTIONS: [[480, 720]]
+  INPUT:
+    IMAGE:
+    ORIGINAL_SIZE_AS_TUPLE: [480, 720]
+    TARGET_SIZE_AS_TUPLE: [480, 720]
+    PROMPT: ""
+    NEGATIVE_PROMPT: ""
+    PROMPT_PREFIX: ""
+    SAMPLE: ddim
+    SAMPLE_STEPS: 50
+    GUIDE_SCALE: 6.0
+    GUIDE_RESCALE: 0.0
+    DISCRETIZATION: trailing
+    NUM_FRAMES:
+      DEFAULT: 49
+      VISIBLE: True
+    FPS:
+      DEFAULT: 8
+      VISIBLE: True
+  OUTPUT:
+    VIDEOS:
+    SEED:
+  MODULES_PARAS:
+    FIRST_STAGE_MODEL:
+      FUNCTION:
+        -
+          NAME: decode
+          DTYPE: bfloat16
+          INPUT: ["LATENT"]
+      PARAS:
+          SCALING_FACTOR_IMAGE: 1.15258426
+    DIFFUSION_MODEL:
+      FUNCTION:
+        -
+          NAME: forward
+          DTYPE: bfloat16
+          INPUT: ["SAMPLE_STEPS", "SAMPLE", "GUIDE_SCALE", "GUIDE_RESCALE", "DISCRETIZATION", "NUM_FRAMES", "FPS"]
+      PARAS:
+        USE_ROTARY_POSITIONAL_EMBEDDINGS: False
+        PATCH_SIZE: 2
+        LATENT_CHANNELS: 16
+        SCALE_FACTOR_SPATIAL: 8
+        SCALE_FACTOR_TEMPORAL: 4
+        ATTENTION_HEAD_DIM: 64
+        SAMPLE_HEIGHT: 480
+        SAMPLE_WIDTH: 720
+    COND_STAGE_MODEL:
+      FUNCTION:
+        -
+          NAME: encode
+          DTYPE: bfloat16
+          INPUT: ["PROMPT"]
+#
+MODEL:
+  PRETRAINED_MODEL:
+  DIFFUSION:
+    NAME: BaseDiffusion
+    PREDICTION_TYPE: v
+    NOISE_SCHEDULER:
+      NAME: ScaledLinearScheduler
+      BETA_MIN: 0.00085
+      BETA_MAX: 0.012
+      SNR_SHIFT_SCALE: 3.0
+      RESCALE_BETAS_ZERO_SNR: True
+    DIFFUSION_SAMPLERS:
+      NAME: DDIMSampler
+      DISCRETIZATION_TYPE: trailing
+      ETA: 0.0
+  #
+  DIFFUSION_MODEL:
+    NAME: CogVideoXTransformer3DModel
+    DTYPE: bfloat16
+    PRETRAINED_MODEL: ms://AI-ModelScope/CogVideoX-2b@transformer/diffusion_pytorch_model.safetensors
+    NUM_ATTENTION_HEADS: 30
+    ATTENTION_HEAD_DIM: 64
+    IN_CHANNELS: 16
+    OUT_CHANNELS: 16
+    FLIP_SIN_TO_COS: True
+    FREQ_SHIFT: 0
+    TIME_EMBED_DIM: 512
+    TEXT_EMBED_DIM: 4096
+    NUM_LAYERS: 30
+    DROPOUT: 0.0
+    ATTENTION_BIAS: True
+    SAMPLE_WIDTH: 90
+    SAMPLE_HEIGHT: 60
+    SAMPLE_FRAMES: 49
+    PATCH_SIZE: 2
+    TEMPORAL_COMPRESSION_RATIO: 4
+    MAX_TEXT_SEQ_LENGTH: 226
+    ACTIVATION_FN: "gelu-approximate"
+    TIMESTEP_ACTIVATION_FN: "silu"
+    NORM_ELEMENTWISE_AFFINE: True
+    NORM_EPS: 1e-5
+    SPATIAL_INTERPOLATION_SCALE: 1.875
+    TEMPORAL_INTERPOLATION_SCALE: 1.0
+    USE_ROTARY_POSITIONAL_EMBEDDINGS: False
+    USE_LEARNED_POSITIONAL_EMBEDDINGS: False
+    GRADIENT_CHECKPOINTING: False
+  #
+  FIRST_STAGE_MODEL:
+    NAME: AutoencoderKLCogVideoX
+    DTYPE: bfloat16
+    PRETRAINED_MODEL: ms://AI-ModelScope/CogVideoX-2b@vae/diffusion_pytorch_model.safetensors
+    SAMPLE_HEIGHT: 480
+    SAMPLE_WIDTH: 720
+    USE_QUANT_CONV: False
+    USE_POST_QUANT_CONV: False
+    USE_SLICING: True
+    USE_TILING: True
+    GRADIENT_CHECKPOINTING: False
+    ENCODER:
+      NAME: CogVideoXEncoder3D
+      IN_CHANNELS: 3
+      OUT_CHANNELS: 16
+      UP_BLOCK_TYPES: [ "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D" ]
+      BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+      LAYERS_PER_BLOCK: 3
+      ACT_FN: "silu"
+      NORM_EPS: 1e-6
+      NORM_NUM_GROUPS: 32
+      DROPOUT: 0.0
+      PAD_MODE: "first"
+      TEMPORAL_COMPRESSION_RATIO: 4
+      GRADIENT_CHECKPOINTING: False
+    DECODER:
+      NAME: CogVideoXDecoder3D
+      IN_CHANNELS: 16
+      OUT_CHANNELS: 3
+      UP_BLOCK_TYPES: [ "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D" ]
+      BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+      LAYERS_PER_BLOCK: 3
+      ACT_FN: "silu"
+      NORM_EPS: 1e-6
+      NORM_NUM_GROUPS: 32
+      DROPOUT: 0.0
+      PAD_MODE: "first"
+      TEMPORAL_COMPRESSION_RATIO: 4
+      GRADIENT_CHECKPOINTING: False
+  #
+  COND_STAGE_MODEL:
+    NAME: T5EmbedderHF
+    PRETRAINED_MODEL: ms://AI-ModelScope/t5-v1_1-xxl
+    TOKENIZER_PATH: ms://AI-ModelScope/t5-v1_1-xxl
+    LENGTH: 226
+    CLEAN:
+    USE_GRAD: False
+    T5_DTYPE: bfloat16

scepter/methods/studio/inference/dit/cogvideox_5b_pro.yaml

@@ -0,0 +1,154 @@
+NAME: COGVIDEOX_5B
+IS_DEFAULT: False
+DEFAULT_PARAS:
+  PARAS:
+    RESOLUTIONS: [[480, 720]]
+  INPUT:
+    IMAGE:
+    ORIGINAL_SIZE_AS_TUPLE: [480, 720]
+    TARGET_SIZE_AS_TUPLE: [480, 720]
+    PROMPT: ""
+    NEGATIVE_PROMPT: ""
+    PROMPT_PREFIX: ""
+    SAMPLE: ddim
+    SAMPLE_STEPS: 50
+    GUIDE_SCALE: 6.0
+    GUIDE_RESCALE: 0.0
+    DISCRETIZATION: trailing
+    NUM_FRAMES:
+      DEFAULT: 49
+      VISIBLE: True
+    FPS:
+      DEFAULT: 8
+      VISIBLE: True
+  OUTPUT:
+    VIDEOS:
+    SEED:
+  MODULES_PARAS:
+    FIRST_STAGE_MODEL:
+      FUNCTION:
+        -
+          NAME: decode
+          DTYPE: bfloat16
+          INPUT: ["LATENT"]
+      PARAS:
+          SCALING_FACTOR_IMAGE: 0.7  # 5b diff
+    DIFFUSION_MODEL:
+      FUNCTION:
+        -
+          NAME: forward
+          DTYPE: bfloat16
+          INPUT: ["SAMPLE_STEPS", "SAMPLE", "GUIDE_SCALE", "GUIDE_RESCALE", "DISCRETIZATION", "NUM_FRAMES", "FPS"]
+      PARAS:
+        USE_ROTARY_POSITIONAL_EMBEDDINGS: True
+        PATCH_SIZE: 2
+        LATENT_CHANNELS: 16
+        SCALE_FACTOR_SPATIAL: 8
+        SCALE_FACTOR_TEMPORAL: 4
+        ATTENTION_HEAD_DIM: 64
+        SAMPLE_HEIGHT: 480
+        SAMPLE_WIDTH: 720
+    COND_STAGE_MODEL:
+      FUNCTION:
+        -
+          NAME: encode
+          DTYPE: bfloat16
+          INPUT: ["PROMPT"]
+#
+MODEL:
+  PRETRAINED_MODEL:
+  DIFFUSION:
+    NAME: BaseDiffusion
+    PREDICTION_TYPE: v
+    NOISE_SCHEDULER:
+      NAME: ScaledLinearScheduler
+      BETA_MIN: 0.00085
+      BETA_MAX: 0.012
+      SNR_SHIFT_SCALE: 1.0  # 5b diff
+      RESCALE_BETAS_ZERO_SNR: True
+    DIFFUSION_SAMPLERS:
+      NAME: DDIMSampler
+      DISCRETIZATION_TYPE: trailing
+      ETA: 0.0
+  #
+  DIFFUSION_MODEL:
+    NAME: CogVideoXTransformer3DModel
+    DTYPE: bfloat16
+    PRETRAINED_MODEL: # 5b diff
+      - ms://AI-ModelScope/CogVideoX-5b@transformer/diffusion_pytorch_model-00001-of-00002.safetensors
+      - ms://AI-ModelScope/CogVideoX-5b@transformer/diffusion_pytorch_model-00002-of-00002.safetensors
+    NUM_ATTENTION_HEADS: 48  # 5b diff
+    ATTENTION_HEAD_DIM: 64
+    IN_CHANNELS: 16
+    OUT_CHANNELS: 16
+    FLIP_SIN_TO_COS: True
+    FREQ_SHIFT: 0
+    TIME_EMBED_DIM: 512
+    TEXT_EMBED_DIM: 4096
+    NUM_LAYERS: 42  # 5b diff
+    DROPOUT: 0.0
+    ATTENTION_BIAS: True
+    SAMPLE_WIDTH: 90
+    SAMPLE_HEIGHT: 60
+    SAMPLE_FRAMES: 49
+    PATCH_SIZE: 2
+    TEMPORAL_COMPRESSION_RATIO: 4
+    MAX_TEXT_SEQ_LENGTH: 226
+    ACTIVATION_FN: "gelu-approximate"
+    TIMESTEP_ACTIVATION_FN: "silu"
+    NORM_ELEMENTWISE_AFFINE: True
+    NORM_EPS: 1e-5
+    SPATIAL_INTERPOLATION_SCALE: 1.875
+    TEMPORAL_INTERPOLATION_SCALE: 1.0
+    USE_ROTARY_POSITIONAL_EMBEDDINGS: True # 5b diff
+    USE_LEARNED_POSITIONAL_EMBEDDINGS: False
+    GRADIENT_CHECKPOINTING: True
+  #
+  FIRST_STAGE_MODEL:
+    NAME: AutoencoderKLCogVideoX
+    DTYPE: bfloat16
+    PRETRAINED_MODEL: ms://AI-ModelScope/CogVideoX-5b@vae/diffusion_pytorch_model.safetensors # 5b diff
+    SAMPLE_HEIGHT: 480
+    SAMPLE_WIDTH: 720
+    USE_QUANT_CONV: False
+    USE_POST_QUANT_CONV: False
+    USE_SLICING: True
+    USE_TILING: True
+    GRADIENT_CHECKPOINTING: True
+    ENCODER:
+      NAME: CogVideoXEncoder3D
+      IN_CHANNELS: 3
+      OUT_CHANNELS: 16
+      UP_BLOCK_TYPES: [ "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D" ]
+      BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+      LAYERS_PER_BLOCK: 3
+      ACT_FN: "silu"
+      NORM_EPS: 1e-6
+      NORM_NUM_GROUPS: 32
+      DROPOUT: 0.0
+      PAD_MODE: "first"
+      TEMPORAL_COMPRESSION_RATIO: 4
+      GRADIENT_CHECKPOINTING: True
+    DECODER:
+      NAME: CogVideoXDecoder3D
+      IN_CHANNELS: 16
+      OUT_CHANNELS: 3
+      UP_BLOCK_TYPES: [ "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D" ]
+      BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+      LAYERS_PER_BLOCK: 3
+      ACT_FN: "silu"
+      NORM_EPS: 1e-6
+      NORM_NUM_GROUPS: 32
+      DROPOUT: 0.0
+      PAD_MODE: "first"
+      TEMPORAL_COMPRESSION_RATIO: 4
+      GRADIENT_CHECKPOINTING: True
+  #
+  COND_STAGE_MODEL:
+    NAME: T5EmbedderHF
+    PRETRAINED_MODEL: ms://AI-ModelScope/t5-v1_1-xxl
+    TOKENIZER_PATH: ms://AI-ModelScope/t5-v1_1-xxl
+    LENGTH: 226
+    CLEAN:
+    USE_GRAD: False
+    T5_DTYPE: bfloat16

scepter/methods/studio/inference/dit/flux1.0_dev_pro.yaml

@@ -0,0 +1,201 @@
+NAME: FLUX1.0_DEV
+IS_DEFAULT: False
+DEFAULT_PARAS:
+  PARAS:
+    RESOLUTIONS: [[1024, 1024]]
+  INPUT:
+    IMAGE:
+    ORIGINAL_SIZE_AS_TUPLE: [1024, 1024]
+    TARGET_SIZE_AS_TUPLE: [1024, 1024]
+    PROMPT: ""
+    NEGATIVE_PROMPT:
+      DEFAULT: ""
+      VISIBLE: False
+    PROMPT_PREFIX: ""
+    SAMPLE:
+      VALUES: ["flow_euler"]
+      DEFAULT: "flow_euler"
+    SAMPLE_STEPS: 50
+    GUIDE_SCALE: 3.5
+    GUIDE_RESCALE:
+      DEFAULT: 0.0
+      VISIBLE: False
+    DISCRETIZATION:
+      VALUES: []
+      DEFAULT:
+      VISIBLE: False
+  OUTPUT:
+    LATENT:
+    IMAGES:
+    SEED:
+  MODULES_PARAS:
+    FIRST_STAGE_MODEL:
+      FUNCTION:
+        -
+          NAME: encode
+          DTYPE: bfloat16
+          INPUT: ["IMAGE"]
+        -
+          NAME: decode
+          DTYPE: bfloat16
+          INPUT: ["LATENT"]
+      PARAS:
+        SCALE_FACTOR: 1.5305
+        SHIFT_FACTOR: 0.0609
+        SIZE_FACTOR: 8
+    DIFFUSION_MODEL:
+      FUNCTION:
+        -
+          NAME: forward
+          DTYPE: bfloat16
+          INPUT: ["SAMPLE_STEPS", "SAMPLE", "GUIDE_SCALE"]
+    COND_STAGE_MODEL:
+      FUNCTION:
+        -
+          NAME: encode
+          DTYPE: bfloat16
+          INPUT: ["PROMPT"]
+#
+MODEL:
+  NAME: LatentDiffusionFlux
+  PARAMETERIZATION: rf
+  TIMESTEPS: 1000
+  MIN_SNR_GAMMA:
+  ZERO_TERMINAL_SNR: False
+  PRETRAINED_MODEL:
+  IGNORE_KEYS: [ ]
+  DEFAULT_N_PROMPT:
+  USE_EMA: False
+  EVAL_EMA: False
+  DIFFUSION:
+    # NAME DESCRIPTION:  TYPE:  default: 'DiffusionFluxRF'
+    NAME: DiffusionFluxRF
+    PREDICTION_TYPE: raw
+    # NOISE_SCHEDULER DESCRIPTION:  TYPE:  default: ''
+    NOISE_SCHEDULER:
+      # NAME DESCRIPTION:  TYPE:  default: 'FlowMatchSigmaScheduler'
+      NAME: FlowMatchSigmaScheduler
+      # WEIGHTING_SCHEME DESCRIPTION: The weighting scheme for sampling timesteps, choose from ['sigma_sqrt', 'logit_normal', 'mode', 'cosmap', 'none']. TYPE: str default: 'logit_normal'
+      WEIGHTING_SCHEME: logit_normal
+      SHIFT: 3.0
+      # LOGIT_MEAN DESCRIPTION: The mean of the logit distribution for sampling timesteps. TYPE: float default: 0.0
+      LOGIT_MEAN: 0.0
+      # LOGIT_STD DESCRIPTION: The standard deviation of the logit distribution for sampling timesteps. TYPE: float default: 1.0
+      LOGIT_STD: 1.0
+      # MODE_SCALE DESCRIPTION: The scale factor for the mode of the logit distribution for sampling timesteps. TYPE: float default: 1.29
+      MODE_SCALE: 1.29
+      #
+  DIFFUSION_MODEL:
+    # NAME DESCRIPTION:  TYPE:  default: 'Flux'
+    NAME: Flux
+    PRETRAINED_MODEL: ms://AI-ModelScope/FLUX.1-dev@flux1-dev.safetensors
+    # IN_CHANNELS DESCRIPTION: model's input channels. TYPE: int default: 64
+    IN_CHANNELS: 64
+    # HIDDEN_SIZE DESCRIPTION: model's hidden size. TYPE: int default: 1024
+    HIDDEN_SIZE: 3072
+    # NUM_HEADS DESCRIPTION: number of heads in the transformer. TYPE: int default: 16
+    NUM_HEADS: 24
+    # AXES_DIM DESCRIPTION: dimensions of the axes of the positional encoding. TYPE: list default: [16, 56, 56]
+    AXES_DIM: [ 16, 56, 56 ]
+    # THETA DESCRIPTION: theta for positional encoding. TYPE: int default: 10000
+    THETA: 10000
+    # VEC_IN_DIM DESCRIPTION: dimension of the vector input. TYPE: int default: 768
+    VEC_IN_DIM: 768
+    # GUIDANCE_EMBED DESCRIPTION: whether to use guidance embedding. TYPE: bool default: False
+    GUIDANCE_EMBED: True
+    # CONTEXT_IN_DIM DESCRIPTION: dimension of the context input. TYPE: int default: 4096
+    CONTEXT_IN_DIM: 4096
+    # MLP_RATIO DESCRIPTION: ratio of mlp hidden size to hidden size. TYPE: float default: 4.0
+    MLP_RATIO: 4.0
+    # QKV_BIAS DESCRIPTION: whether to use bias in qkv projection. TYPE: bool default: True
+    QKV_BIAS: True
+    # DEPTH DESCRIPTION: number of transformer blocks. TYPE: int default: 19
+    DEPTH: 19
+    # DEPTH_SINGLE_BLOCKS DESCRIPTION: number of transformer blocks in the single stream block. TYPE: int default: 38
+    DEPTH_SINGLE_BLOCKS: 38
+
+  #
+  FIRST_STAGE_MODEL:
+    NAME: AutoencoderKLFlux
+    EMBED_DIM: 16
+    PRETRAINED_MODEL: ms://AI-ModelScope/FLUX.1-dev@ae.safetensors
+    IGNORE_KEYS: [ ]
+    BATCH_SIZE: 8
+    USE_CONV: False
+    SCALE_FACTOR: 0.3611
+    SHIFT_FACTOR: 0.1159
+    #
+    ENCODER:
+      NAME: Encoder
+      USE_CHECKPOINT: True
+      CH: 128
+      OUT_CH: 3
+      NUM_RES_BLOCKS: 2
+      IN_CHANNELS: 3
+      ATTN_RESOLUTIONS: [ ]
+      CH_MULT: [ 1, 2, 4, 4 ]
+      Z_CHANNELS: 16
+      DOUBLE_Z: True
+      DROPOUT: 0.0
+      RESAMP_WITH_CONV: True
+    #
+    DECODER:
+      NAME: Decoder
+      USE_CHECKPOINT: True
+      CH: 128
+      OUT_CH: 3
+      NUM_RES_BLOCKS: 2
+      IN_CHANNELS: 3
+      ATTN_RESOLUTIONS: [ ]
+      CH_MULT: [ 1, 2, 4, 4 ]
+      Z_CHANNELS: 16
+      DROPOUT: 0.0
+      RESAMP_WITH_CONV: True
+      GIVE_PRE_END: False
+      TANH_OUT: False
+  #
+  COND_STAGE_MODEL:
+    # NAME DESCRIPTION:  TYPE:  default: 'T5PlusClipFluxEmbedder'
+    NAME: T5PlusClipFluxEmbedder
+    # T5_MODEL DESCRIPTION:  TYPE:  default: ''
+    T5_MODEL:
+      # NAME DESCRIPTION:  TYPE:  default: 'HFEmbedder'
+      NAME: HFEmbedder
+      # HF_MODEL_CLS DESCRIPTION: huggingface cls in transfomer TYPE: NoneType default: None
+      HF_MODEL_CLS: T5EncoderModel
+      # MODEL_PATH DESCRIPTION: model folder path TYPE: NoneType default: None
+      MODEL_PATH: ms://AI-ModelScope/FLUX.1-dev@text_encoder_2/
+      # HF_TOKENIZER_CLS DESCRIPTION: huggingface cls in transfomer TYPE: NoneType default: None
+      HF_TOKENIZER_CLS: T5Tokenizer
+      # TOKENIZER_PATH DESCRIPTION: tokenizer folder path TYPE: NoneType default: None
+      TOKENIZER_PATH: ms://AI-ModelScope/FLUX.1-dev@tokenizer_2/
+      # MAX_LENGTH DESCRIPTION: max length of input TYPE: int default: 77
+      MAX_LENGTH: 512
+      # OUTPUT_KEY DESCRIPTION: output key TYPE: str default: 'last_hidden_state'
+      OUTPUT_KEY: last_hidden_state
+      # D_TYPE DESCRIPTION: dtype TYPE: str default: 'bfloat16'
+      D_TYPE: bfloat16
+      # BATCH_INFER DESCRIPTION: batch infer TYPE: bool default: False
+      BATCH_INFER: False
+      CLEAN: whitespace
+    # CLIP_MODEL DESCRIPTION:  TYPE:  default: ''
+    CLIP_MODEL:
+      # NAME DESCRIPTION:  TYPE:  default: 'HFEmbedder'
+      NAME: HFEmbedder
+      # HF_MODEL_CLS DESCRIPTION: huggingface cls in transfomer TYPE: NoneType default: None
+      HF_MODEL_CLS: CLIPTextModel
+      # MODEL_PATH DESCRIPTION: model folder path TYPE: NoneType default: None
+      MODEL_PATH: ms://AI-ModelScope/FLUX.1-dev@text_encoder/
+      # HF_TOKENIZER_CLS DESCRIPTION: huggingface cls in transfomer TYPE: NoneType default: None
+      HF_TOKENIZER_CLS: CLIPTokenizer
+      # TOKENIZER_PATH DESCRIPTION: tokenizer folder path TYPE: NoneType default: None
+      TOKENIZER_PATH: ms://AI-ModelScope/FLUX.1-dev@tokenizer/
+      # MAX_LENGTH DESCRIPTION: max length of input TYPE: int default: 77
+      MAX_LENGTH: 77
+      # OUTPUT_KEY DESCRIPTION: output key TYPE: str default: 'last_hidden_state'
+      OUTPUT_KEY: pooler_output
+      # D_TYPE DESCRIPTION: dtype TYPE: str default: 'bfloat16'
+      D_TYPE: bfloat16
+      # BATCH_INFER DESCRIPTION: batch infer TYPE: bool default: False
+      BATCH_INFER: True
+      CLEAN: whitespace

scepter/methods/studio/inference/dit/flux1.0_schnell_pro.yaml

@@ -0,0 +1,212 @@
+NAME: FLUX1.0_SCHNELL
+IS_DEFAULT: False
+DEFAULT_PARAS:
+  PARAS:
+    RESOLUTIONS: [[1024, 1024]]
+  INPUT:
+    IMAGE:
+    ORIGINAL_SIZE_AS_TUPLE: [1024, 1024]
+    TARGET_SIZE_AS_TUPLE: [1024, 1024]
+    PROMPT: ""
+    NEGATIVE_PROMPT:
+      DEFAULT: ""
+      VISIBLE: False
+    PROMPT_PREFIX: ""
+    SAMPLE:
+      VALUES: ["flow_euler"]
+      DEFAULT: "flow_euler"
+    SAMPLE_STEPS: 4
+    GUIDE_SCALE: 3.5
+    GUIDE_RESCALE:
+      DEFAULT: 0.0
+      VISIBLE: False
+    DISCRETIZATION:
+      VALUES: []
+      DEFAULT:
+      VISIBLE: False
+  OUTPUT:
+    LATENT:
+    IMAGES:
+    SEED:
+  MODULES_PARAS:
+    FIRST_STAGE_MODEL:
+      FUNCTION:
+        -
+          NAME: encode
+          DTYPE: bfloat16
+          INPUT: ["IMAGE"]
+        -
+          NAME: decode
+          DTYPE: bfloat16
+          INPUT: ["LATENT"]
+      PARAS:
+        SCALE_FACTOR: 1.5305
+        SHIFT_FACTOR: 0.0609
+        SIZE_FACTOR: 8
+    DIFFUSION_MODEL:
+      FUNCTION:
+        -
+          NAME: forward
+          DTYPE: bfloat16
+          INPUT: ["SAMPLE_STEPS", "SAMPLE", "GUIDE_SCALE"]
+    COND_STAGE_MODEL:
+      FUNCTION:
+        -
+          NAME: encode
+          DTYPE: bfloat16
+          INPUT: ["PROMPT"]
+#
+MODEL:
+  NAME: LatentDiffusionFlux
+  PARAMETERIZATION: rf
+  TIMESTEPS: 1000
+  MIN_SNR_GAMMA:
+  ZERO_TERMINAL_SNR: False
+  PRETRAINED_MODEL:
+  IGNORE_KEYS: [ ]
+  DEFAULT_N_PROMPT:
+  USE_EMA: False
+  EVAL_EMA: False
+  DIFFUSION:
+    # NAME DESCRIPTION:  TYPE:  default: 'DiffusionFluxRF'
+    NAME: DiffusionFluxRF
+    PREDICTION_TYPE: raw
+    # NOISE_SCHEDULER DESCRIPTION:  TYPE:  default: ''
+    NOISE_SCHEDULER:
+      # NAME DESCRIPTION:  TYPE:  default: 'FlowMatchSigmaScheduler'
+      NAME: FlowMatchSigmaScheduler
+      # WEIGHTING_SCHEME DESCRIPTION: The weighting scheme for sampling timesteps, choose from ['sigma_sqrt', 'logit_normal', 'mode', 'cosmap', 'none']. TYPE: str default: 'logit_normal'
+      WEIGHTING_SCHEME: logit_normal
+      SHIFT: 3.0
+      # LOGIT_MEAN DESCRIPTION: The mean of the logit distribution for sampling timesteps. TYPE: float default: 0.0
+      LOGIT_MEAN: 0.0
+      # LOGIT_STD DESCRIPTION: The standard deviation of the logit distribution for sampling timesteps. TYPE: float default: 1.0
+      LOGIT_STD: 1.0
+      # MODE_SCALE DESCRIPTION: The scale factor for the mode of the logit distribution for sampling timesteps. TYPE: float default: 1.29
+      MODE_SCALE: 1.29
+    SAMPLER_SCHEDULER:
+      # NAME DESCRIPTION:  TYPE:  default: 'FlowMatchFluxShiftScheduler'
+      NAME: FlowMatchFluxShiftScheduler
+      # SHIFT DESCRIPTION: Use timestamp shift or not, default is True. TYPE: bool default: True
+      SHIFT: False
+      # SIGMOID_SCALE DESCRIPTION: The scale of sigmoid function for sampling timesteps. TYPE: int default: 1
+      SIGMOID_SCALE: 1
+      # BASE_SHIFT DESCRIPTION: The base shift factor for the timestamp. TYPE: float default: 0.5
+      BASE_SHIFT: 0.5
+      # MAX_SHIFT DESCRIPTION: The max shift factor for the timestamp. TYPE: float default: 1.15
+      MAX_SHIFT: 1.15
+      #
+  DIFFUSION_MODEL:
+    # NAME DESCRIPTION:  TYPE:  default: 'Flux'
+    NAME: Flux
+    PRETRAINED_MODEL: ms://AI-ModelScope/FLUX.1-schnell@flux1-schnell.safetensors
+    # IN_CHANNELS DESCRIPTION: model's input channels. TYPE: int default: 64
+    IN_CHANNELS: 64
+    # HIDDEN_SIZE DESCRIPTION: model's hidden size. TYPE: int default: 1024
+    HIDDEN_SIZE: 3072
+    # NUM_HEADS DESCRIPTION: number of heads in the transformer. TYPE: int default: 16
+    NUM_HEADS: 24
+    # AXES_DIM DESCRIPTION: dimensions of the axes of the positional encoding. TYPE: list default: [16, 56, 56]
+    AXES_DIM: [ 16, 56, 56 ]
+    # THETA DESCRIPTION: theta for positional encoding. TYPE: int default: 10000
+    THETA: 10000
+    # VEC_IN_DIM DESCRIPTION: dimension of the vector input. TYPE: int default: 768
+    VEC_IN_DIM: 768
+    # GUIDANCE_EMBED DESCRIPTION: whether to use guidance embedding. TYPE: bool default: False
+    GUIDANCE_EMBED: False
+    # CONTEXT_IN_DIM DESCRIPTION: dimension of the context input. TYPE: int default: 4096
+    CONTEXT_IN_DIM: 4096
+    # MLP_RATIO DESCRIPTION: ratio of mlp hidden size to hidden size. TYPE: float default: 4.0
+    MLP_RATIO: 4.0
+    # QKV_BIAS DESCRIPTION: whether to use bias in qkv projection. TYPE: bool default: True
+    QKV_BIAS: True
+    # DEPTH DESCRIPTION: number of transformer blocks. TYPE: int default: 19
+    DEPTH: 19
+    # DEPTH_SINGLE_BLOCKS DESCRIPTION: number of transformer blocks in the single stream block. TYPE: int default: 38
+    DEPTH_SINGLE_BLOCKS: 38
+
+  #
+  FIRST_STAGE_MODEL:
+    NAME: AutoencoderKLFlux
+    EMBED_DIM: 16
+    PRETRAINED_MODEL: ms://AI-ModelScope/FLUX.1-schnell@ae.safetensors
+    IGNORE_KEYS: [ ]
+    BATCH_SIZE: 8
+    USE_CONV: False
+    SCALE_FACTOR: 0.3611
+    SHIFT_FACTOR: 0.1159
+    #
+    ENCODER:
+      NAME: Encoder
+      USE_CHECKPOINT: True
+      CH: 128
+      OUT_CH: 3
+      NUM_RES_BLOCKS: 2
+      IN_CHANNELS: 3
+      ATTN_RESOLUTIONS: [ ]
+      CH_MULT: [ 1, 2, 4, 4 ]
+      Z_CHANNELS: 16
+      DOUBLE_Z: True
+      DROPOUT: 0.0
+      RESAMP_WITH_CONV: True
+    #
+    DECODER:
+      NAME: Decoder
+      USE_CHECKPOINT: True
+      CH: 128
+      OUT_CH: 3
+      NUM_RES_BLOCKS: 2
+      IN_CHANNELS: 3
+      ATTN_RESOLUTIONS: [ ]
+      CH_MULT: [ 1, 2, 4, 4 ]
+      Z_CHANNELS: 16
+      DROPOUT: 0.0
+      RESAMP_WITH_CONV: True
+      GIVE_PRE_END: False
+      TANH_OUT: False
+  #
+  COND_STAGE_MODEL:
+    # NAME DESCRIPTION:  TYPE:  default: 'T5PlusClipFluxEmbedder'
+    NAME: T5PlusClipFluxEmbedder
+    # T5_MODEL DESCRIPTION:  TYPE:  default: ''
+    T5_MODEL:
+      # NAME DESCRIPTION:  TYPE:  default: 'HFEmbedder'
+      NAME: HFEmbedder
+      # HF_MODEL_CLS DESCRIPTION: huggingface cls in transfomer TYPE: NoneType default: None
+      HF_MODEL_CLS: T5EncoderModel
+      # MODEL_PATH DESCRIPTION: model folder path TYPE: NoneType default: None
+      MODEL_PATH: ms://AI-ModelScope/FLUX.1-schnell@text_encoder_2/
+      # HF_TOKENIZER_CLS DESCRIPTION: huggingface cls in transfomer TYPE: NoneType default: None
+      HF_TOKENIZER_CLS: T5Tokenizer
+      # TOKENIZER_PATH DESCRIPTION: tokenizer folder path TYPE: NoneType default: None
+      TOKENIZER_PATH: ms://AI-ModelScope/FLUX.1-schnell@tokenizer_2/
+      # MAX_LENGTH DESCRIPTION: max length of input TYPE: int default: 77
+      MAX_LENGTH: 256
+      # OUTPUT_KEY DESCRIPTION: output key TYPE: str default: 'last_hidden_state'
+      OUTPUT_KEY: last_hidden_state
+      # D_TYPE DESCRIPTION: dtype TYPE: str default: 'bfloat16'
+      D_TYPE: bfloat16
+      # BATCH_INFER DESCRIPTION: batch infer TYPE: bool default: False
+      BATCH_INFER: False
+      CLEAN: whitespace
+    # CLIP_MODEL DESCRIPTION:  TYPE:  default: ''
+    CLIP_MODEL:
+      # NAME DESCRIPTION:  TYPE:  default: 'HFEmbedder'
+      NAME: HFEmbedder
+      # HF_MODEL_CLS DESCRIPTION: huggingface cls in transfomer TYPE: NoneType default: None
+      HF_MODEL_CLS: CLIPTextModel
+      # MODEL_PATH DESCRIPTION: model folder path TYPE: NoneType default: None
+      MODEL_PATH: ms://AI-ModelScope/FLUX.1-schnell@text_encoder/
+      # HF_TOKENIZER_CLS DESCRIPTION: huggingface cls in transfomer TYPE: NoneType default: None
+      HF_TOKENIZER_CLS: CLIPTokenizer
+      # TOKENIZER_PATH DESCRIPTION: tokenizer folder path TYPE: NoneType default: None
+      TOKENIZER_PATH: ms://AI-ModelScope/FLUX.1-schnell@tokenizer/
+      # MAX_LENGTH DESCRIPTION: max length of input TYPE: int default: 77
+      MAX_LENGTH: 77
+      # OUTPUT_KEY DESCRIPTION: output key TYPE: str default: 'last_hidden_state'
+      OUTPUT_KEY: pooler_output
+      # D_TYPE DESCRIPTION: dtype TYPE: str default: 'bfloat16'
+      D_TYPE: bfloat16
+      # BATCH_INFER DESCRIPTION: batch infer TYPE: bool default: False
+      BATCH_INFER: True
+      CLEAN: whitespace

scepter/methods/studio/inference/inference.yaml

@@ -1,4 +1,5 @@
 WORK_DIR: "inference"
+SKIP_EXAMPLES: True
 DIFFUSION_PARAS:
   SAMPLE:
     VALUES: ['ddim', 'euler', 'euler_ancestral', 'heun', 'dpm2',
@@ -6,65 +7,91 @@ DIFFUSION_PARAS:
              'dpm2_karras', 'dpm2_ancestral_karras', 'dpmpp_2s_ancestral_karras', 'dpmpp_2m_karras',
              'dpmpp_sde_karras', 'dpmpp_2m_sde_karras']
     DEFAULT: 'dpmpp_2s_ancestral'
+    VISIBLE: True
   NEGATIVE_PROMPT:
     DEFAULT:
+    VISIBLE: True
   PROMPT_PREFIX:
     DEFAULT:
+    VISIBLE: True
   SAMPLES:
     MIN: 1
     MAX: 4
     DEFAULT: 1
+    VISIBLE: True
+  NUM_FRAMES:
+    MIN: 1
+    MAX: 100
+    DEFAULT: 49
+    VISIBLE: False
+  FPS:
+    MIN: 1
+    MAX: 50
+    DEFAULT: 8
+    VISIBLE: False
   SAMPLE_STEPS:
     MIN: 1
     MAX: 100
     DEFAULT: 30
+    VISIBLE: True
   GUIDE_SCALE:
     MIN: 0
     MAX: 10
     DEFAULT: 5.0
+    VISIBLE: True
   GUIDE_RESCALE:
     MIN: 0
     MAX: 1.0
     DEFAULT: 0.5
+    VISIBLE: True
   DISCRETIZATION:
     VALUES: ["trailing", "leading", "linspace"]
     DEFAULT: "linspace"
+    VISIBLE: True
   REFINE_SAMPLERS:
     VALUES: [ 'ddim', 'euler', 'euler_ancestral', 'heun', 'dpm2',
               'dpm2_ancestral', 'dpmpp_2m', 'dpmpp_sde', 'dpmpp_2m_sde', 'dpmpp_2s_ancestral',
               'dpm2_karras', 'dpm2_ancestral_karras', 'dpmpp_2s_ancestral_karras', 'dpmpp_2m_karras',
               'dpmpp_sde_karras', 'dpmpp_2m_sde_karras' ]
     DEFAULT: 'dpmpp_2s_ancestral'
+    VISIBLE: True
   REFINE_SAMPLE_STEPS:
     MIN: 0
     MAX: 100
     DEFAULT: 30
+    VISIBLE: True
   REFINE_GUIDE_SCALE:
     MIN: 0
     MAX: 10
     DEFAULT: 5.0
+    VISIBLE: True
   REFINE_GUIDE_RESCALE:
     MIN: 0
     MAX: 1.0
     DEFAULT: 0.5
+    VISIBLE: True
   REFINE_DISCRETIZATION:
     VALUES: [ "trailing", "leading", "linspace" ]
     DEFAULT: "linspace"
+    VISIBLE: True
   AESTHETIC_SCORE:
     MIN: 0.0
     MAX: 10.0
     DEFAULT: 6.0
+    VISIBLE: True
   NEGATIVE_AESTHETIC_SCORE:
     MIN: 0.0
     MAX: 10.0
     DEFAULT: 2.5
+    VISIBLE: True
   REFINE_STRENGTH:
     MIN: 0
     MAX: 1.0
     DEFAULT: 0.15
   RESOLUTIONS:
-    VALUES: [ [512, 512], [768, 768],
-              [704, 1408], [704, 1344], [768, 1344],
+    VALUES: [
+             [512, 512], [768, 768],
+             [704, 1408], [704, 1344], [768, 1344],
              [720, 1280],
              [768, 1280], [832, 1216], [832, 1152],
              [896, 1152], [896, 1088], [960, 1088],
@@ -74,8 +101,14 @@ DIFFUSION_PARAS:
              [1280, 768],
              [1344, 768], [1344, 704], [1408, 704],
              [1472, 704], [1536, 640], [1600, 640],
-             [1664, 576], [1728, 576]]
+             [1664, 576], [1728, 576],
+             [2048, 2048], [2048, 1920], [1920, 2048],
+             [1536, 2560], [2560, 1536], [2560, 1440],
+             [2560, 1440],
+             [480, 720], [720, 480]
+            ]
     DEFAULT: [1024, 1024]
+    VISIBLE: True
 EXTENSION_PARAS:
   MANTRA_BOOK: scepter/methods/studio/extensions/mantra_book/mantra_book.yaml
   OFFICIAL_TUNERS: scepter/methods/studio/extensions/tuners/official_tuners.yaml

scepter/methods/studio/preprocess/preprocess.yaml

@@ -2,11 +2,10 @@ WORK_DIR: datasets
 EXPORT_DIR: export_datasets
 FILE_SYSTEM:
   -
-    # NAME DESCRIPTION:  TYPE:  default: ''
     NAME: LocalFs
     AUTO_CLEAN: False
-
 PROCESSORS:
+  # Caption processor
   - NAME: BlipImageBase
     TYPE: caption
     MODEL_PATH: ms://cubeai/blip-image-captioning-base
@@ -15,6 +14,81 @@ PROCESSORS:
     PARAS:
       - LANGUAGE_NAME: English
         LANGUAGE_ZH_NAME: 英语
+  - NAME: InternVL15
+    TYPE: caption
+    MODEL_PATH: ms://AI-ModelScope/InternVL-Chat-V1-5
+    DEVICE: "gpu"
+    MEMORY: 49968
+    PARAS:
+      - PROMPT: 用中文描述这张图片
+        LANGUAGE_NAME: Chinese
+        LANGUAGE_ZH_NAME: 中文
+      - PROMPT: Generate the caption in English
+        LANGUAGE_NAME: English
+        LANGUAGE_ZH_NAME: 英语
+  - NAME: QWVLQuantize
+    TYPE: caption
+    DEVICE: "gpu"
+    MEMORY: 7885
+    MODEL_PATH: ms://qwen/Qwen-VL:v1.0.3
+    PARAS:
+      - PROMPT: 用中文描述这张图片
+        LANGUAGE_NAME: Chinese
+        LANGUAGE_ZH_NAME: 中文
+        MAX_NEW_TOKENS:
+          VALUE: 1024
+          MAX: 2048
+          STEP: 128
+          MIN: 256
+        MIN_NEW_TOKENS:
+          VALUE: 16
+          MAX: 1024
+          STEP: 16
+          MIN: 0
+        NUM_BEAMS:
+          VALUE: 1
+          MAX: 12
+          STEP: 1
+          MIN: 1
+        REPETITION_PENALTY:
+          VALUE: 1.0
+          MAX: 100.0
+          STEP: 1.0
+          MIN: 1.0
+        TEMPERATURE:
+          VALUE: 1.0
+          MAX: 100.0
+          STEP: 1.0
+          MIN: 1.0
+      - PROMPT: Generate the caption in English
+        LANGUAGE_NAME: English
+        LANGUAGE_ZH_NAME: 英语
+        MAX_NEW_TOKENS:
+          VALUE: 1024
+          MAX: 2048
+          STEP: 128
+          MIN: 256
+        MIN_NEW_TOKENS:
+          VALUE: 16
+          MAX: 1024
+          STEP: 16
+          MIN: 0
+        NUM_BEAMS:
+          VALUE: 1
+          MAX: 12
+          STEP: 1
+          MIN: 1
+        REPETITION_PENALTY:
+          VALUE: 1.0
+          MAX: 100.0
+          STEP: 1.0
+          MIN: 1.0
+        TEMPERATURE:
+          VALUE: 1.0
+          MAX: 100.0
+          STEP: 1.0
+          MIN: 1.0
+
   - NAME: QWVL
     TYPE: caption
     MODEL_PATH: ms://qwen/Qwen-VL:v1.0.3
@@ -77,75 +151,14 @@ PROCESSORS:
           MAX: 100.0
           STEP: 1.0
           MIN: 1.0
-  -
-    NAME: QWVLQuantize
-    TYPE: caption
-    DEVICE: "gpu"
-    MEMORY: 7885
-    MODEL_PATH: ms://qwen/Qwen-VL:v1.0.3
-    PARAS:
-      - PROMPT: 用中文描述这张图片
-        LANGUAGE_NAME: Chinese
-        LANGUAGE_ZH_NAME: 中文
-        MAX_NEW_TOKENS:
-          VALUE: 1024
-          MAX: 2048
-          STEP: 128
-          MIN: 256
-        MIN_NEW_TOKENS:
-          VALUE: 16
-          MAX: 1024
-          STEP: 16
-          MIN: 0
-        NUM_BEAMS:
-          VALUE: 1
-          MAX: 12
-          STEP: 1
-          MIN: 1
-        REPETITION_PENALTY:
-          VALUE: 1.0
-          MAX: 100.0
-          STEP: 1.0
-          MIN: 1.0
-        TEMPERATURE:
-          VALUE: 1.0
-          MAX: 100.0
-          STEP: 1.0
-          MIN: 1.0
-      - PROMPT: Generate the caption in English
-        LANGUAGE_NAME: English
-        LANGUAGE_ZH_NAME: 英语
-        MAX_NEW_TOKENS:
-          VALUE: 1024
-          MAX: 2048
-          STEP: 128
-          MIN: 256
-        MIN_NEW_TOKENS:
-          VALUE: 16
-          MAX: 1024
-          STEP: 16
-          MIN: 0
-        NUM_BEAMS:
-          VALUE: 1
-          MAX: 12
-          STEP: 1
-          MIN: 1
-        REPETITION_PENALTY:
-          VALUE: 1.0
-          MAX: 100.0
-          STEP: 1.0
-          MIN: 1.0
-        TEMPERATURE:
-          VALUE: 1.0
-          MAX: 100.0
-          STEP: 1.0
-          MIN: 1.0
-  -
-    NAME: CenterCrop
+
+  # Simple processor
+  - NAME: CenterCrop
     TYPE: image
     DEVICE: "cpu"
     MEMORY: 10
     PARAS:
+      CAPTION_INTERACTIVE: False
       HEIGHT_RATIO:
         VALUE: 1
         MAX: 20
@@ -156,18 +169,309 @@ PROCESSORS:
         MAX: 20
         STEP: 1
         MIN: 1
-#  - NAME: PaddingCrop
-#    TYPE: image
-#    DEVICE: "cpu"
-#    MEMORY: 10
-#    PARAS:
-#      HEIGHT_RATIO:
-#        VALUE: 3
-#        MAX: 25
-#        STEP: 1
-#        MIN: 1
-#      WIDTH_RATIO:
-#        VALUE: 4
-#        MAX: 20
-#        STEP: 1
-#        MIN: 1
+  - NAME: ChangeSample
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    PARAS:
+      CAPTION_INTERACTIVE: False
+      SRC_IMAGE_INTERACTIVE: True
+      SRC_IMAGE_MASK_INTERACTIVE: True
+      TARGET_IMAGE_INTERACTIVE: True
+      PREVIEW_BTN_VISIBLE: False
+  - NAME: MaskEditSample
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    PARAS:
+      SRC_IMAGE_INTERACTIVE: True
+      SRC_IMAGE_TOOL: sketch
+      CAPTION_INTERACTIVE: False
+  - NAME: SourceMaskSample
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    PARAS:
+      SRC_IMAGE_INTERACTIVE: True
+      SRC_IMAGE_TOOL: sketch
+      CAPTION_INTERACTIVE: False
+  - NAME: MaskSwapEditSample
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    PARAS:
+      PREVIEW_BTN_VISIBLE: False
+      SRC_IMAGE_INTERACTIVE: True
+      SRC_IMAGE_TOOL: sketch
+      CAPTION_INTERACTIVE: False
+  - NAME: SwapMaskSwapEditSample
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    PARAS:
+      PREVIEW_BTN_VISIBLE: False
+      SRC_IMAGE_INTERACTIVE: True
+      SRC_IMAGE_TOOL: sketch
+      CAPTION_INTERACTIVE: False
+  - NAME: SwapSample
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    PARAS:
+      PREVIEW_BTN_VISIBLE: True
+      SRC_IMAGE_INTERACTIVE: True
+      TARGET_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+  - NAME: DefaultMaskSample
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    PARAS:
+      PREVIEW_BTN_VISIBLE: True
+      SRC_IMAGE_INTERACTIVE: False
+      TARGET_IMAGE_INTERACTIVE: False
+      CAPTION_INTERACTIVE: False
+  #  - NAME: PaddingCrop
+  #    TYPE: image
+  #    DEVICE: "cpu"
+  #    MEMORY: 10
+  #    PARAS:
+  #      HEIGHT_RATIO:
+  #        VALUE: 3
+  #        MAX: 25
+  #        STEP: 1
+  #        MIN: 1
+  #      WIDTH_RATIO:
+  #        VALUE: 4
+  #        MAX: 20
+  #        STEP: 1
+  #        MIN: 1
+  # Annotator processor
+  - NAME: CannyExtractor
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "CannyAnnotator"
+      LOW_THRESHOLD: 100
+      HIGH_THRESHOLD: 200
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+  - NAME: ColorExtractor
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "ColorAnnotator"
+      RATIO: 64
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+  - NAME: InfoDrawContourExtractor
+    TYPE: image
+    DEVICE: "gpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "InfoDrawContourAnnotator"
+      INPUT_NC: 3
+      OUTPUT_NC: 1
+      N_RESIDUAL_BLOCKS: 3
+      SIGMOID: True
+      PRETRAINED_MODEL: "ms://iic/scepter_annotator@annotator/ckpts/informative_drawing_contour_style.pth"
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+  - NAME: DegradationExtractor
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "DegradationAnnotator"
+      RANDOM_DEGRADATION: True
+      PARAMS:
+        gaussian_noise: { }
+        resize: { 'scale': [ 0.4, 0.8 ] }
+        jpeg: { 'jpeg_level': [ 25, 75 ] }
+        gaussian_blur: { 'kernel_size': [ 7, 9, 11, 13, 15 ], 'sigma': [ 0.9, 1.8 ] }
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+  - NAME: MidasExtractor
+    TYPE: image
+    DEVICE: "gpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "MidasDetector"
+      PRETRAINED_MODEL: "ms://iic/scepter_annotator@annotator/ckpts/dpt_hybrid-midas-501f0c75.pt"
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+  - NAME: DoodleExtractor
+    TYPE: image
+    DEVICE: "gpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "DoodleAnnotator"
+      PROCESSOR_TYPE: "pidinet_sketch"
+      PROCESSOR_CFG:
+        - NAME: "PiDiAnnotator"
+          PRETRAINED_MODEL: "ms://iic/scepter_annotator@annotator/ckpts/table5_pidinet.pth"
+        - NAME: "SketchAnnotator"
+          PRETRAINED_MODEL: "ms://iic/scepter_annotator@annotator/ckpts/sketch_simplification_gan.pth"
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+  - NAME: GrayExtractor
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "GrayAnnotator"
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+  - NAME: InpaintingExtractor
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "InpaintingAnnotator"
+      RETURN_MASK: False
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+
+  - NAME: InpaintingSourceExtractor
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+  - NAME: OpenposeExtractor
+    TYPE: image
+    DEVICE: "gpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "OpenposeAnnotator"
+      BODY_MODEL_PATH: "ms://iic/scepter_annotator@annotator/ckpts/body_pose_model.pth"
+      HAND_MODEL_PATH: "ms://iic/scepter_annotator@annotator/ckpts/hand_pose_model.pth"
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+  - NAME: OutpaintingExtractor
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "OutpaintingAnnotator"
+      RETURN_MASK: False
+      KEEP_PADDING_RATIO: 1
+      RANDOM_CFG:
+        DIRECTION_RANGE: [ 'left', 'right', 'up', 'down' ]
+        RATIO_RANGE: [ 0.1, 0.7 ]
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+      USE_MASK_VISIBLE: True
+      PREVIEW_BTN_VISIBLE: False
+  - NAME: OutpaintingResize
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "OutpaintingResize"
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+      PREVIEW_BTN_VISIBLE: False
+  - NAME: InfoDrawAnimeAnnotator
+    TYPE: image
+    DEVICE: "gpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "InfoDrawAnimeAnnotator"
+      INPUT_NC: 3
+      OUTPUT_NC: 1
+      N_RESIDUAL_BLOCKS: 3
+      SIGMOID: True
+      PRETRAINED_MODEL: "ms://iic/scepter_annotator@annotator/ckpts/informative_drawing_anime_style.pth"
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+  - NAME: ESAMExtractor
+    TYPE: image
+    DEVICE: "gpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "ESAMAnnotator"
+      PRETRAINED_MODEL: "ms://iic/scepter_annotator@annotator/ckpts/efficient_sam_vits.pt"
+      SAVE_MODE: 'P'
+      GRID_SIZE: 32
+      USE_DOMINANT_COLOR: True
+      RETURN_MASK: False
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+  - NAME: InvertExtractor
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "InvertAnnotator"
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+
+  - NAME: LamaExtractor
+    TYPE: image
+    DEVICE: "cpu"
+    MEMORY: 10
+    MODEL:
+      NAME: "LamaAnnotator"
+      PRETRAINED_MODEL: "ms:///iic/cv_fft_inpainting_lama/"
+    PARAS:
+      SRC_IMAGE_TOOL: sketch
+      SRC_IMAGE_INTERACTIVE: True
+      CAPTION_INTERACTIVE: False
+
+VIDEO_PROCESSORS:
+  - NAME: CogVLM2Llama3Caption
+    TYPE: caption
+    MODEL_PATH: ms://ZhipuAI/cogvlm2-llama3-caption
+    DEVICE: "gpu"
+    MEMORY: 20000
+    PROMPT: Please describe this video in detail.
+    TEMPERATURE: 0.1
+    MAX_NEW_TOKENS: 2048
+    PAD_TOKEN_ID: 128002
+    TOP_K: 1
+    TOP_P: 0.1
+
+TRANSLATION_PROCESSORS:
+  - NAME: OpusMtZhEn
+    TYPE: caption
+    MODEL_PATH: ms://cubeai/trans-opus-mt-zh-en
+    DEVICE: "gpu"
+    MEMORY: 5000
+  - NAME: OpusMtEnZh
+    TYPE: caption
+    MODEL_PATH: ms://cubeai/trans-opus-mt-en-zh
+    DEVICE: "gpu"
+    MEMORY: 5000

scepter/methods/studio/scepter_ui.yaml

@@ -87,3 +87,7 @@ INTERFACE:
     NAME_EN: Inference
     IFID: inference
     CONFIG: scepter/methods/studio/inference/inference.yaml
+  - NAME: 对话式编辑
+    NAME_EN: ChatBot
+    IFID: chatbot
+    CONFIG: scepter/methods/studio/chatbot/chatbot.yaml

scepter/methods/studio/self_train/dit/cogvideox_2b_pro.yaml

@@ -0,0 +1,315 @@
+ENV:
+  BACKEND: nccl
+  SEED: 42
+  TENSOR_PARALLEL_SIZE: 1
+  PIPELINE_PARALLEL_SIZE: 1
+  SYS_ENVS:
+    TORCH_CUDNN_V8_API_ENABLED: '1'
+    TOKENIZERS_PARALLELISM: 'false'
+    TF_CPP_MIN_LOG_LEVEL: '3'
+    PYTORCH_CUDA_ALLOC_CONF: 'expandable_segments:True'
+META:
+  VERSION: 'COGVIDEOX_2B'
+  DESCRIPTION: "cogvideox 2b"
+  IS_DEFAULT: False
+  IS_SHARE: True
+  INFERENCE_PARAS:
+    INFERENCE_BATCH_SIZE: 1
+    INFERENCE_PREFIX: ""
+    DEFAULT_SAMPLER: "ddim"
+    DEFAULT_SAMPLE_STEPS: 50
+    INFERENCE_N_PROMPT: ""
+    RESOLUTION: [ 480, 720 ]
+  PARAS:
+    - TRAIN_BATCH_SIZE: 1
+      TRAIN_PREFIX: ""
+      TRAIN_N_PROMPT: ""
+      RESOLUTION: [ 480, 720 ]
+      MEMORY: 89000
+      EPOCHS: 50
+      SAVE_INTERVAL: 25
+      EPSEC: 0.818
+      LEARNING_RATE: 4e-4
+      IS_DEFAULT: False
+      TUNER: FULL
+    - TRAIN_BATCH_SIZE: 1
+      TRAIN_PREFIX: ""
+      TRAIN_N_PROMPT: ""
+      RESOLUTION: [ 480, 720 ]
+      MEMORY: 89000
+      EPOCHS: 50
+      SAVE_INTERVAL: 25
+      EPSEC: 0.818
+      LEARNING_RATE: 4e-4
+      IS_DEFAULT: True
+      TUNER: LORA
+  #
+  TUNERS:
+    LORA:
+      - NAME: SwiftLoRA
+        R: 64
+        LORA_ALPHA: 64
+        LORA_DROPOUT: 0.0
+        BIAS: "none"
+        TARGET_MODULES: "model.*(.to_k|.to_q|.to_v|.to_out.0)$"
+#
+SOLVER:
+  NAME: LatentDiffusionVideoSolver
+  MAX_STEPS: 2000
+  USE_AMP: True
+  DTYPE: bfloat16
+  USE_FAIRSCALE: False
+  USE_FSDP: True
+  LOAD_MODEL_ONLY: False
+  ENABLE_GRADSCALER: False
+  USE_SCALER: False
+  RESUME_FROM:
+  WORK_DIR: ./cache/save_data/dit_cogvideox_2b_lora
+  LOG_FILE: std_log.txt
+  EVAL_INTERVAL: 100
+  LOG_TRAIN_NUM: 4
+  FPS: 8
+  SHARDING_STRATEGY: full_shard
+  FSDP_REDUCE_DTYPE: float32
+  FSDP_BUFFER_DTYPE: float32
+  FSDP_SHARD_MODULES: [ 'model', 'cond_stage_model.model']
+  SAVE_MODULES: [ 'model', 'cond_stage_model.model']
+  TRAIN_MODULES: ['model']
+  #
+  FILE_SYSTEM:
+    NAME: "ModelscopeFs"
+    TEMP_DIR: "./cache/cache_data"
+  #
+  TUNER:
+  #
+  MODEL:
+    NAME: LatentDiffusionCogVideoX
+    PRETRAINED_MODEL:
+    PARAMETERIZATION: v
+    TIMESTEPS: 1000
+    MIN_SNR_GAMMA: 3.0
+    ZERO_TERMINAL_SNR: True
+    SCALE_FACTOR_SPATIAL: 8
+    SCALE_FACTOR_TEMPORAL: 4
+    SCALING_FACTOR_IMAGE: 1.15258426
+    IGNORE_KEYS: [ ]
+    DEFAULT_N_PROMPT:
+    USE_EMA: False
+    EVAL_EMA: False
+    DIFFUSION:
+      NAME: BaseDiffusion
+      PREDICTION_TYPE: v
+      NOISE_SCHEDULER:
+        NAME: ScaledLinearScheduler
+        BETA_MIN: 0.00085
+        BETA_MAX: 0.012
+        SNR_SHIFT_SCALE: 3.0
+        RESCALE_BETAS_ZERO_SNR: True
+      DIFFUSION_SAMPLERS:
+        NAME: DDIMSampler
+        DISCRETIZATION_TYPE: trailing
+        ETA: 0.0
+    #
+    DIFFUSION_MODEL:
+      NAME: CogVideoXTransformer3DModel
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: ms://AI-ModelScope/CogVideoX-2b@transformer/diffusion_pytorch_model.safetensors
+      NUM_ATTENTION_HEADS: 30
+      ATTENTION_HEAD_DIM: 64
+      IN_CHANNELS: 16
+      OUT_CHANNELS: 16
+      FLIP_SIN_TO_COS: True
+      FREQ_SHIFT: 0
+      TIME_EMBED_DIM: 512
+      TEXT_EMBED_DIM: 4096
+      NUM_LAYERS: 30
+      DROPOUT: 0.0
+      ATTENTION_BIAS: True
+      SAMPLE_WIDTH: 90
+      SAMPLE_HEIGHT: 60
+      SAMPLE_FRAMES: 49
+      PATCH_SIZE: 2
+      TEMPORAL_COMPRESSION_RATIO: 4
+      MAX_TEXT_SEQ_LENGTH: 226
+      ACTIVATION_FN: "gelu-approximate"
+      TIMESTEP_ACTIVATION_FN: "silu"
+      NORM_ELEMENTWISE_AFFINE: True
+      NORM_EPS: 1e-5
+      SPATIAL_INTERPOLATION_SCALE: 1.875
+      TEMPORAL_INTERPOLATION_SCALE: 1.0
+      USE_ROTARY_POSITIONAL_EMBEDDINGS: False
+      USE_LEARNED_POSITIONAL_EMBEDDINGS: False
+      GRADIENT_CHECKPOINTING: False
+    #
+    FIRST_STAGE_MODEL:
+      NAME: AutoencoderKLCogVideoX
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: ms://AI-ModelScope/CogVideoX-2b@vae/diffusion_pytorch_model.safetensors
+      SAMPLE_HEIGHT: 480
+      SAMPLE_WIDTH: 720
+      USE_QUANT_CONV: False
+      USE_POST_QUANT_CONV: False
+      USE_SLICING: True
+      USE_TILING: True
+      GRADIENT_CHECKPOINTING: False
+      ENCODER:
+        NAME: CogVideoXEncoder3D
+        IN_CHANNELS: 3
+        OUT_CHANNELS: 16
+        UP_BLOCK_TYPES: [ "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: False
+      DECODER:
+        NAME: CogVideoXDecoder3D
+        IN_CHANNELS: 16
+        OUT_CHANNELS: 3
+        UP_BLOCK_TYPES: [ "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: False
+    #
+    COND_STAGE_MODEL:
+      NAME: T5EmbedderHF
+      PRETRAINED_MODEL: ms://AI-ModelScope/t5-v1_1-xxl
+      TOKENIZER_PATH: ms://AI-ModelScope/t5-v1_1-xxl
+      LENGTH: 226
+      CLEAN:
+      USE_GRAD: False
+    #
+    LOSS:
+      NAME: ReconstructLoss
+      LOSS_TYPE: l2
+  #
+  SAMPLE_ARGS:
+    SAMPLER: ddim
+    SAMPLE_STEPS: 50
+    SEED: 42
+    GUIDE_SCALE: 6.0
+    GUIDE_RESCALE: 0.0
+    NUM_FRAMES: 49
+  #
+  OPTIMIZER:
+    NAME: Adam
+    LEARNING_RATE: 1e-3
+    BETAS: [ 0.9, 0.95 ]
+    EPS: 1e-8
+    WEIGHT_DECAY: 0.0
+    AMSGRAD: False
+  #
+#  LR_SCHEDULER:
+#    NAME: StepAnnealingLR
+#    WARMUP_STEPS: 200
+#    TOTAL_STEPS: 2000
+#    DECAY_MODE: 'cosine'
+  #
+  TRAIN_DATA:
+    NAME: VideoGenDatasetOTF
+    MODE: train
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    NUM_WORKERS: 4
+    PROMPT_PREFIX: ''
+    DELIMITER: '#;#'
+    FIELDS: [ 'video_path', 'width', 'height', 'prompt' ]
+    PATH_PREFIX:
+    DATA_FILE:
+    SAMPLER:
+      NAME: LoopSampler
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ 'video', 'video_latent', "prompt" ]
+        META_KEYS: [ ]
+    MODEL:
+      NAME: AutoencoderKLCogVideoX
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: ms://AI-ModelScope/CogVideoX-2b@vae/diffusion_pytorch_model.safetensors
+      SAMPLE_HEIGHT: 480
+      SAMPLE_WIDTH: 720
+      USE_QUANT_CONV: False
+      USE_POST_QUANT_CONV: False
+      USE_SLICING: True
+      USE_TILING: True
+      GRADIENT_CHECKPOINTING: True
+      ENCODER:
+        NAME: CogVideoXEncoder3D
+        IN_CHANNELS: 3
+        OUT_CHANNELS: 16
+        UP_BLOCK_TYPES: [ "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+      DECODER:
+        NAME: CogVideoXDecoder3D
+        IN_CHANNELS: 16
+        OUT_CHANNELS: 3
+        UP_BLOCK_TYPES: [ "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+  #
+  EVAL_DATA:
+    NAME: Text2ImageDataset
+    MODE: eval
+    PROMPT_FILE:
+    PROMPT_DATA: [ "A girl riding a bike.", "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical atmosphere of this unique musical performance." ]
+    IMAGE_SIZE: [ 480, 720 ]
+    FIELDS: [ "prompt" ]
+    DELIMITER: '#;#'
+    PROMPT_PREFIX: ''
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+#    USE_NUM: 8
+    NUM_WORKERS: 4
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ 'index', 'prompt' ]
+        META_KEYS: [ 'image_size' ]
+  #
+  TRAIN_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0
+    - NAME: BackwardHook
+      PRIORITY: 10
+    - NAME: LogHook
+      LOG_INTERVAL: 10
+      PRIORITY: 20
+    - NAME: CheckpointHook
+      INTERVAL: 1000
+      PRIORITY: 40
+      SAVE_LAST: True
+      SAVE_NAME_PREFIX: 'step'
+      DISABLE_SNAPSHOT: True
+  #
+  EVAL_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0
+      SAVE_LAST: True
+      SAVE_NAME_PREFIX: 'step'
+      SAVE_PROBE_PREFIX: 'image'

scepter/methods/studio/self_train/dit/cogvideox_5b_pro.yaml

@@ -0,0 +1,317 @@
+ENV:
+  BACKEND: nccl
+  SEED: 42
+  TENSOR_PARALLEL_SIZE: 1
+  PIPELINE_PARALLEL_SIZE: 1
+  SYS_ENVS:
+    TORCH_CUDNN_V8_API_ENABLED: '1'
+    TOKENIZERS_PARALLELISM: 'false'
+    TF_CPP_MIN_LOG_LEVEL: '3'
+    PYTORCH_CUDA_ALLOC_CONF: 'expandable_segments:True'
+META:
+  VERSION: 'COGVIDEOX_5B'
+  DESCRIPTION: "cogvideox 5b"
+  IS_DEFAULT: False
+  IS_SHARE: True
+  INFERENCE_PARAS:
+    INFERENCE_BATCH_SIZE: 1
+    INFERENCE_PREFIX: ""
+    DEFAULT_SAMPLER: "ddim"
+    DEFAULT_SAMPLE_STEPS: 50
+    INFERENCE_N_PROMPT: ""
+    RESOLUTION: [ 480, 720 ]
+  PARAS:
+    - TRAIN_BATCH_SIZE: 1
+      TRAIN_PREFIX: ""
+      TRAIN_N_PROMPT: ""
+      RESOLUTION: [ 480, 720 ]
+      MEMORY: 89000
+      EPOCHS: 50
+      SAVE_INTERVAL: 25
+      EPSEC: 0.818
+      LEARNING_RATE: 4e-4
+      IS_DEFAULT: False
+      TUNER: FULL
+    - TRAIN_BATCH_SIZE: 1
+      TRAIN_PREFIX: ""
+      TRAIN_N_PROMPT: ""
+      RESOLUTION: [ 480, 720 ]
+      MEMORY: 89000
+      EPOCHS: 50
+      SAVE_INTERVAL: 25
+      EPSEC: 0.818
+      LEARNING_RATE: 4e-4
+      IS_DEFAULT: True
+      TUNER: LORA
+  #
+  TUNERS:
+    LORA:
+      - NAME: SwiftLoRA
+        R: 64
+        LORA_ALPHA: 64
+        LORA_DROPOUT: 0.0
+        BIAS: "none"
+        TARGET_MODULES: "model.*(.to_k|.to_q|.to_v|.to_out.0)$"
+#
+SOLVER:
+  NAME: LatentDiffusionVideoSolver
+  MAX_STEPS: 2000
+  USE_AMP: True
+  DTYPE: bfloat16
+  USE_FAIRSCALE: False
+  USE_FSDP: True
+  LOAD_MODEL_ONLY: False
+  ENABLE_GRADSCALER: False
+  USE_SCALER: False
+  RESUME_FROM:
+  WORK_DIR: ./cache/save_data/dit_cogvideox_5b_lora
+  LOG_FILE: std_log.txt
+  EVAL_INTERVAL: 100
+  LOG_TRAIN_NUM: 4
+  FPS: 8
+  SHARDING_STRATEGY: full_shard
+  FSDP_REDUCE_DTYPE: float32
+  FSDP_BUFFER_DTYPE: float32
+  FSDP_SHARD_MODULES: [ 'model', 'cond_stage_model.model']
+  SAVE_MODULES: [ 'model', 'cond_stage_model.model']
+  TRAIN_MODULES: ['model']
+  #
+  FILE_SYSTEM:
+    NAME: "ModelscopeFs"
+    TEMP_DIR: "./cache/cache_data"
+  #
+  TUNER:
+  #
+  MODEL:
+    NAME: LatentDiffusionCogVideoX
+    PRETRAINED_MODEL:
+    PARAMETERIZATION: v
+    TIMESTEPS: 1000
+    MIN_SNR_GAMMA: 3.0
+    ZERO_TERMINAL_SNR: True
+    SCALE_FACTOR_SPATIAL: 8
+    SCALE_FACTOR_TEMPORAL: 4
+    SCALING_FACTOR_IMAGE: 0.7  # 5b diff
+    IGNORE_KEYS: [ ]
+    DEFAULT_N_PROMPT:
+    USE_EMA: False
+    EVAL_EMA: False
+    DIFFUSION:
+      NAME: BaseDiffusion
+      PREDICTION_TYPE: v
+      NOISE_SCHEDULER:
+        NAME: ScaledLinearScheduler
+        BETA_MIN: 0.00085
+        BETA_MAX: 0.012
+        SNR_SHIFT_SCALE: 1.0  # 5b diff
+        RESCALE_BETAS_ZERO_SNR: True
+      DIFFUSION_SAMPLERS:
+        NAME: DDIMSampler
+        DISCRETIZATION_TYPE: trailing
+        ETA: 0.0
+    #
+    DIFFUSION_MODEL:
+      NAME: CogVideoXTransformer3DModel
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: # 5b diff
+        - ms://AI-ModelScope/CogVideoX-5b@transformer/diffusion_pytorch_model-00001-of-00002.safetensors
+        - ms://AI-ModelScope/CogVideoX-5b@transformer/diffusion_pytorch_model-00002-of-00002.safetensors
+      NUM_ATTENTION_HEADS: 48  # 5b diff
+      ATTENTION_HEAD_DIM: 64
+      IN_CHANNELS: 16
+      OUT_CHANNELS: 16
+      FLIP_SIN_TO_COS: True
+      FREQ_SHIFT: 0
+      TIME_EMBED_DIM: 512
+      TEXT_EMBED_DIM: 4096
+      NUM_LAYERS: 42  # 5b diff
+      DROPOUT: 0.0
+      ATTENTION_BIAS: True
+      SAMPLE_WIDTH: 90
+      SAMPLE_HEIGHT: 60
+      SAMPLE_FRAMES: 49
+      PATCH_SIZE: 2
+      TEMPORAL_COMPRESSION_RATIO: 4
+      MAX_TEXT_SEQ_LENGTH: 226
+      ACTIVATION_FN: "gelu-approximate"
+      TIMESTEP_ACTIVATION_FN: "silu"
+      NORM_ELEMENTWISE_AFFINE: True
+      NORM_EPS: 1e-5
+      SPATIAL_INTERPOLATION_SCALE: 1.875
+      TEMPORAL_INTERPOLATION_SCALE: 1.0
+      USE_ROTARY_POSITIONAL_EMBEDDINGS: True # 5b diff
+      USE_LEARNED_POSITIONAL_EMBEDDINGS: False
+      GRADIENT_CHECKPOINTING: True
+    #
+    FIRST_STAGE_MODEL:
+      NAME: AutoencoderKLCogVideoX
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: ms://AI-ModelScope/CogVideoX-5b@vae/diffusion_pytorch_model.safetensors # 5b diff
+      SAMPLE_HEIGHT: 480
+      SAMPLE_WIDTH: 720
+      USE_QUANT_CONV: False
+      USE_POST_QUANT_CONV: False
+      USE_SLICING: True
+      USE_TILING: True
+      GRADIENT_CHECKPOINTING: True
+      ENCODER:
+        NAME: CogVideoXEncoder3D
+        IN_CHANNELS: 3
+        OUT_CHANNELS: 16
+        UP_BLOCK_TYPES: [ "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+      DECODER:
+        NAME: CogVideoXDecoder3D
+        IN_CHANNELS: 16
+        OUT_CHANNELS: 3
+        UP_BLOCK_TYPES: [ "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+    #
+    COND_STAGE_MODEL:
+      NAME: T5EmbedderHF
+      PRETRAINED_MODEL: ms://AI-ModelScope/t5-v1_1-xxl
+      TOKENIZER_PATH: ms://AI-ModelScope/t5-v1_1-xxl
+      LENGTH: 226
+      CLEAN:
+      USE_GRAD: False
+    #
+    LOSS:
+      NAME: ReconstructLoss
+      LOSS_TYPE: l2
+  #
+  SAMPLE_ARGS:
+    SAMPLER: ddim
+    SAMPLE_STEPS: 50
+    SEED: 42
+    GUIDE_SCALE: 6.0
+    GUIDE_RESCALE: 0.0
+    NUM_FRAMES: 49
+  #
+  OPTIMIZER:
+    NAME: Adam
+    LEARNING_RATE: 1e-3
+    BETAS: [ 0.9, 0.95 ]
+    EPS: 1e-8
+    WEIGHT_DECAY: 0.0
+    AMSGRAD: False
+  #
+#  LR_SCHEDULER:
+#    NAME: StepAnnealingLR
+#    WARMUP_STEPS: 200
+#    TOTAL_STEPS: 2000
+#    DECAY_MODE: 'cosine'
+  #
+  TRAIN_DATA:
+    NAME: VideoGenDatasetOTF
+    MODE: train
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    NUM_WORKERS: 4
+    PROMPT_PREFIX: ''
+    DELIMITER: '#;#'
+    FIELDS: [ 'video_path', 'width', 'height', 'prompt' ]
+    PATH_PREFIX:
+    DATA_FILE:
+    SAMPLER:
+      NAME: LoopSampler
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ 'video', 'video_latent', "prompt" ]
+        META_KEYS: [ ]
+    MODEL:
+      NAME: AutoencoderKLCogVideoX
+      DTYPE: bfloat16
+      PRETRAINED_MODEL: ms://AI-ModelScope/CogVideoX-5b@vae/diffusion_pytorch_model.safetensors
+      SAMPLE_HEIGHT: 480
+      SAMPLE_WIDTH: 720
+      USE_QUANT_CONV: False
+      USE_POST_QUANT_CONV: False
+      USE_SLICING: True
+      USE_TILING: True
+      GRADIENT_CHECKPOINTING: True
+      ENCODER:
+        NAME: CogVideoXEncoder3D
+        IN_CHANNELS: 3
+        OUT_CHANNELS: 16
+        UP_BLOCK_TYPES: [ "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D", "CogVideoXDownBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+      DECODER:
+        NAME: CogVideoXDecoder3D
+        IN_CHANNELS: 16
+        OUT_CHANNELS: 3
+        UP_BLOCK_TYPES: [ "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D", "CogVideoXUpBlock3D" ]
+        BLOCK_OUT_CHANNELS: [ 128, 256, 256, 512 ]
+        LAYERS_PER_BLOCK: 3
+        ACT_FN: "silu"
+        NORM_EPS: 1e-6
+        NORM_NUM_GROUPS: 32
+        DROPOUT: 0.0
+        PAD_MODE: "first"
+        TEMPORAL_COMPRESSION_RATIO: 4
+        GRADIENT_CHECKPOINTING: True
+  #
+  EVAL_DATA:
+    NAME: Text2ImageDataset
+    MODE: eval
+    PROMPT_FILE:
+    PROMPT_DATA: [ "A girl riding a bike.", "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical atmosphere of this unique musical performance." ]
+    IMAGE_SIZE: [ 480, 720 ]
+    FIELDS: [ "prompt" ]
+    DELIMITER: '#;#'
+    PROMPT_PREFIX: ''
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    #    USE_NUM: 8
+    NUM_WORKERS: 4
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ 'index', 'prompt' ]
+        META_KEYS: [ 'image_size' ]
+  #
+  TRAIN_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0
+    - NAME: BackwardHook
+      PRIORITY: 10
+    - NAME: LogHook
+      LOG_INTERVAL: 10
+      PRIORITY: 20
+    - NAME: CheckpointHook
+      INTERVAL: 1000
+      PRIORITY: 40
+      SAVE_LAST: True
+      SAVE_NAME_PREFIX: 'step'
+      DISABLE_SNAPSHOT: True
+  #
+  EVAL_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0
+      SAVE_LAST: True
+      SAVE_NAME_PREFIX: 'step'
+      SAVE_PROBE_PREFIX: 'image'

scepter/methods/studio/self_train/dit/flux1.0_dv_pro.yaml

@@ -0,0 +1,285 @@
+ENV:
+  BACKEND: nccl
+META:
+  VERSION: 'FLUX1.0_DEV'
+  DESCRIPTION: "flux 1.0 dev"
+  IS_DEFAULT: True
+  IS_SHARE: True
+  INFERENCE_PARAS:
+    INFERENCE_BATCH_SIZE: 1
+    INFERENCE_PREFIX: ""
+    DEFAULT_SAMPLER: "flow_euler"
+    DEFAULT_SAMPLE_STEPS: 50
+    INFERENCE_N_PROMPT: ""
+    RESOLUTION: [1024, 1024]
+  PARAS:
+    -
+      TRAIN_BATCH_SIZE: 1
+      TRAIN_PREFIX: ""
+      TRAIN_N_PROMPT: ""
+      RESOLUTION: [1024, 1024]
+      MEMORY: 89000
+      EPOCHS: 50
+      SAVE_INTERVAL: 25
+      EPSEC: 0.818
+      LEARNING_RATE: 4e-4
+      IS_DEFAULT: False
+      TUNER: FULL
+    -
+      TRAIN_BATCH_SIZE: 1
+      TRAIN_PREFIX: ""
+      TRAIN_N_PROMPT: ""
+      RESOLUTION: [1024, 1024]
+      MEMORY: 89000
+      EPOCHS: 50
+      SAVE_INTERVAL: 25
+      EPSEC: 0.818
+      LEARNING_RATE: 4e-4
+      IS_DEFAULT: True
+      TUNER: LORA
+  #
+  TUNERS:
+    LORA:
+      -
+        NAME: SwiftLoRA
+        R: 4
+        LORA_ALPHA: 4
+        LORA_DROPOUT: 0.0
+        BIAS: "none"
+        TARGET_MODULES: "(model.double_blocks.*(.qkv|.proj|.img_mod.lin|.txt_mod.lin))|(model.single_blocks.*(.linear1|.linear2|.modulation.lin))$"
+#
+SOLVER:
+  NAME: LatentDiffusionSolver
+  MAX_STEPS: 100000
+  USE_AMP: True
+  DTYPE: bfloat16
+  USE_FAIRSCALE: False
+  USE_FSDP: True
+  LOAD_MODEL_ONLY: False
+  RESUME_FROM:
+  WORK_DIR: ./cache/save_data/dit_flux_dev_1024_lora
+  LOG_FILE: std_log.txt
+  EVAL_INTERVAL: 100
+  LOG_TRAIN_NUM: 16
+  ENABLE_GRADSCALER: False
+  USE_SCALER: False
+  FSDP_REDUCE_DTYPE: float32
+  FSDP_BUFFER_DTYPE: float32
+  FSDP_SHARD_MODULES: [ 'model', 'cond_stage_model.t5_model' ] #
+  SAVE_MODULES: [ 'model']
+  TRAIN_MODULES: ['model']
+
+  FILE_SYSTEM:
+    NAME: "ModelscopeFs"
+    TEMP_DIR: "./cache/cache_data"
+
+  FREEZE:
+
+  TUNER:
+
+  MODEL:
+    NAME: LatentDiffusionFlux
+    PARAMETERIZATION: rf
+    TIMESTEPS: 1000
+    MIN_SNR_GAMMA:
+    ZERO_TERMINAL_SNR: False
+    PRETRAINED_MODEL:
+    IGNORE_KEYS: [ ]
+    DEFAULT_N_PROMPT:
+    USE_EMA: False
+    EVAL_EMA: False
+    DIFFUSION:
+      NAME: DiffusionFluxRF
+      PREDICTION_TYPE: raw
+      NOISE_SCHEDULER:
+        NAME: FlowMatchSigmaScheduler
+        WEIGHTING_SCHEME: logit_normal
+        SHIFT: 3.0
+        LOGIT_MEAN: 0.0
+        LOGIT_STD: 1.0
+        MODE_SCALE: 1.29
+      SAMPLER_SCHEDULER:
+        NAME: FlowMatchFluxShiftScheduler
+        SHIFT: False
+        SIGMOID_SCALE: 1
+        BASE_SHIFT: 0.5
+        MAX_SHIFT: 1.15
+
+    DIFFUSION_MODEL:
+      NAME: Flux
+      PRETRAINED_MODEL: ms://AI-ModelScope/FLUX.1-dev@flux1-dev.safetensors
+      IN_CHANNELS: 64
+      HIDDEN_SIZE: 3072
+      NUM_HEADS: 24
+      AXES_DIM: [ 16, 56, 56 ]
+      THETA: 10000
+      VEC_IN_DIM: 768
+      GUIDANCE_EMBED: False
+      CONTEXT_IN_DIM: 4096
+      MLP_RATIO: 4.0
+      QKV_BIAS: True
+      DEPTH: 19
+      DEPTH_SINGLE_BLOCKS: 38
+      USE_GRAD_CHECKPOINT: True
+
+    FIRST_STAGE_MODEL:
+      NAME: AutoencoderKLFlux
+      EMBED_DIM: 16
+      PRETRAINED_MODEL: ms://AI-ModelScope/FLUX.1-dev@ae.safetensors
+      IGNORE_KEYS: [ ]
+      BATCH_SIZE: 8
+      USE_CONV: False
+      SCALE_FACTOR: 0.3611
+      SHIFT_FACTOR: 0.1159
+
+      ENCODER:
+        NAME: Encoder
+        USE_CHECKPOINT: True
+        CH: 128
+        OUT_CH: 3
+        NUM_RES_BLOCKS: 2
+        IN_CHANNELS: 3
+        ATTN_RESOLUTIONS: [ ]
+        CH_MULT: [ 1, 2, 4, 4 ]
+        Z_CHANNELS: 16
+        DOUBLE_Z: True
+        DROPOUT: 0.0
+        RESAMP_WITH_CONV: True
+
+      DECODER:
+        NAME: Decoder
+        USE_CHECKPOINT: True
+        CH: 128
+        OUT_CH: 3
+        NUM_RES_BLOCKS: 2
+        IN_CHANNELS: 3
+        ATTN_RESOLUTIONS: [ ]
+        CH_MULT: [ 1, 2, 4, 4 ]
+        Z_CHANNELS: 16
+        DROPOUT: 0.0
+        RESAMP_WITH_CONV: True
+        GIVE_PRE_END: False
+        TANH_OUT: False
+
+    COND_STAGE_MODEL:
+      NAME: T5PlusClipFluxEmbedder
+      T5_MODEL:
+        NAME: HFEmbedder
+        HF_MODEL_CLS: T5EncoderModel
+        MODEL_PATH: ms://AI-ModelScope/FLUX.1-dev@text_encoder_2/
+        HF_TOKENIZER_CLS: T5Tokenizer
+        TOKENIZER_PATH: ms://AI-ModelScope/FLUX.1-dev@tokenizer_2/
+        MAX_LENGTH: 512
+        OUTPUT_KEY: last_hidden_state
+        D_TYPE: bfloat16
+        BATCH_INFER: False
+        CLEAN: whitespace
+      CLIP_MODEL:
+        NAME: HFEmbedder
+        HF_MODEL_CLS: CLIPTextModel
+        MODEL_PATH: ms://AI-ModelScope/FLUX.1-dev@text_encoder/
+        HF_TOKENIZER_CLS: CLIPTokenizer
+        TOKENIZER_PATH: ms://AI-ModelScope/FLUX.1-dev@tokenizer/
+        MAX_LENGTH: 77
+        OUTPUT_KEY: pooler_output
+        D_TYPE: bfloat16
+        BATCH_INFER: True
+        CLEAN: whitespace
+
+  SAMPLE_ARGS:
+    SAMPLE_STEPS: 50
+    SAMPLER: flow_euler
+    SEED: 2024
+    IMAGE_SIZE: [ 1024, 1024 ]
+    SHIFT: True
+    GUIDE_SCALE: 3.5
+
+  OPTIMIZER:
+    NAME: AdamW
+    LEARNING_RATE: 4e-4
+    BETAS: [ 0.9, 0.999 ]
+    EPS: 1e-8
+    WEIGHT_DECAY: 1e-2
+    AMSGRAD: False
+
+  TRAIN_DATA:
+    NAME: ImageTextPairMSDataset
+    MODE: train
+    MS_DATASET_NAME: style_custom_dataset
+    MS_DATASET_NAMESPACE: damo
+    MS_DATASET_SUBNAME: 3D
+    PROMPT_PREFIX: ""
+    MS_DATASET_SPLIT: train
+    MS_REMAP_KEYS: { 'Image:FILE': 'Target:FILE' }
+    REPLACE_STYLE: False
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    NUM_WORKERS: 4
+    SAMPLER:
+      NAME: LoopSampler
+    TRANSFORMS:
+      - NAME: LoadImageFromFile
+        RGB_ORDER: RGB
+        BACKEND: pillow
+      - NAME: FlexibleResize
+        INTERPOLATION: bilinear
+        SIZE: [ 1024, 1024 ]
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'img' ]
+        BACKEND: pillow
+      - NAME: FlexibleCenterCrop
+        SIZE: [ 1024, 1024 ]
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'img' ]
+        BACKEND: pillow
+      - NAME: ImageToTensor
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'img' ]
+        BACKEND: pillow
+      - NAME: Normalize
+        MEAN: [ 0.5,  0.5,  0.5 ]
+        STD: [ 0.5,  0.5,  0.5 ]
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'image' ]
+        BACKEND: torchvision
+      - NAME: Select
+        KEYS: [ 'image', 'prompt' ]
+        META_KEYS: [ 'data_key' ]
+
+  EVAL_DATA:
+    NAME: Text2ImageDataset
+    MODE: eval
+    PROMPT_FILE:
+    PROMPT_DATA: [ "a cat holds a blackboard that writes \"hello world\"", "a dog running on the lawn" ]
+    IMAGE_SIZE: [ 1024, 1024 ]
+    FIELDS: [ "prompt" ]
+    DELIMITER: '#;#'
+    PROMPT_PREFIX: ''
+    PIN_MEMORY: True
+    BATCH_SIZE: 2
+    NUM_WORKERS: 4
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ 'index', 'prompt' ]
+        META_KEYS: [ 'image_size' ]
+
+  TRAIN_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0
+    - NAME: BackwardHook
+      PRIORITY: 10
+    - NAME: LogHook
+      LOG_INTERVAL: 10
+    - NAME: CheckpointHook
+      INTERVAL: 10000
+      PRIORITY: 200
+      SAVE_LAST: True
+      SAVE_NAME_PREFIX: 'step'
+      DISABLE_SNAPSHOT: True
+  EVAL_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      SAVE_LAST: True
+      SAVE_NAME_PREFIX: 'step'
+      SAVE_PROBE_PREFIX: 'image'

scepter/methods/studio/self_train/dit/flux1.0_schnell_pro.yaml

@@ -0,0 +1,284 @@
+ENV:
+  BACKEND: nccl
+META:
+  VERSION: 'FLUX1.0_SCHNELL'
+  DESCRIPTION: "flux 1.0 schnell"
+  IS_DEFAULT: False
+  IS_SHARE: True
+  INFERENCE_PARAS:
+    INFERENCE_BATCH_SIZE: 1
+    INFERENCE_PREFIX: ""
+    DEFAULT_SAMPLER: "flow_euler"
+    DEFAULT_SAMPLE_STEPS: 4
+    INFERENCE_N_PROMPT: ""
+    RESOLUTION: [1024, 1024]
+  PARAS:
+    -
+      TRAIN_BATCH_SIZE: 1
+      TRAIN_PREFIX: ""
+      TRAIN_N_PROMPT: ""
+      RESOLUTION: [1024, 1024]
+      MEMORY: 89000
+      EPOCHS: 50
+      SAVE_INTERVAL: 25
+      EPSEC: 0.818
+      LEARNING_RATE: 4e-4
+      IS_DEFAULT: False
+      TUNER: FULL
+    -
+      TRAIN_BATCH_SIZE: 1
+      TRAIN_PREFIX: ""
+      TRAIN_N_PROMPT: ""
+      RESOLUTION: [1024, 1024]
+      MEMORY: 89000
+      EPOCHS: 50
+      SAVE_INTERVAL: 25
+      EPSEC: 0.818
+      LEARNING_RATE: 4e-4
+      IS_DEFAULT: True
+      TUNER: LORA
+  #
+  TUNERS:
+    LORA:
+      -
+        NAME: SwiftLoRA
+        R: 4
+        LORA_ALPHA: 4
+        LORA_DROPOUT: 0.0
+        BIAS: "none"
+        TARGET_MODULES: "(model.double_blocks.*(.qkv|.proj|.img_mod.lin|.txt_mod.lin))|(model.single_blocks.*(.linear1|.linear2|.modulation.lin))$"
+#
+SOLVER:
+  NAME: LatentDiffusionSolver
+  MAX_STEPS: 100000
+  USE_AMP: True
+  DTYPE: bfloat16
+  USE_FAIRSCALE: False
+  USE_FSDP: True
+  LOAD_MODEL_ONLY: False
+  RESUME_FROM:
+  WORK_DIR: ./cache/save_data/dit_flux_schnell_1024_lora
+  LOG_FILE: std_log.txt
+  EVAL_INTERVAL: 100
+  LOG_TRAIN_NUM: 16
+  ENABLE_GRADSCALER: False
+  USE_SCALER: False
+  FSDP_REDUCE_DTYPE: float32
+  FSDP_BUFFER_DTYPE: float32
+  FSDP_SHARD_MODULES: [ 'model', 'cond_stage_model.t5_model' ]
+  SAVE_MODULES: [ 'model']
+  TRAIN_MODULES: ['model']
+  #
+  FILE_SYSTEM:
+    NAME: "ModelscopeFs"
+    TEMP_DIR: "./cache/cache_data"
+
+  FREEZE:
+
+  TUNER:
+
+  MODEL:
+    NAME: LatentDiffusionFlux
+    PARAMETERIZATION: rf
+    TIMESTEPS: 1000
+    MIN_SNR_GAMMA:
+    ZERO_TERMINAL_SNR: False
+    PRETRAINED_MODEL:
+    IGNORE_KEYS: [ ]
+    DEFAULT_N_PROMPT:
+    USE_EMA: False
+    EVAL_EMA: False
+    DIFFUSION:
+      NAME: DiffusionFluxRF
+      PREDICTION_TYPE: raw
+      NOISE_SCHEDULER:
+        NAME: FlowMatchSigmaScheduler
+        WEIGHTING_SCHEME: logit_normal
+        SHIFT: 3.0
+        LOGIT_MEAN: 0.0
+        LOGIT_STD: 1.0
+        MODE_SCALE: 1.29
+      SAMPLER_SCHEDULER:
+        NAME: FlowMatchFluxShiftScheduler
+        SHIFT: False
+        SIGMOID_SCALE: 1
+        BASE_SHIFT: 0.5
+        MAX_SHIFT: 1.15
+
+    DIFFUSION_MODEL:
+      NAME: Flux
+      PRETRAINED_MODEL: ms://AI-ModelScope/FLUX.1-schnell@flux1-schnell.safetensors
+      IN_CHANNELS: 64
+      HIDDEN_SIZE: 3072
+      NUM_HEADS: 24
+      AXES_DIM: [ 16, 56, 56 ]
+      THETA: 10000
+      VEC_IN_DIM: 768
+      GUIDANCE_EMBED: False
+      CONTEXT_IN_DIM: 4096
+      MLP_RATIO: 4.0
+      QKV_BIAS: True
+      DEPTH: 19
+      DEPTH_SINGLE_BLOCKS: 38
+      USE_GRAD_CHECKPOINT: True
+
+    FIRST_STAGE_MODEL:
+      NAME: AutoencoderKLFlux
+      EMBED_DIM: 16
+      PRETRAINED_MODEL: ms://AI-ModelScope/FLUX.1-schnell@ae.safetensors
+      IGNORE_KEYS: [ ]
+      BATCH_SIZE: 8
+      USE_CONV: False
+      SCALE_FACTOR: 0.3611
+      SHIFT_FACTOR: 0.1159
+
+      ENCODER:
+        NAME: Encoder
+        USE_CHECKPOINT: True
+        CH: 128
+        OUT_CH: 3
+        NUM_RES_BLOCKS: 2
+        IN_CHANNELS: 3
+        ATTN_RESOLUTIONS: [ ]
+        CH_MULT: [ 1, 2, 4, 4 ]
+        Z_CHANNELS: 16
+        DOUBLE_Z: True
+        DROPOUT: 0.0
+        RESAMP_WITH_CONV: True
+
+      DECODER:
+        NAME: Decoder
+        USE_CHECKPOINT: True
+        CH: 128
+        OUT_CH: 3
+        NUM_RES_BLOCKS: 2
+        IN_CHANNELS: 3
+        ATTN_RESOLUTIONS: [ ]
+        CH_MULT: [ 1, 2, 4, 4 ]
+        Z_CHANNELS: 16
+        DROPOUT: 0.0
+        RESAMP_WITH_CONV: True
+        GIVE_PRE_END: False
+        TANH_OUT: False
+
+    COND_STAGE_MODEL:
+      NAME: T5PlusClipFluxEmbedder
+      T5_MODEL:
+        NAME: HFEmbedder
+        HF_MODEL_CLS: T5EncoderModel
+        MODEL_PATH: ms://AI-ModelScope/FLUX.1-schnell@text_encoder_2/
+        HF_TOKENIZER_CLS: T5Tokenizer
+        TOKENIZER_PATH: ms://AI-ModelScope/FLUX.1-schnell@tokenizer_2/
+        MAX_LENGTH: 256
+        OUTPUT_KEY: last_hidden_state
+        D_TYPE: bfloat16
+        BATCH_INFER: False
+        CLEAN: whitespace
+      CLIP_MODEL:
+        NAME: HFEmbedder
+        HF_MODEL_CLS: CLIPTextModel
+        MODEL_PATH: ms://AI-ModelScope/FLUX.1-schnell@text_encoder/
+        HF_TOKENIZER_CLS: CLIPTokenizer
+        TOKENIZER_PATH: ms://AI-ModelScope/FLUX.1-schnell@tokenizer/
+        MAX_LENGTH: 77
+        OUTPUT_KEY: pooler_output
+        D_TYPE: bfloat16
+        BATCH_INFER: True
+        CLEAN: whitespace
+
+  SAMPLE_ARGS:
+    SAMPLE_STEPS: 4
+    SAMPLER: flow_euler
+    SEED: 2024
+    IMAGE_SIZE: [ 1024, 1024 ]
+    GUIDE_SCALE: 3.5
+
+  OPTIMIZER:
+    NAME: AdamW
+    LEARNING_RATE: 4e-4
+    BETAS: [ 0.9, 0.999 ]
+    EPS: 1e-8
+    WEIGHT_DECAY: 1e-2
+    AMSGRAD: False
+
+  TRAIN_DATA:
+    NAME: ImageTextPairMSDataset
+    MODE: train
+    MS_DATASET_NAME: style_custom_dataset
+    MS_DATASET_NAMESPACE: damo
+    MS_DATASET_SUBNAME: 3D
+    PROMPT_PREFIX: ""
+    MS_DATASET_SPLIT: train
+    MS_REMAP_KEYS: { 'Image:FILE': 'Target:FILE' }
+    REPLACE_STYLE: False
+    PIN_MEMORY: True
+    BATCH_SIZE: 1
+    NUM_WORKERS: 4
+    SAMPLER:
+      NAME: LoopSampler
+    TRANSFORMS:
+      - NAME: LoadImageFromFile
+        RGB_ORDER: RGB
+        BACKEND: pillow
+      - NAME: FlexibleResize
+        INTERPOLATION: bilinear
+        SIZE: [ 1024, 1024 ]
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'img' ]
+        BACKEND: pillow
+      - NAME: FlexibleCenterCrop
+        SIZE: [ 1024, 1024 ]
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'img' ]
+        BACKEND: pillow
+      - NAME: ImageToTensor
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'img' ]
+        BACKEND: pillow
+      - NAME: Normalize
+        MEAN: [ 0.5,  0.5,  0.5 ]
+        STD: [ 0.5,  0.5,  0.5 ]
+        INPUT_KEY: [ 'img' ]
+        OUTPUT_KEY: [ 'image' ]
+        BACKEND: torchvision
+      - NAME: Select
+        KEYS: [ 'image', 'prompt' ]
+        META_KEYS: [ 'data_key' ]
+
+  EVAL_DATA:
+    NAME: Text2ImageDataset
+    MODE: eval
+    PROMPT_FILE:
+    PROMPT_DATA: [ "a cat holds a blackboard that writes \"hello world\"", "a dog running on the lawn" ]
+    IMAGE_SIZE: [ 1024, 1024 ]
+    FIELDS: [ "prompt" ]
+    DELIMITER: '#;#'
+    PROMPT_PREFIX: ''
+    PIN_MEMORY: True
+    BATCH_SIZE: 2
+    NUM_WORKERS: 4
+    TRANSFORMS:
+      - NAME: Select
+        KEYS: [ 'index', 'prompt' ]
+        META_KEYS: [ 'image_size' ]
+
+  TRAIN_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      PRIORITY: 0
+    - NAME: BackwardHook
+      PRIORITY: 10
+    - NAME: LogHook
+      LOG_INTERVAL: 10
+    - NAME: CheckpointHook
+      INTERVAL: 10000
+      PRIORITY: 200
+      SAVE_LAST: True
+      SAVE_NAME_PREFIX: 'step'
+      DISABLE_SNAPSHOT: True
+  EVAL_HOOKS:
+    - NAME: ProbeDataHook
+      PROB_INTERVAL: 100
+      SAVE_LAST: True
+      SAVE_NAME_PREFIX: 'step'
+      SAVE_PROBE_PREFIX: 'image'

scepter/methods/studio/self_train/dit/pixart_alpha_pro.yaml

@@ -35,7 +35,7 @@ META:
       SAVE_INTERVAL: 25
       EPSEC: 0.818
       LEARNING_RATE: 0.0001
-      IS_DEFAULT: False
+      IS_DEFAULT: True
       TUNER: LORA
   #
   TUNERS:

scepter/methods/studio/self_train/self_train.yaml

@@ -13,8 +13,10 @@ TRAIN_PARAS:
     VALUES: [[256, 256], [320, 180], [180, 320],
              [512, 512], [640, 360], [360, 640],
              [768, 768], [960, 540], [540, 960],
-             [1024, 1024], [1280, 720], [720, 1280]]
+             [1024, 1024], [1280, 720], [720, 1280],
+             [720, 480], [480, 720]]
     DEFAULT: [1024, 1024]
   EVAL_PROMPTS:
     - a boy wearing a jacket
     - a dog running on the lawn
+SAVE_FILE_LOCAL_PATH: "cache/scepter_ui/datasets/train_data_from_list"

scepter/methods/studio/tuner_manager/tuner_manager.yaml

@@ -13,3 +13,5 @@ BASE_MODEL_VERSION:
     TUNER_TYPE: [ 'LORA', 'SCE', 'FULL' ]
   - BASE_MODEL: 'EDIT'
     TUNER_TYPE: [ 'LORA', 'SCE', 'FULL' ]
+  - BASE_MODEL: 'FLUX1.0_DEV'
+    TUNER_TYPE: [ 'LORA' ]

scepter/modules/__init__.py

@@ -1,4 +1,23 @@
 # -*- coding: utf-8 -*-
 # Copyright (c) Alibaba, Inc. and its affiliates.
-from scepter.modules import (data, inference, model, opt, solver, transform,
-                             utils)
+from typing import TYPE_CHECKING
+from scepter.modules.utils.import_utils import LazyImportModule
+
+
+if TYPE_CHECKING:
+    from scepter.modules import (data, inference, model, opt, solver, transform,
+                                 utils)
+else:
+    _import_structure = {
+        'modules': ['data', 'inference', 'model', 'opt', 'solver',
+                    'transform', 'utils']
+    }
+
+    import sys
+    sys.modules[__name__] = LazyImportModule(
+        __name__,
+        globals()['__file__'],
+        _import_structure,
+        module_spec=__spec__,
+        extra_objects={},
+    )

scepter/modules/annotator/__init__.py

@@ -1,11 +1,64 @@
 # -*- coding: utf-8 -*-
 # Copyright (c) Alibaba, Inc. and its affiliates.
-from scepter.modules.annotator.base_annotator import GeneralAnnotator
-from scepter.modules.annotator.canny import CannyAnnotator
-from scepter.modules.annotator.color import ColorAnnotator
-from scepter.modules.annotator.hed import HedAnnotator
-from scepter.modules.annotator.identity import IdentityAnnotator
-from scepter.modules.annotator.invert import InvertAnnotator
-from scepter.modules.annotator.midas_op import MidasDetector
-from scepter.modules.annotator.mlsd_op import MLSDdetector
-from scepter.modules.annotator.openpose import OpenposeAnnotator
+from typing import TYPE_CHECKING
+from scepter.modules.utils.import_utils import LazyImportModule
+
+if TYPE_CHECKING:
+    from scepter.modules.annotator.base_annotator import GeneralAnnotator
+    from scepter.modules.annotator.canny import CannyAnnotator
+    from scepter.modules.annotator.color import ColorAnnotator
+    from scepter.modules.annotator.degradation import DegradationAnnotator
+    from scepter.modules.annotator.doodle import DoodleAnnotator
+    from scepter.modules.annotator.gray import GrayAnnotator
+    from scepter.modules.annotator.hed import HedAnnotator
+    from scepter.modules.annotator.identity import IdentityAnnotator
+    from scepter.modules.annotator.informative_drawing import (
+        InfoDrawAnimeAnnotator, InfoDrawContourAnnotator,
+        InfoDrawOpenSketchAnnotator)
+    from scepter.modules.annotator.inpainting import InpaintingAnnotator
+    from scepter.modules.annotator.invert import InvertAnnotator
+    from scepter.modules.annotator.midas_op import MidasDetector
+    from scepter.modules.annotator.mlsd_op import MLSDdetector
+    from scepter.modules.annotator.openpose import OpenposeAnnotator
+    from scepter.modules.annotator.outpainting import OutpaintingAnnotator, OutpaintingResize
+    from scepter.modules.annotator.pidinet import PiDiAnnotator
+    from scepter.modules.annotator.segmentation import ESAMAnnotator
+    from scepter.modules.annotator.sketch import SketchAnnotator
+    from scepter.modules.annotator.lama import LamaAnnotator
+    from scepter.modules.annotator.mask_aug import MaskAugAnnotator, MaskDrawAnnotator, MaskLayoutAnnotator
+    from scepter.modules.annotator.raft import RAFTAnnotator, RAFTVisAnnotator
+else:
+    _import_structure = {
+        'base_annotator': ['GeneralAnnotator'],
+        'canny': ['CannyAnnotator'],
+        'color': ['ColorAnnotator'],
+        'degradation': ['DegradationAnnotator'],
+        'doodle': ['DoodleAnnotator'],
+        'gray': ['GrayAnnotator'],
+        'hed': ['HedAnnotator'],
+        'identity': ['IdentityAnnotator'],
+        'informative_drawing': ['InfoDrawAnimeAnnotator',
+                                'InfoDrawContourAnnotator',
+                                'InfoDrawOpenSketchAnnotator'],
+        'inpainting': ['InpaintingAnnotator'],
+        'invert': ['InvertAnnotator'],
+        'midas_op': ['MidasDetector'],
+        'mlsd_op': ['MLSDdetector'],
+        'openpose': ['OpenposeAnnotator'],
+        'outpainting': ['OutpaintingAnnotator', 'OutpaintingResize'],
+        'pidinet': ['PiDiAnnotator'],
+        'segmentation': ['ESAMAnnotator'],
+        'sketch': ['SketchAnnotator'],
+        'lama': ['LamaAnnotator'],
+        'mask_aug': ['MaskAugAnnotator', 'MaskDrawAnnotator', 'MaskLayoutAnnotator'],
+        'raft': ['RAFTAnnotator', 'RAFTVisAnnotator'],
+    }
+
+    import sys
+    sys.modules[__name__] = LazyImportModule(
+        __name__,
+        globals()['__file__'],
+        _import_structure,
+        module_spec=__spec__,
+        extra_objects={},
+    )

scepter/modules/annotator/degradation.py

@@ -0,0 +1,135 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import math
+import random
+from abc import ABCMeta
+
+import cv2
+import numpy as np
+import torch
+
+from PIL import Image
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import Config, dict_to_yaml
+
+
+def gaussian_noise_op(im, v):
+    from basicsr.data.degradations import random_add_gaussian_noise
+    noise_level = v.get('noise_level', [10, 20])
+    out = random_add_gaussian_noise(
+        im,
+        sigma_range=noise_level,
+        clip=True,
+        rounds=False,
+        gray_prob=0.4,
+    )
+    out = np.clip(out, 0.0, 1.0)
+    return out
+
+
+def resize_op(im, v):
+    scale = v.get('scale', [0.5, 0.8])
+    h, w = im.shape[:2]
+    scale = random.uniform(scale[0], scale[1])
+    h_, w_ = int(h * scale), int(w * scale)
+    mode = v.get('mode', 'nearest')
+    if mode == 'nearest':
+        interpolation = cv2.INTER_NEAREST
+    elif mode == 'bilinear':
+        interpolation = cv2.INTER_LINEAR
+    elif mode == 'bicubic':
+        interpolation = cv2.INTER_CUBIC
+    else:
+        interpolation = cv2.INTER_NEAREST
+    im = cv2.resize(im, (w_, h_), interpolation=interpolation)
+    out = cv2.resize(im, (w, h), interpolation=interpolation)
+    out = np.clip(out, 0.0, 1.0)
+    return out
+
+
+def jpeg_op(im, v):
+    from basicsr.data.degradations import add_jpg_compression
+    jpeg_level = v.get('jpeg_level', [50, 75])
+    v = int(random.uniform(jpeg_level[0], jpeg_level[1]))
+    out = add_jpg_compression(im, v)
+    out = np.clip(out, 0.0, 1.0)
+    return out
+
+
+def gaussian_blur_op(im, v):
+    from basicsr.data.degradations import random_mixed_kernels
+    kernel_range = v.get('kernel_size', [7, 9])
+    kernel_size = random.choice(kernel_range)
+    kernel_size = min(int(kernel_size) // 2 * 2 + 1, 21)
+    blur_sigma = v.get('sigma', [0.9, 1.0])
+    kernel = random_mixed_kernels(
+        ('iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso',
+         'plateau_aniso'), (0.45, 0.25, 0.12, 0.03, 0.12, 0.03),
+        kernel_size,
+        blur_sigma,
+        blur_sigma, [-math.pi, math.pi], [0.5, 2.0], [1, 1.5],
+        noise_range=None)
+
+    pad_size = (21 - kernel_size) // 2
+    kernel = np.pad(kernel, ((pad_size, pad_size), (pad_size, pad_size)))
+    out = cv2.filter2D(im, -1, kernel)
+    out = np.clip(out, 0.0, 1.0)
+    return out
+
+
+@ANNOTATORS.register_class()
+class DegradationAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        self.params = cfg.get('PARAMS', {
+            'gaussian_noise': {},
+            'resize': {},
+            'jpeg': {},
+            'gaussian_blur': {},
+        })
+        if not isinstance(self.params, dict):
+            self.params = Config.get_dict(self.params)
+        self.random_degradation = cfg.get('RANDOM_DEGRADATION', False)
+
+    def forward(self, image):
+        if isinstance(image, Image.Image):
+            image = np.array(image)
+        elif isinstance(image, torch.Tensor):
+            image = image.detach().cpu().numpy()
+        elif isinstance(image, np.ndarray):
+            image = image.copy()
+        else:
+            raise f'Unsurpport datatype{type(image)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+        if np.max(image) > 1.0:
+            image = (image / 255.).astype(np.float32)
+
+        degradation_list = list(self.params.keys())
+        if self.random_degradation:
+            random.shuffle(degradation_list)
+
+        for degradation_type in degradation_list:
+            if degradation_type == 'gaussian_noise':
+                image = gaussian_noise_op(image, self.params[degradation_type])
+            elif degradation_type == 'resize':
+                image = resize_op(image, self.params[degradation_type])
+            elif degradation_type == 'jpeg':
+                image = jpeg_op(image, self.params[degradation_type])
+            elif degradation_type == 'gaussian_blur':
+                image = gaussian_blur_op(image, self.params[degradation_type])
+            else:
+                raise NotImplementedError(
+                    f'ERROR: degradation_type: {degradation_type} is invalid.')
+        image = (image * 255.0).astype(np.uint8)
+
+        assert len(image.shape) < 4
+        return image
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('ANNOTATORS',
+                            __class__.__name__,
+                            DegradationAnnotator.para_dict,
+                            set_name=True)

scepter/modules/annotator/doodle.py

@@ -0,0 +1,42 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+from abc import ABCMeta
+
+import torch
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import dict_to_yaml
+
+
+@ANNOTATORS.register_class()
+class DoodleAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        self.processor_type = cfg.get('PROCESSOR_TYPE', 'pidinet_sketch')
+        processor_cfg = cfg.get('PROCESSOR_CFG', None)
+        if self.processor_type == 'pidinet_sketch':
+            self.pidinet_ins = ANNOTATORS.build(processor_cfg[0])
+            self.sketch_ins = ANNOTATORS.build(processor_cfg[1])
+        else:
+            raise 'Unsurpport PROCESSOR for DoodleAnnotator'
+
+    @torch.no_grad()
+    @torch.inference_mode()
+    @torch.autocast('cuda', enabled=False)
+    def forward(self, image):
+        if self.processor_type == 'pidinet_sketch':
+            pidinet_res = self.pidinet_ins(image)
+            sketch_res = self.sketch_ins(pidinet_res)
+            doodle_res = sketch_res
+        else:
+            raise 'Unsurpport PROCESSOR for DoodleAnnotator'
+        return doodle_res
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('ANNOTATORS',
+                            __class__.__name__,
+                            DoodleAnnotator.para_dict,
+                            set_name=True)

scepter/modules/annotator/dwpose/__init__.py

@@ -0,0 +1,2 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.

scepter/modules/annotator/dwpose/onnxdet.py

@@ -0,0 +1,127 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import cv2
+import numpy as np
+
+import onnxruntime
+
+def nms(boxes, scores, nms_thr):
+    """Single class NMS implemented in Numpy."""
+    x1 = boxes[:, 0]
+    y1 = boxes[:, 1]
+    x2 = boxes[:, 2]
+    y2 = boxes[:, 3]
+
+    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
+    order = scores.argsort()[::-1]
+
+    keep = []
+    while order.size > 0:
+        i = order[0]
+        keep.append(i)
+        xx1 = np.maximum(x1[i], x1[order[1:]])
+        yy1 = np.maximum(y1[i], y1[order[1:]])
+        xx2 = np.minimum(x2[i], x2[order[1:]])
+        yy2 = np.minimum(y2[i], y2[order[1:]])
+
+        w = np.maximum(0.0, xx2 - xx1 + 1)
+        h = np.maximum(0.0, yy2 - yy1 + 1)
+        inter = w * h
+        ovr = inter / (areas[i] + areas[order[1:]] - inter)
+
+        inds = np.where(ovr <= nms_thr)[0]
+        order = order[inds + 1]
+
+    return keep
+
+def multiclass_nms(boxes, scores, nms_thr, score_thr):
+    """Multiclass NMS implemented in Numpy. Class-aware version."""
+    final_dets = []
+    num_classes = scores.shape[1]
+    for cls_ind in range(num_classes):
+        cls_scores = scores[:, cls_ind]
+        valid_score_mask = cls_scores > score_thr
+        if valid_score_mask.sum() == 0:
+            continue
+        else:
+            valid_scores = cls_scores[valid_score_mask]
+            valid_boxes = boxes[valid_score_mask]
+            keep = nms(valid_boxes, valid_scores, nms_thr)
+            if len(keep) > 0:
+                cls_inds = np.ones((len(keep), 1)) * cls_ind
+                dets = np.concatenate(
+                    [valid_boxes[keep], valid_scores[keep, None], cls_inds], 1
+                )
+                final_dets.append(dets)
+    if len(final_dets) == 0:
+        return None
+    return np.concatenate(final_dets, 0)
+
+def demo_postprocess(outputs, img_size, p6=False):
+    grids = []
+    expanded_strides = []
+    strides = [8, 16, 32] if not p6 else [8, 16, 32, 64]
+
+    hsizes = [img_size[0] // stride for stride in strides]
+    wsizes = [img_size[1] // stride for stride in strides]
+
+    for hsize, wsize, stride in zip(hsizes, wsizes, strides):
+        xv, yv = np.meshgrid(np.arange(wsize), np.arange(hsize))
+        grid = np.stack((xv, yv), 2).reshape(1, -1, 2)
+        grids.append(grid)
+        shape = grid.shape[:2]
+        expanded_strides.append(np.full((*shape, 1), stride))
+
+    grids = np.concatenate(grids, 1)
+    expanded_strides = np.concatenate(expanded_strides, 1)
+    outputs[..., :2] = (outputs[..., :2] + grids) * expanded_strides
+    outputs[..., 2:4] = np.exp(outputs[..., 2:4]) * expanded_strides
+
+    return outputs
+
+def preprocess(img, input_size, swap=(2, 0, 1)):
+    if len(img.shape) == 3:
+        padded_img = np.ones((input_size[0], input_size[1], 3), dtype=np.uint8) * 114
+    else:
+        padded_img = np.ones(input_size, dtype=np.uint8) * 114
+
+    r = min(input_size[0] / img.shape[0], input_size[1] / img.shape[1])
+    resized_img = cv2.resize(
+        img,
+        (int(img.shape[1] * r), int(img.shape[0] * r)),
+        interpolation=cv2.INTER_LINEAR,
+    ).astype(np.uint8)
+    padded_img[: int(img.shape[0] * r), : int(img.shape[1] * r)] = resized_img
+
+    padded_img = padded_img.transpose(swap)
+    padded_img = np.ascontiguousarray(padded_img, dtype=np.float32)
+    return padded_img, r
+
+def inference_detector(session, oriImg):
+    input_shape = (640,640)
+    img, ratio = preprocess(oriImg, input_shape)
+
+    ort_inputs = {session.get_inputs()[0].name: img[None, :, :, :]}
+    output = session.run(None, ort_inputs)
+    predictions = demo_postprocess(output[0], input_shape)[0]
+
+    boxes = predictions[:, :4]
+    scores = predictions[:, 4:5] * predictions[:, 5:]
+
+    boxes_xyxy = np.ones_like(boxes)
+    boxes_xyxy[:, 0] = boxes[:, 0] - boxes[:, 2]/2.
+    boxes_xyxy[:, 1] = boxes[:, 1] - boxes[:, 3]/2.
+    boxes_xyxy[:, 2] = boxes[:, 0] + boxes[:, 2]/2.
+    boxes_xyxy[:, 3] = boxes[:, 1] + boxes[:, 3]/2.
+    boxes_xyxy /= ratio
+    dets = multiclass_nms(boxes_xyxy, scores, nms_thr=0.45, score_thr=0.1)
+    if dets is not None:
+        final_boxes, final_scores, final_cls_inds = dets[:, :4], dets[:, 4], dets[:, 5]
+        isscore = final_scores>0.3
+        iscat = final_cls_inds == 0
+        isbbox = [ i and j for (i, j) in zip(isscore, iscat)]
+        final_boxes = final_boxes[isbbox]
+    else:
+        final_boxes = np.array([])
+
+    return final_boxes

scepter/modules/annotator/dwpose/onnxpose.py

@@ -0,0 +1,362 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+from typing import List, Tuple
+
+import cv2
+import numpy as np
+import onnxruntime as ort
+
+def preprocess(
+    img: np.ndarray, out_bbox, input_size: Tuple[int, int] = (192, 256)
+) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """Do preprocessing for RTMPose model inference.
+
+    Args:
+        img (np.ndarray): Input image in shape.
+        input_size (tuple): Input image size in shape (w, h).
+
+    Returns:
+        tuple:
+        - resized_img (np.ndarray): Preprocessed image.
+        - center (np.ndarray): Center of image.
+        - scale (np.ndarray): Scale of image.
+    """
+    # get shape of image
+    img_shape = img.shape[:2]
+    out_img, out_center, out_scale = [], [], []
+    if len(out_bbox) == 0:
+        out_bbox = [[0, 0, img_shape[1], img_shape[0]]]
+    for i in range(len(out_bbox)):
+        x0 = out_bbox[i][0]
+        y0 = out_bbox[i][1]
+        x1 = out_bbox[i][2]
+        y1 = out_bbox[i][3]
+        bbox = np.array([x0, y0, x1, y1])
+
+        # get center and scale
+        center, scale = bbox_xyxy2cs(bbox, padding=1.25)
+
+        # do affine transformation
+        resized_img, scale = top_down_affine(input_size, scale, center, img)
+
+        # normalize image
+        mean = np.array([123.675, 116.28, 103.53])
+        std = np.array([58.395, 57.12, 57.375])
+        resized_img = (resized_img - mean) / std
+
+        out_img.append(resized_img)
+        out_center.append(center)
+        out_scale.append(scale)
+
+    return out_img, out_center, out_scale
+
+
+def inference(sess: ort.InferenceSession, img: np.ndarray) -> np.ndarray:
+    """Inference RTMPose model.
+
+    Args:
+        sess (ort.InferenceSession): ONNXRuntime session.
+        img (np.ndarray): Input image in shape.
+
+    Returns:
+        outputs (np.ndarray): Output of RTMPose model.
+    """
+    all_out = []
+    # build input
+    for i in range(len(img)):
+        input = [img[i].transpose(2, 0, 1)]
+
+        # build output
+        sess_input = {sess.get_inputs()[0].name: input}
+        sess_output = []
+        for out in sess.get_outputs():
+            sess_output.append(out.name)
+
+        # run model
+        outputs = sess.run(sess_output, sess_input)
+        all_out.append(outputs)
+
+    return all_out
+
+
+def postprocess(outputs: List[np.ndarray],
+                model_input_size: Tuple[int, int],
+                center: Tuple[int, int],
+                scale: Tuple[int, int],
+                simcc_split_ratio: float = 2.0
+                ) -> Tuple[np.ndarray, np.ndarray]:
+    """Postprocess for RTMPose model output.
+
+    Args:
+        outputs (np.ndarray): Output of RTMPose model.
+        model_input_size (tuple): RTMPose model Input image size.
+        center (tuple): Center of bbox in shape (x, y).
+        scale (tuple): Scale of bbox in shape (w, h).
+        simcc_split_ratio (float): Split ratio of simcc.
+
+    Returns:
+        tuple:
+        - keypoints (np.ndarray): Rescaled keypoints.
+        - scores (np.ndarray): Model predict scores.
+    """
+    all_key = []
+    all_score = []
+    for i in range(len(outputs)):
+        # use simcc to decode
+        simcc_x, simcc_y = outputs[i]
+        keypoints, scores = decode(simcc_x, simcc_y, simcc_split_ratio)
+
+        # rescale keypoints
+        keypoints = keypoints / model_input_size * scale[i] + center[i] - scale[i] / 2
+        all_key.append(keypoints[0])
+        all_score.append(scores[0])
+
+    return np.array(all_key), np.array(all_score)
+
+
+def bbox_xyxy2cs(bbox: np.ndarray,
+                 padding: float = 1.) -> Tuple[np.ndarray, np.ndarray]:
+    """Transform the bbox format from (x,y,w,h) into (center, scale)
+
+    Args:
+        bbox (ndarray): Bounding box(es) in shape (4,) or (n, 4), formatted
+            as (left, top, right, bottom)
+        padding (float): BBox padding factor that will be multilied to scale.
+            Default: 1.0
+
+    Returns:
+        tuple: A tuple containing center and scale.
+        - np.ndarray[float32]: Center (x, y) of the bbox in shape (2,) or
+            (n, 2)
+        - np.ndarray[float32]: Scale (w, h) of the bbox in shape (2,) or
+            (n, 2)
+    """
+    # convert single bbox from (4, ) to (1, 4)
+    dim = bbox.ndim
+    if dim == 1:
+        bbox = bbox[None, :]
+
+    # get bbox center and scale
+    x1, y1, x2, y2 = np.hsplit(bbox, [1, 2, 3])
+    center = np.hstack([x1 + x2, y1 + y2]) * 0.5
+    scale = np.hstack([x2 - x1, y2 - y1]) * padding
+
+    if dim == 1:
+        center = center[0]
+        scale = scale[0]
+
+    return center, scale
+
+
+def _fix_aspect_ratio(bbox_scale: np.ndarray,
+                      aspect_ratio: float) -> np.ndarray:
+    """Extend the scale to match the given aspect ratio.
+
+    Args:
+        scale (np.ndarray): The image scale (w, h) in shape (2, )
+        aspect_ratio (float): The ratio of ``w/h``
+
+    Returns:
+        np.ndarray: The reshaped image scale in (2, )
+    """
+    w, h = np.hsplit(bbox_scale, [1])
+    bbox_scale = np.where(w > h * aspect_ratio,
+                          np.hstack([w, w / aspect_ratio]),
+                          np.hstack([h * aspect_ratio, h]))
+    return bbox_scale
+
+
+def _rotate_point(pt: np.ndarray, angle_rad: float) -> np.ndarray:
+    """Rotate a point by an angle.
+
+    Args:
+        pt (np.ndarray): 2D point coordinates (x, y) in shape (2, )
+        angle_rad (float): rotation angle in radian
+
+    Returns:
+        np.ndarray: Rotated point in shape (2, )
+    """
+    sn, cs = np.sin(angle_rad), np.cos(angle_rad)
+    rot_mat = np.array([[cs, -sn], [sn, cs]])
+    return rot_mat @ pt
+
+
+def _get_3rd_point(a: np.ndarray, b: np.ndarray) -> np.ndarray:
+    """To calculate the affine matrix, three pairs of points are required. This
+    function is used to get the 3rd point, given 2D points a & b.
+
+    The 3rd point is defined by rotating vector `a - b` by 90 degrees
+    anticlockwise, using b as the rotation center.
+
+    Args:
+        a (np.ndarray): The 1st point (x,y) in shape (2, )
+        b (np.ndarray): The 2nd point (x,y) in shape (2, )
+
+    Returns:
+        np.ndarray: The 3rd point.
+    """
+    direction = a - b
+    c = b + np.r_[-direction[1], direction[0]]
+    return c
+
+
+def get_warp_matrix(center: np.ndarray,
+                    scale: np.ndarray,
+                    rot: float,
+                    output_size: Tuple[int, int],
+                    shift: Tuple[float, float] = (0., 0.),
+                    inv: bool = False) -> np.ndarray:
+    """Calculate the affine transformation matrix that can warp the bbox area
+    in the input image to the output size.
+
+    Args:
+        center (np.ndarray[2, ]): Center of the bounding box (x, y).
+        scale (np.ndarray[2, ]): Scale of the bounding box
+            wrt [width, height].
+        rot (float): Rotation angle (degree).
+        output_size (np.ndarray[2, ] | list(2,)): Size of the
+            destination heatmaps.
+        shift (0-100%): Shift translation ratio wrt the width/height.
+            Default (0., 0.).
+        inv (bool): Option to inverse the affine transform direction.
+            (inv=False: src->dst or inv=True: dst->src)
+
+    Returns:
+        np.ndarray: A 2x3 transformation matrix
+    """
+    shift = np.array(shift)
+    src_w = scale[0]
+    dst_w = output_size[0]
+    dst_h = output_size[1]
+
+    # compute transformation matrix
+    rot_rad = np.deg2rad(rot)
+    src_dir = _rotate_point(np.array([0., src_w * -0.5]), rot_rad)
+    dst_dir = np.array([0., dst_w * -0.5])
+
+    # get four corners of the src rectangle in the original image
+    src = np.zeros((3, 2), dtype=np.float32)
+    src[0, :] = center + scale * shift
+    src[1, :] = center + src_dir + scale * shift
+    src[2, :] = _get_3rd_point(src[0, :], src[1, :])
+
+    # get four corners of the dst rectangle in the input image
+    dst = np.zeros((3, 2), dtype=np.float32)
+    dst[0, :] = [dst_w * 0.5, dst_h * 0.5]
+    dst[1, :] = np.array([dst_w * 0.5, dst_h * 0.5]) + dst_dir
+    dst[2, :] = _get_3rd_point(dst[0, :], dst[1, :])
+
+    if inv:
+        warp_mat = cv2.getAffineTransform(np.float32(dst), np.float32(src))
+    else:
+        warp_mat = cv2.getAffineTransform(np.float32(src), np.float32(dst))
+
+    return warp_mat
+
+
+def top_down_affine(input_size: dict, bbox_scale: dict, bbox_center: dict,
+                    img: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
+    """Get the bbox image as the model input by affine transform.
+
+    Args:
+        input_size (dict): The input size of the model.
+        bbox_scale (dict): The bbox scale of the img.
+        bbox_center (dict): The bbox center of the img.
+        img (np.ndarray): The original image.
+
+    Returns:
+        tuple: A tuple containing center and scale.
+        - np.ndarray[float32]: img after affine transform.
+        - np.ndarray[float32]: bbox scale after affine transform.
+    """
+    w, h = input_size
+    warp_size = (int(w), int(h))
+
+    # reshape bbox to fixed aspect ratio
+    bbox_scale = _fix_aspect_ratio(bbox_scale, aspect_ratio=w / h)
+
+    # get the affine matrix
+    center = bbox_center
+    scale = bbox_scale
+    rot = 0
+    warp_mat = get_warp_matrix(center, scale, rot, output_size=(w, h))
+
+    # do affine transform
+    img = cv2.warpAffine(img, warp_mat, warp_size, flags=cv2.INTER_LINEAR)
+
+    return img, bbox_scale
+
+
+def get_simcc_maximum(simcc_x: np.ndarray,
+                      simcc_y: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
+    """Get maximum response location and value from simcc representations.
+
+    Note:
+        instance number: N
+        num_keypoints: K
+        heatmap height: H
+        heatmap width: W
+
+    Args:
+        simcc_x (np.ndarray): x-axis SimCC in shape (K, Wx) or (N, K, Wx)
+        simcc_y (np.ndarray): y-axis SimCC in shape (K, Wy) or (N, K, Wy)
+
+    Returns:
+        tuple:
+        - locs (np.ndarray): locations of maximum heatmap responses in shape
+            (K, 2) or (N, K, 2)
+        - vals (np.ndarray): values of maximum heatmap responses in shape
+            (K,) or (N, K)
+    """
+    N, K, Wx = simcc_x.shape
+    simcc_x = simcc_x.reshape(N * K, -1)
+    simcc_y = simcc_y.reshape(N * K, -1)
+
+    # get maximum value locations
+    x_locs = np.argmax(simcc_x, axis=1)
+    y_locs = np.argmax(simcc_y, axis=1)
+    locs = np.stack((x_locs, y_locs), axis=-1).astype(np.float32)
+    max_val_x = np.amax(simcc_x, axis=1)
+    max_val_y = np.amax(simcc_y, axis=1)
+
+    # get maximum value across x and y axis
+    mask = max_val_x > max_val_y
+    max_val_x[mask] = max_val_y[mask]
+    vals = max_val_x
+    locs[vals <= 0.] = -1
+
+    # reshape
+    locs = locs.reshape(N, K, 2)
+    vals = vals.reshape(N, K)
+
+    return locs, vals
+
+
+def decode(simcc_x: np.ndarray, simcc_y: np.ndarray,
+           simcc_split_ratio) -> Tuple[np.ndarray, np.ndarray]:
+    """Modulate simcc distribution with Gaussian.
+
+    Args:
+        simcc_x (np.ndarray[K, Wx]): model predicted simcc in x.
+        simcc_y (np.ndarray[K, Wy]): model predicted simcc in y.
+        simcc_split_ratio (int): The split ratio of simcc.
+
+    Returns:
+        tuple: A tuple containing center and scale.
+        - np.ndarray[float32]: keypoints in shape (K, 2) or (n, K, 2)
+        - np.ndarray[float32]: scores in shape (K,) or (n, K)
+    """
+    keypoints, scores = get_simcc_maximum(simcc_x, simcc_y)
+    keypoints /= simcc_split_ratio
+
+    return keypoints, scores
+
+
+def inference_pose(session, out_bbox, oriImg):
+    h, w = session.get_inputs()[0].shape[2:]
+    model_input_size = (w, h)
+    resized_img, center, scale = preprocess(oriImg, out_bbox, model_input_size)
+    outputs = inference(session, resized_img)
+    keypoints, scores = postprocess(outputs, model_input_size, center, scale)
+
+    return keypoints, scores

scepter/modules/annotator/dwpose/util.py

@@ -0,0 +1,299 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import math
+import numpy as np
+import matplotlib
+import cv2
+
+
+eps = 0.01
+
+
+def smart_resize(x, s):
+    Ht, Wt = s
+    if x.ndim == 2:
+        Ho, Wo = x.shape
+        Co = 1
+    else:
+        Ho, Wo, Co = x.shape
+    if Co == 3 or Co == 1:
+        k = float(Ht + Wt) / float(Ho + Wo)
+        return cv2.resize(x, (int(Wt), int(Ht)), interpolation=cv2.INTER_AREA if k < 1 else cv2.INTER_LANCZOS4)
+    else:
+        return np.stack([smart_resize(x[:, :, i], s) for i in range(Co)], axis=2)
+
+
+def smart_resize_k(x, fx, fy):
+    if x.ndim == 2:
+        Ho, Wo = x.shape
+        Co = 1
+    else:
+        Ho, Wo, Co = x.shape
+    Ht, Wt = Ho * fy, Wo * fx
+    if Co == 3 or Co == 1:
+        k = float(Ht + Wt) / float(Ho + Wo)
+        return cv2.resize(x, (int(Wt), int(Ht)), interpolation=cv2.INTER_AREA if k < 1 else cv2.INTER_LANCZOS4)
+    else:
+        return np.stack([smart_resize_k(x[:, :, i], fx, fy) for i in range(Co)], axis=2)
+
+
+def padRightDownCorner(img, stride, padValue):
+    h = img.shape[0]
+    w = img.shape[1]
+
+    pad = 4 * [None]
+    pad[0] = 0 # up
+    pad[1] = 0 # left
+    pad[2] = 0 if (h % stride == 0) else stride - (h % stride) # down
+    pad[3] = 0 if (w % stride == 0) else stride - (w % stride) # right
+
+    img_padded = img
+    pad_up = np.tile(img_padded[0:1, :, :]*0 + padValue, (pad[0], 1, 1))
+    img_padded = np.concatenate((pad_up, img_padded), axis=0)
+    pad_left = np.tile(img_padded[:, 0:1, :]*0 + padValue, (1, pad[1], 1))
+    img_padded = np.concatenate((pad_left, img_padded), axis=1)
+    pad_down = np.tile(img_padded[-2:-1, :, :]*0 + padValue, (pad[2], 1, 1))
+    img_padded = np.concatenate((img_padded, pad_down), axis=0)
+    pad_right = np.tile(img_padded[:, -2:-1, :]*0 + padValue, (1, pad[3], 1))
+    img_padded = np.concatenate((img_padded, pad_right), axis=1)
+
+    return img_padded, pad
+
+
+def transfer(model, model_weights):
+    transfered_model_weights = {}
+    for weights_name in model.state_dict().keys():
+        transfered_model_weights[weights_name] = model_weights['.'.join(weights_name.split('.')[1:])]
+    return transfered_model_weights
+
+
+def draw_bodypose(canvas, candidate, subset):
+    H, W, C = canvas.shape
+    candidate = np.array(candidate)
+    subset = np.array(subset)
+
+    stickwidth = 4
+
+    limbSeq = [[2, 3], [2, 6], [3, 4], [4, 5], [6, 7], [7, 8], [2, 9], [9, 10], \
+               [10, 11], [2, 12], [12, 13], [13, 14], [2, 1], [1, 15], [15, 17], \
+               [1, 16], [16, 18], [3, 17], [6, 18]]
+
+    colors = [[255, 0, 0], [255, 85, 0], [255, 170, 0], [255, 255, 0], [170, 255, 0], [85, 255, 0], [0, 255, 0], \
+              [0, 255, 85], [0, 255, 170], [0, 255, 255], [0, 170, 255], [0, 85, 255], [0, 0, 255], [85, 0, 255], \
+              [170, 0, 255], [255, 0, 255], [255, 0, 170], [255, 0, 85]]
+
+    for i in range(17):
+        for n in range(len(subset)):
+            index = subset[n][np.array(limbSeq[i]) - 1]
+            if -1 in index:
+                continue
+            Y = candidate[index.astype(int), 0] * float(W)
+            X = candidate[index.astype(int), 1] * float(H)
+            mX = np.mean(X)
+            mY = np.mean(Y)
+            length = ((X[0] - X[1]) ** 2 + (Y[0] - Y[1]) ** 2) ** 0.5
+            angle = math.degrees(math.atan2(X[0] - X[1], Y[0] - Y[1]))
+            polygon = cv2.ellipse2Poly((int(mY), int(mX)), (int(length / 2), stickwidth), int(angle), 0, 360, 1)
+            cv2.fillConvexPoly(canvas, polygon, colors[i])
+
+    canvas = (canvas * 0.6).astype(np.uint8)
+
+    for i in range(18):
+        for n in range(len(subset)):
+            index = int(subset[n][i])
+            if index == -1:
+                continue
+            x, y = candidate[index][0:2]
+            x = int(x * W)
+            y = int(y * H)
+            cv2.circle(canvas, (int(x), int(y)), 4, colors[i], thickness=-1)
+
+    return canvas
+
+
+def draw_handpose(canvas, all_hand_peaks):
+    H, W, C = canvas.shape
+
+    edges = [[0, 1], [1, 2], [2, 3], [3, 4], [0, 5], [5, 6], [6, 7], [7, 8], [0, 9], [9, 10], \
+             [10, 11], [11, 12], [0, 13], [13, 14], [14, 15], [15, 16], [0, 17], [17, 18], [18, 19], [19, 20]]
+
+    for peaks in all_hand_peaks:
+        peaks = np.array(peaks)
+
+        for ie, e in enumerate(edges):
+            x1, y1 = peaks[e[0]]
+            x2, y2 = peaks[e[1]]
+            x1 = int(x1 * W)
+            y1 = int(y1 * H)
+            x2 = int(x2 * W)
+            y2 = int(y2 * H)
+            if x1 > eps and y1 > eps and x2 > eps and y2 > eps:
+                cv2.line(canvas, (x1, y1), (x2, y2), matplotlib.colors.hsv_to_rgb([ie / float(len(edges)), 1.0, 1.0]) * 255, thickness=2)
+
+        for i, keyponit in enumerate(peaks):
+            x, y = keyponit
+            x = int(x * W)
+            y = int(y * H)
+            if x > eps and y > eps:
+                cv2.circle(canvas, (x, y), 4, (0, 0, 255), thickness=-1)
+    return canvas
+
+
+def draw_facepose(canvas, all_lmks):
+    H, W, C = canvas.shape
+    for lmks in all_lmks:
+        lmks = np.array(lmks)
+        for lmk in lmks:
+            x, y = lmk
+            x = int(x * W)
+            y = int(y * H)
+            if x > eps and y > eps:
+                cv2.circle(canvas, (x, y), 3, (255, 255, 255), thickness=-1)
+    return canvas
+
+
+# detect hand according to body pose keypoints
+# please refer to https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/src/openpose/hand/handDetector.cpp
+def handDetect(candidate, subset, oriImg):
+    # right hand: wrist 4, elbow 3, shoulder 2
+    # left hand: wrist 7, elbow 6, shoulder 5
+    ratioWristElbow = 0.33
+    detect_result = []
+    image_height, image_width = oriImg.shape[0:2]
+    for person in subset.astype(int):
+        # if any of three not detected
+        has_left = np.sum(person[[5, 6, 7]] == -1) == 0
+        has_right = np.sum(person[[2, 3, 4]] == -1) == 0
+        if not (has_left or has_right):
+            continue
+        hands = []
+        #left hand
+        if has_left:
+            left_shoulder_index, left_elbow_index, left_wrist_index = person[[5, 6, 7]]
+            x1, y1 = candidate[left_shoulder_index][:2]
+            x2, y2 = candidate[left_elbow_index][:2]
+            x3, y3 = candidate[left_wrist_index][:2]
+            hands.append([x1, y1, x2, y2, x3, y3, True])
+        # right hand
+        if has_right:
+            right_shoulder_index, right_elbow_index, right_wrist_index = person[[2, 3, 4]]
+            x1, y1 = candidate[right_shoulder_index][:2]
+            x2, y2 = candidate[right_elbow_index][:2]
+            x3, y3 = candidate[right_wrist_index][:2]
+            hands.append([x1, y1, x2, y2, x3, y3, False])
+
+        for x1, y1, x2, y2, x3, y3, is_left in hands:
+            # pos_hand = pos_wrist + ratio * (pos_wrist - pos_elbox) = (1 + ratio) * pos_wrist - ratio * pos_elbox
+            # handRectangle.x = posePtr[wrist*3] + ratioWristElbow * (posePtr[wrist*3] - posePtr[elbow*3]);
+            # handRectangle.y = posePtr[wrist*3+1] + ratioWristElbow * (posePtr[wrist*3+1] - posePtr[elbow*3+1]);
+            # const auto distanceWristElbow = getDistance(poseKeypoints, person, wrist, elbow);
+            # const auto distanceElbowShoulder = getDistance(poseKeypoints, person, elbow, shoulder);
+            # handRectangle.width = 1.5f * fastMax(distanceWristElbow, 0.9f * distanceElbowShoulder);
+            x = x3 + ratioWristElbow * (x3 - x2)
+            y = y3 + ratioWristElbow * (y3 - y2)
+            distanceWristElbow = math.sqrt((x3 - x2) ** 2 + (y3 - y2) ** 2)
+            distanceElbowShoulder = math.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2)
+            width = 1.5 * max(distanceWristElbow, 0.9 * distanceElbowShoulder)
+            # x-y refers to the center --> offset to topLeft point
+            # handRectangle.x -= handRectangle.width / 2.f;
+            # handRectangle.y -= handRectangle.height / 2.f;
+            x -= width / 2
+            y -= width / 2  # width = height
+            # overflow the image
+            if x < 0: x = 0
+            if y < 0: y = 0
+            width1 = width
+            width2 = width
+            if x + width > image_width: width1 = image_width - x
+            if y + width > image_height: width2 = image_height - y
+            width = min(width1, width2)
+            # the max hand box value is 20 pixels
+            if width >= 20:
+                detect_result.append([int(x), int(y), int(width), is_left])
+
+    '''
+    return value: [[x, y, w, True if left hand else False]].
+    width=height since the network require squared input.
+    x, y is the coordinate of top left 
+    '''
+    return detect_result
+
+
+# Written by Lvmin
+def faceDetect(candidate, subset, oriImg):
+    # left right eye ear 14 15 16 17
+    detect_result = []
+    image_height, image_width = oriImg.shape[0:2]
+    for person in subset.astype(int):
+        has_head = person[0] > -1
+        if not has_head:
+            continue
+
+        has_left_eye = person[14] > -1
+        has_right_eye = person[15] > -1
+        has_left_ear = person[16] > -1
+        has_right_ear = person[17] > -1
+
+        if not (has_left_eye or has_right_eye or has_left_ear or has_right_ear):
+            continue
+
+        head, left_eye, right_eye, left_ear, right_ear = person[[0, 14, 15, 16, 17]]
+
+        width = 0.0
+        x0, y0 = candidate[head][:2]
+
+        if has_left_eye:
+            x1, y1 = candidate[left_eye][:2]
+            d = max(abs(x0 - x1), abs(y0 - y1))
+            width = max(width, d * 3.0)
+
+        if has_right_eye:
+            x1, y1 = candidate[right_eye][:2]
+            d = max(abs(x0 - x1), abs(y0 - y1))
+            width = max(width, d * 3.0)
+
+        if has_left_ear:
+            x1, y1 = candidate[left_ear][:2]
+            d = max(abs(x0 - x1), abs(y0 - y1))
+            width = max(width, d * 1.5)
+
+        if has_right_ear:
+            x1, y1 = candidate[right_ear][:2]
+            d = max(abs(x0 - x1), abs(y0 - y1))
+            width = max(width, d * 1.5)
+
+        x, y = x0, y0
+
+        x -= width
+        y -= width
+
+        if x < 0:
+            x = 0
+
+        if y < 0:
+            y = 0
+
+        width1 = width * 2
+        width2 = width * 2
+
+        if x + width > image_width:
+            width1 = image_width - x
+
+        if y + width > image_height:
+            width2 = image_height - y
+
+        width = min(width1, width2)
+
+        if width >= 20:
+            detect_result.append([int(x), int(y), int(width)])
+
+    return detect_result
+
+
+# get max index of 2d array
+def npmax(array):
+    arrayindex = array.argmax(1)
+    arrayvalue = array.max(1)
+    i = arrayvalue.argmax()
+    j = arrayindex[i]
+    return i, j

scepter/modules/annotator/dwpose/wholebody.py

@@ -0,0 +1,80 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import cv2
+import numpy as np
+import onnxruntime as ort
+from .onnxdet import inference_detector
+from .onnxpose import inference_pose
+
+def HWC3(x):
+    assert x.dtype == np.uint8
+    if x.ndim == 2:
+        x = x[:, :, None]
+    assert x.ndim == 3
+    H, W, C = x.shape
+    assert C == 1 or C == 3 or C == 4
+    if C == 3:
+        return x
+    if C == 1:
+        return np.concatenate([x, x, x], axis=2)
+    if C == 4:
+        color = x[:, :, 0:3].astype(np.float32)
+        alpha = x[:, :, 3:4].astype(np.float32) / 255.0
+        y = color * alpha + 255.0 * (1.0 - alpha)
+        y = y.clip(0, 255).astype(np.uint8)
+        return y
+
+
+def resize_image(input_image, resolution):
+    H, W, C = input_image.shape
+    H = float(H)
+    W = float(W)
+    k = float(resolution) / min(H, W)
+    H *= k
+    W *= k
+    H = int(np.round(H / 64.0)) * 64
+    W = int(np.round(W / 64.0)) * 64
+    img = cv2.resize(input_image, (W, H), interpolation=cv2.INTER_LANCZOS4 if k > 1 else cv2.INTER_AREA)
+    return img
+
+class Wholebody:
+    def __init__(self, onnx_det, onnx_pose, device = 'cuda:0'):
+
+        providers = ['CPUExecutionProvider'
+                 ] if device == 'cpu' else ['CUDAExecutionProvider']
+        # onnx_det = 'annotator/ckpts/yolox_l.onnx'
+        # onnx_pose = 'annotator/ckpts/dw-ll_ucoco_384.onnx'
+
+        self.session_det = ort.InferenceSession(path_or_bytes=onnx_det, providers=providers)
+        self.session_pose = ort.InferenceSession(path_or_bytes=onnx_pose, providers=providers)
+    
+    def __call__(self, ori_img):
+        det_result = inference_detector(self.session_det, ori_img)
+        keypoints, scores = inference_pose(self.session_pose, det_result, ori_img)
+
+        keypoints_info = np.concatenate(
+            (keypoints, scores[..., None]), axis=-1)
+        # compute neck joint
+        neck = np.mean(keypoints_info[:, [5, 6]], axis=1)
+        # neck score when visualizing pred
+        neck[:, 2:4] = np.logical_and(
+            keypoints_info[:, 5, 2:4] > 0.3,
+            keypoints_info[:, 6, 2:4] > 0.3).astype(int)
+        new_keypoints_info = np.insert(
+            keypoints_info, 17, neck, axis=1)
+        mmpose_idx = [
+            17, 6, 8, 10, 7, 9, 12, 14, 16, 13, 15, 2, 1, 4, 3
+        ]
+        openpose_idx = [
+            1, 2, 3, 4, 6, 7, 8, 9, 10, 12, 13, 14, 15, 16, 17
+        ]
+        new_keypoints_info[:, openpose_idx] = \
+            new_keypoints_info[:, mmpose_idx]
+        keypoints_info = new_keypoints_info
+
+        keypoints, scores = keypoints_info[
+            ..., :2], keypoints_info[..., 2]
+        
+        return keypoints, scores, det_result
+
+

scepter/modules/annotator/dwpose_op.py

@@ -0,0 +1,203 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+# Openpose
+# Original from CMU https://github.com/CMU-Perceptual-Computing-Lab/openpose
+# 2nd Edited by https://github.com/Hzzone/pytorch-openpose
+# 3rd Edited by ControlNet
+# 4th Edited by ControlNet (added face and correct hands)
+
+# ``` requirements for cuda 12.1:
+#     onnxruntime==1.19
+#     onnxruntime-gpu==1.19
+# ```
+
+import os
+
+import numpy as np
+import torch
+from PIL import Image
+
+import cv2
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.dwpose import util
+from scepter.modules.annotator.dwpose.wholebody import (HWC3, Wholebody,
+                                                        resize_image)
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.distribute import we
+from scepter.modules.utils.file_system import FS
+
+os.environ['KMP_DUPLICATE_LIB_OK'] = 'TRUE'
+
+
+def draw_pose(pose, H, W, use_hand=False, use_body=False, use_face=False):
+    bodies = pose['bodies']
+    faces = pose['faces']
+    hands = pose['hands']
+    candidate = bodies['candidate']
+    subset = bodies['subset']
+    canvas = np.zeros(shape=(H, W, 3), dtype=np.uint8)
+
+    if use_body:
+        canvas = util.draw_bodypose(canvas, candidate, subset)
+    if use_hand:
+        canvas = util.draw_handpose(canvas, hands)
+    if use_face:
+        canvas = util.draw_facepose(canvas, faces)
+
+    return canvas
+
+
+@ANNOTATORS.register_class()
+class DWposeAnnotator(BaseAnnotator):
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        with FS.get_from(cfg['DETECTION_MODEL'],
+                         wait_finish=True) as onnx_det, FS.get_from(
+                             cfg['POSE_MODEL'], wait_finish=True) as onnx_pose:
+            self.pose_estimation = Wholebody(onnx_det,
+                                             onnx_pose,
+                                             device=f'cuda:{we.device_id}')
+        self.resize_size = cfg.get('RESIZE_SIZE', 1024)
+        self.use_body = cfg.get('USE_BODY', True)
+        self.use_face = cfg.get('USE_FACE', True)
+        self.use_hand = cfg.get('USE_HAND', True)
+
+    @torch.no_grad()
+    @torch.inference_mode
+    def forward(self, image):
+        if isinstance(image, Image.Image):
+            image = np.array(image)
+        elif isinstance(image, torch.Tensor):
+            image = image.detach().cpu().numpy()
+        elif isinstance(image, np.ndarray):
+            image = image.copy()
+        else:
+            raise f'Unsurpport datatype{type(image)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+
+        input_image = HWC3(image[..., ::-1])
+        return self.process(resize_image(input_image, self.resize_size),
+                            image.shape[:2])
+
+    def process(self, ori_img, ori_shape):
+        ori_h, ori_w = ori_shape
+        ori_img = ori_img.copy()
+        H, W, C = ori_img.shape
+        with torch.no_grad():
+            candidate, subset, det_result = self.pose_estimation(ori_img)
+            nums, keys, locs = candidate.shape
+            candidate[..., 0] /= float(W)
+            candidate[..., 1] /= float(H)
+            body = candidate[:, :18].copy()
+            body = body.reshape(nums * 18, locs)
+            score = subset[:, :18]
+            for i in range(len(score)):
+                for j in range(len(score[i])):
+                    if score[i][j] > 0.3:
+                        score[i][j] = int(18 * i + j)
+                    else:
+                        score[i][j] = -1
+
+            un_visible = subset < 0.3
+            candidate[un_visible] = -1
+
+            foot = candidate[:, 18:24]
+
+            faces = candidate[:, 24:92]
+
+            hands = candidate[:, 92:113]
+            hands = np.vstack([hands, candidate[:, 113:]])
+
+            bodies = dict(candidate=body, subset=score)
+            pose = dict(bodies=bodies, hands=hands, faces=faces)
+
+            ret_data = {}
+            if self.use_body:
+                detected_map_body = draw_pose(pose, H, W, use_body=True)
+                detected_map_body = cv2.resize(
+                    detected_map_body[..., ::-1], (ori_w, ori_h),
+                    interpolation=cv2.INTER_LANCZOS4
+                    if ori_h * ori_w > H * W else cv2.INTER_AREA)
+                ret_data['detected_map_body'] = detected_map_body
+
+            if self.use_face:
+                detected_map_face = draw_pose(pose, H, W, use_face=True)
+                detected_map_face = cv2.resize(
+                    detected_map_face[..., ::-1], (ori_w, ori_h),
+                    interpolation=cv2.INTER_LANCZOS4
+                    if ori_h * ori_w > H * W else cv2.INTER_AREA)
+                ret_data['detected_map_face'] = detected_map_face
+
+            if self.use_body and self.use_face:
+                detected_map_bodyface = draw_pose(pose,
+                                                  H,
+                                                  W,
+                                                  use_body=True,
+                                                  use_face=True)
+                detected_map_bodyface = cv2.resize(
+                    detected_map_bodyface[..., ::-1], (ori_w, ori_h),
+                    interpolation=cv2.INTER_LANCZOS4
+                    if ori_h * ori_w > H * W else cv2.INTER_AREA)
+                ret_data['detected_map_bodyface'] = detected_map_bodyface
+
+            if self.use_hand and self.use_body and self.use_face:
+                detected_map_handbodyface = draw_pose(pose,
+                                                      H,
+                                                      W,
+                                                      use_hand=True,
+                                                      use_body=True,
+                                                      use_face=True)
+                detected_map_handbodyface = cv2.resize(
+                    detected_map_handbodyface[..., ::-1], (ori_w, ori_h),
+                    interpolation=cv2.INTER_LANCZOS4
+                    if ori_h * ori_w > H * W else cv2.INTER_AREA)
+                ret_data[
+                    'detected_map_handbodyface'] = detected_map_handbodyface
+
+            # convert_size
+            if det_result.shape[0] > 0:
+                w_ratio, h_ratio = ori_w / W, ori_h / H
+                det_result[..., ::2] *= h_ratio
+                det_result[..., 1::2] *= w_ratio
+                det_result = det_result.astype(np.int32)
+            # for det_tup in det_result:
+            #     cv2.rectangle(detected_map, det_tup[2:].tolist(), det_tup[:2].tolist(), color=(255, 0, 0), thickness=3)
+            return ret_data, det_result
+
+
+@ANNOTATORS.register_class()
+class DWposeBodyAnnotator(DWposeAnnotator):
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        self.use_body, self.use_face, self.use_hand = True, False, False
+
+    @torch.no_grad()
+    @torch.inference_mode
+    def forward(self, image):
+        ret_data, det_result = super().forward(image)
+        return ret_data['detected_map_body']
+
+
+@ANNOTATORS.register_class()
+class DWposeFaceAnnotator(DWposeAnnotator):
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        self.use_body, self.use_face, self.use_hand = False, True, False
+
+    @torch.no_grad()
+    @torch.inference_mode
+    def forward(self, image):
+        ret_data, det_result = super().forward(image)
+        return ret_data['detected_map_face']
+
+
+@ANNOTATORS.register_class()
+class DWposeBodyFaceAnnotator(DWposeAnnotator):
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        self.use_body, self.use_face, self.use_hand = True, True, False
+
+    @torch.no_grad()
+    @torch.inference_mode
+    def forward(self, image):
+        ret_data, det_result = super().forward(image)
+        return ret_data['detected_map_bodyface']

scepter/modules/annotator/face.py

@@ -0,0 +1,63 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import os
+from abc import ABCMeta
+
+import numpy as np
+import torch
+from PIL import Image
+
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.distribute import we
+from scepter.modules.utils.file_system import FS
+
+
+@ANNOTATORS.register_class()
+class FaceAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        from insightface.app import FaceAnalysis
+        local_path = FS.map_to_local(cfg.PRETRAINED_MODEL)[0]
+        local_model_path = os.path.join(local_path, 'models', cfg.MODEL_NAME)
+        FS.get_dir_to_local_dir(cfg.PRETRAINED_MODEL, local_model_path)
+        self.model = FaceAnalysis(name=cfg.MODEL_NAME, root=local_path, providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])
+        self.model.prepare(ctx_id=we.device_id, det_size=(640, 640))
+
+    def forward(self, image=None):
+
+        if isinstance(image, Image.Image):
+            image = np.array(image)
+        elif isinstance(image, torch.Tensor):
+            image = image.detach().cpu().numpy()
+        elif isinstance(image, np.ndarray):
+            image = image.copy()
+        else:
+            raise f'Unsurpport datatype{type(image)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+
+        # [dict_keys(['bbox', 'kps', 'det_score', 'landmark_3d_68', 'pose', 'landmark_2d_106', 'gender', 'age', 'embedding'])]
+        faces = self.model.get(image)
+        return faces
+
+
+@ANNOTATORS.register_class()
+class FaceMaskAnnotator(FaceAnnotator):
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        self.multi_face = cfg.get('MULTI_FACE', True)
+
+    def forward(self, image=None):
+        faces = super().forward(image)
+        if len(faces) > 0:
+            if not self.multi_face:
+                faces = faces[:1]
+            mask = np.zeros_like(image[:, :, 0])
+            for face in faces:
+                x_min, y_min, x_max, y_max = face['bbox'].tolist()
+                mask[int(y_min): int(y_max) + 1, int(x_min): int(x_max) + 1] = 255
+            return mask
+        else:
+            return np.zeros_like(image[:, :, 0])

scepter/modules/annotator/frame_reference.py

@@ -0,0 +1,58 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import random
+import numpy as np
+
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import Config
+
+
+@ANNOTATORS.register_class()
+class FrameReferenceAnnotator(BaseAnnotator):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        # first / last / firstlast / random
+        self.ref_cfg = cfg.get('REF_CFG', [{"mode": "first", "proba": 0.1},
+                                           {"mode": "last", "proba": 0.1},
+                                           {"mode": "firstlast", "proba": 0.1},
+                                           {"mode": "random", "proba": 0.1}])
+        self.ref_num = cfg.get('REF_NUM', 1)
+        self.ref_cfg = Config.get_dict(self.ref_cfg) if isinstance(
+            self.ref_cfg, Config) else self.ref_cfg
+        self.ref_color = cfg.get('REF_COLOR', 127.5)
+
+    def forward(self, frames, ref_cfg=None, ref_num=None):
+        ref_cfg = ref_cfg if ref_cfg is not None else self.ref_cfg
+        ref_cfg = [ref_cfg] if not isinstance(ref_cfg, list) else ref_cfg
+        probas = [item['proba'] if 'proba' in item else 1.0 / len(ref_cfg) for item in ref_cfg]
+        sel_ref_cfg = random.choices(ref_cfg, weights=probas, k=1)[0]
+        mode = sel_ref_cfg['mode'] if 'mode' in sel_ref_cfg else 'original'
+        ref_num = int(ref_num) if ref_num is not None else self.ref_num
+
+        frame_num = len(frames)
+        frame_num_range = list(range(frame_num))
+        if mode == "first":
+            sel_idx = frame_num_range[:ref_num]
+        elif mode == "last":
+            sel_idx = frame_num_range[-ref_num:]
+        elif mode == "firstlast":
+            sel_idx = frame_num_range[:ref_num] + frame_num_range[-ref_num:]
+        elif mode == "random":
+            sel_idx = random.sample(frame_num_range, ref_num)
+        else:
+            raise NotImplementedError
+
+        out_frames, out_masks = [], []
+        for i in range(frame_num):
+            if i in sel_idx:
+                out_frame = frames[i]
+                out_mask = np.zeros_like(frames[i][:, :, 0])
+            else:
+                out_frame = np.ones_like(frames[i]) * self.ref_color
+                out_mask = np.ones_like(frames[i][:, :, 0]) * 255
+            out_frames.append(out_frame)
+            out_masks.append(out_mask)
+        return out_frames, out_masks

scepter/modules/annotator/gray.py

@@ -0,0 +1,38 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+from abc import ABCMeta
+
+import cv2
+import numpy as np
+import torch
+from PIL import Image
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import dict_to_yaml
+
+
+@ANNOTATORS.register_class()
+class GrayAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+
+    def forward(self, image):
+        if isinstance(image, Image.Image):
+            image = np.array(image)
+        elif isinstance(image, torch.Tensor):
+            image = image.detach().cpu().numpy()
+        elif isinstance(image, np.ndarray):
+            image = image.copy()
+        else:
+            raise f'Unsurpport datatype{type(image)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+        gray_map = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
+        return gray_map[..., None].repeat(3, axis=2)
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('ANNOTATORS',
+                            __class__.__name__,
+                            GrayAnnotator.para_dict,
+                            set_name=True)

scepter/modules/annotator/hed.py

@@ -114,7 +114,7 @@ class HedAnnotator(BaseAnnotator, metaclass=ABCMeta):
         pretrained_model = cfg.get('PRETRAINED_MODEL', None)
         if pretrained_model:
             with FS.get_from(pretrained_model, wait_finish=True) as local_path:
-                self.netNetwork.load_state_dict(torch.load(local_path))
+                self.netNetwork.load_state_dict(torch.load(local_path, weights_only=True))
 
     @torch.no_grad()
     @torch.inference_mode()

scepter/modules/annotator/informative_drawing.py

@@ -0,0 +1,175 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+from abc import ABCMeta
+
+import numpy as np
+import torch
+import torch.nn as nn
+from einops import rearrange
+from torchvision.transforms import InterpolationMode
+
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import dict_to_yaml
+from scepter.modules.utils.distribute import we
+from scepter.modules.utils.file_system import FS
+
+norm_layer = nn.InstanceNorm2d
+
+
+class ResidualBlock(nn.Module):
+    def __init__(self, in_features):
+        super(ResidualBlock, self).__init__()
+
+        conv_block = [
+            nn.ReflectionPad2d(1),
+            nn.Conv2d(in_features, in_features, 3),
+            norm_layer(in_features),
+            nn.ReLU(inplace=True),
+            nn.ReflectionPad2d(1),
+            nn.Conv2d(in_features, in_features, 3),
+            norm_layer(in_features)
+        ]
+
+        self.conv_block = nn.Sequential(*conv_block)
+
+    def forward(self, x):
+        return x + self.conv_block(x)
+
+
+class ContourInference(nn.Module):
+    def __init__(self, input_nc, output_nc, n_residual_blocks=9, sigmoid=True):
+        super(ContourInference, self).__init__()
+
+        # Initial convolution block
+        model0 = [
+            nn.ReflectionPad2d(3),
+            nn.Conv2d(input_nc, 64, 7),
+            norm_layer(64),
+            nn.ReLU(inplace=True)
+        ]
+        self.model0 = nn.Sequential(*model0)
+
+        # Downsampling
+        model1 = []
+        in_features = 64
+        out_features = in_features * 2
+        for _ in range(2):
+            model1 += [
+                nn.Conv2d(in_features, out_features, 3, stride=2, padding=1),
+                norm_layer(out_features),
+                nn.ReLU(inplace=True)
+            ]
+            in_features = out_features
+            out_features = in_features * 2
+        self.model1 = nn.Sequential(*model1)
+
+        model2 = []
+        # Residual blocks
+        for _ in range(n_residual_blocks):
+            model2 += [ResidualBlock(in_features)]
+        self.model2 = nn.Sequential(*model2)
+
+        # Upsampling
+        model3 = []
+        out_features = in_features // 2
+        for _ in range(2):
+            model3 += [
+                nn.ConvTranspose2d(in_features,
+                                   out_features,
+                                   3,
+                                   stride=2,
+                                   padding=1,
+                                   output_padding=1),
+                norm_layer(out_features),
+                nn.ReLU(inplace=True)
+            ]
+            in_features = out_features
+            out_features = in_features // 2
+        self.model3 = nn.Sequential(*model3)
+
+        # Output layer
+        model4 = [nn.ReflectionPad2d(3), nn.Conv2d(64, output_nc, 7)]
+        if sigmoid:
+            model4 += [nn.Sigmoid()]
+
+        self.model4 = nn.Sequential(*model4)
+
+    def forward(self, x, cond=None):
+        out = self.model0(x)
+        out = self.model1(out)
+        out = self.model2(out)
+        out = self.model3(out)
+        out = self.model4(out)
+
+        return out
+
+
+@ANNOTATORS.register_class()
+class InfoDrawContourAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        input_nc = cfg.get('INPUT_NC', 3)
+        output_nc = cfg.get('OUTPUT_NC', 1)
+        n_residual_blocks = cfg.get('N_RESIDUAL_BLOCKS', 3)
+        sigmoid = cfg.get('SIGMOID', True)
+        pretrained_model = cfg.get('PRETRAINED_MODEL', None)
+
+        self.model = ContourInference(input_nc, output_nc, n_residual_blocks,
+                                      sigmoid)
+        with FS.get_from(pretrained_model, wait_finish=True) as local_path:
+            self.model.load_state_dict(torch.load(local_path, weights_only=True))
+        self.model = self.model.eval().requires_grad_(False).to(we.device_id)
+
+    @torch.no_grad()
+    @torch.inference_mode()
+    @torch.autocast('cuda', enabled=False)
+    def forward(self, image):
+        is_batch = False if len(image.shape) == 3 else True
+        if isinstance(image, torch.Tensor):
+            if len(image.shape) == 3:
+                image = rearrange(image, 'h w c -> 1 c h w')
+                B, C, H, W = image.shape
+            elif len(image.shape) == 4:
+                B, C, H, W = image.shape
+            else:
+                raise "Unsurpport input image's shape"
+        elif isinstance(image, np.ndarray):
+            image = torch.from_numpy(image.copy()).float()
+            if len(image.shape) == 3:
+                image = rearrange(image, 'h w c -> 1 c h w')
+                B, C, H, W = image.shape
+            elif len(image.shape) == 4:
+                B, C, H, W = image.shape
+            else:
+                raise "Unsurpport input image's shape"
+        else:
+            raise "Unsurpport input image's type"
+
+        image = image.float().div(255).to(we.device_id)
+        contour_map = self.model(image)
+        contour_map = (contour_map.squeeze(dim=1) * 255.0).clip(
+            0, 255).cpu().numpy().astype(np.uint8)
+        contour_map = contour_map[..., None].repeat(3, -1)
+        if not is_batch:
+            contour_map = contour_map.squeeze()
+        return contour_map
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('ANNOTATORS',
+                            __class__.__name__,
+                            InfoDrawContourAnnotator.para_dict,
+                            set_name=True)
+
+
+@ANNOTATORS.register_class()
+class InfoDrawAnimeAnnotator(InfoDrawContourAnnotator):
+    pass
+
+
+@ANNOTATORS.register_class()
+class InfoDrawOpenSketchAnnotator(InfoDrawContourAnnotator):
+    pass

scepter/modules/annotator/inpainting.py

@@ -0,0 +1,275 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import random
+from abc import ABCMeta
+from enum import Enum
+
+import cv2
+import numpy as np
+import torch
+from PIL import Image
+
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import Config, dict_to_yaml
+
+
+def invert_image(im):
+    im_arr = np.array(im)
+    mask_1 = im_arr == 0
+    mask_2 = im_arr == 255
+    im_arr[mask_1] = 255
+    im_arr[mask_2] = 0
+    new_im = im_arr
+    return new_im
+
+
+class DrawMethod(Enum):
+    LINE = 'line'
+    CIRCLE = 'circle'
+    SQUARE = 'square'
+
+
+def make_random_irregular_mask(shape,
+                               max_angle=4,
+                               max_len=60,
+                               max_width=20,
+                               min_times=0,
+                               max_times=10,
+                               draw_method=DrawMethod.LINE):
+    draw_method = DrawMethod(draw_method)
+
+    height, width = shape
+    mask = np.zeros((height, width), np.float32)
+    times = np.random.randint(min_times, max_times + 1)
+    for i in range(times):
+        start_x = np.random.randint(width)
+        start_y = np.random.randint(height)
+        for j in range(1 + np.random.randint(5)):
+            angle = 0.01 + np.random.randint(max_angle)
+            if i % 2 == 0:
+                angle = 2 * 3.1415926 - angle
+            length = 10 + np.random.randint(max_len)
+            brush_w = 5 + np.random.randint(max_width)
+            end_x = np.clip(
+                (start_x + length * np.sin(angle)).astype(np.int32), 0, width)
+            end_y = np.clip(
+                (start_y + length * np.cos(angle)).astype(np.int32), 0, height)
+            if draw_method == DrawMethod.LINE:
+                cv2.line(mask, (start_x, start_y), (end_x, end_y), 1.0,
+                         brush_w)
+            elif draw_method == DrawMethod.CIRCLE:
+                cv2.circle(mask, (start_x, start_y),
+                           radius=brush_w,
+                           color=1.,
+                           thickness=-1)
+            elif draw_method == DrawMethod.SQUARE:
+                radius = brush_w // 2
+                mask[start_y - radius:start_y + radius,
+                     start_x - radius:start_x + radius] = 1
+            start_x, start_y = end_x, end_y
+    return mask[None, ...]
+
+
+class RandomIrregularMaskGenerator:
+    def __init__(self,
+                 max_angle=4,
+                 max_len=60,
+                 max_width=20,
+                 min_times=0,
+                 max_times=10,
+                 ramp_kwargs=None,
+                 draw_method=DrawMethod.LINE):
+        self.max_angle = max_angle
+        self.max_len = max_len
+        self.max_width = max_width
+        self.min_times = min_times
+        self.max_times = max_times
+        self.draw_method = draw_method
+        self.ramp = None
+        # self.ramp = LinearRamp(**ramp_kwargs) if ramp_kwargs is not None else None
+
+    def __call__(self, img, iter_i=None, raw_image=None):
+        coef = self.ramp(iter_i) if (self.ramp is not None) and (
+            iter_i is not None) else 1
+        cur_max_len = int(max(1, self.max_len * coef))
+        cur_max_width = int(max(1, self.max_width * coef))
+        cur_max_times = int(self.min_times + 1 +
+                            (self.max_times - self.min_times) * coef)
+        return make_random_irregular_mask(img.shape[1:],
+                                          max_angle=self.max_angle,
+                                          max_len=cur_max_len,
+                                          max_width=cur_max_width,
+                                          min_times=self.min_times,
+                                          max_times=cur_max_times,
+                                          draw_method=self.draw_method)
+
+
+def make_random_rectangle_mask(shape,
+                               margin=10,
+                               bbox_min_size=30,
+                               bbox_max_size=100,
+                               min_times=0,
+                               max_times=3):
+    height, width = shape
+    mask = np.zeros((height, width), np.float32)
+    bbox_max_size = min(bbox_max_size, height - margin * 2, width - margin * 2)
+    times = np.random.randint(min_times, max_times + 1)
+    for i in range(times):
+        box_width = np.random.randint(bbox_min_size, bbox_max_size)
+        box_height = np.random.randint(bbox_min_size, bbox_max_size)
+        start_x = np.random.randint(margin, width - margin - box_width + 1)
+        start_y = np.random.randint(margin, height - margin - box_height + 1)
+        mask[start_y:start_y + box_height, start_x:start_x + box_width] = 1
+    return mask[None, ...]
+
+
+class RandomRectangleMaskGenerator:
+    def __init__(self,
+                 margin=10,
+                 bbox_min_size=30,
+                 bbox_max_size=100,
+                 min_times=0,
+                 max_times=3,
+                 ramp_kwargs=None):
+        self.margin = margin
+        self.bbox_min_size = bbox_min_size
+        self.bbox_max_size = bbox_max_size
+        self.min_times = min_times
+        self.max_times = max_times
+        self.ramp = None
+        # self.ramp = LinearRamp(**ramp_kwargs) if ramp_kwargs is not None else None
+
+    def __call__(self, img, iter_i=None, raw_image=None):
+        coef = self.ramp(iter_i) if (self.ramp is not None) and (
+            iter_i is not None) else 1
+        cur_bbox_max_size = int(self.bbox_min_size + 1 +
+                                (self.bbox_max_size - self.bbox_min_size) *
+                                coef)
+        cur_max_times = int(self.min_times +
+                            (self.max_times - self.min_times) * coef)
+        return make_random_rectangle_mask(img.shape[1:],
+                                          margin=self.margin,
+                                          bbox_min_size=self.bbox_min_size,
+                                          bbox_max_size=cur_bbox_max_size,
+                                          min_times=self.min_times,
+                                          max_times=cur_max_times)
+
+
+class MixedMaskGenerator:
+    def __init__(self,
+                 irregular_proba=1 / 3,
+                 irregular_kwargs=None,
+                 box_proba=1 / 3,
+                 box_kwargs=None,
+                 invert_proba=0):
+        self.probas = []
+        self.gens = []
+
+        if irregular_proba > 0:
+            self.probas.append(irregular_proba)
+            if irregular_kwargs is None:
+                irregular_kwargs = {}
+            else:
+                irregular_kwargs = dict(irregular_kwargs)
+            irregular_kwargs['draw_method'] = DrawMethod.LINE
+            self.gens.append(RandomIrregularMaskGenerator(**irregular_kwargs))
+
+        if box_proba > 0:
+            self.probas.append(box_proba)
+            if box_kwargs is None:
+                box_kwargs = {}
+            self.gens.append(RandomRectangleMaskGenerator(**box_kwargs))
+
+        self.probas = np.array(self.probas, dtype='float32')
+        self.probas /= self.probas.sum()
+        self.invert_proba = invert_proba
+
+    def __call__(self, img, iter_i=None, raw_image=None):
+        kind = np.random.choice(len(self.probas), p=self.probas)
+        gen = self.gens[kind]
+        result = gen(img, iter_i=iter_i, raw_image=raw_image)
+        if self.invert_proba > 0 and random.random() < self.invert_proba:
+            result = 1 - result
+        return result
+
+
+@ANNOTATORS.register_class()
+class InpaintingAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        self.mask_cfg = cfg.get(
+            'MASK_CFG', {
+                'irregular_proba': 0.5,
+                'irregular_kwargs': {
+                    'min_times': 4,
+                    'max_times': 10,
+                    'max_width': 100,
+                    'max_angle': 4,
+                    'max_len': 200
+                },
+                'box_proba': 0.5,
+                'box_kwargs': {
+                    'margin': 0,
+                    'bbox_min_size': 30,
+                    'bbox_max_size': 150,
+                    'max_times': 5,
+                    'min_times': 1
+                }
+            })
+        self.mask_cfg = Config.get_dict(self.mask_cfg) if not isinstance(
+            self.mask_cfg, dict) else self.mask_cfg
+        self.mask_generator = MixedMaskGenerator(**self.mask_cfg)
+        self.return_mask = cfg.get('RETURN_MASK', False)
+        self.return_invert = cfg.get('RETURN_INVERT', True)
+        self.mask_color = cfg.get('MASK_COLOR', 0)
+
+    def forward(self,
+                image,
+                mask=None,
+                return_mask=None,
+                return_invert=None,
+                mask_color=None):
+        return_mask = return_mask if return_mask is not None else self.return_mask
+        return_invert = return_invert if return_invert is not None else self.return_invert
+        mask_color = mask_color if mask_color is not None else self.mask_color
+        if isinstance(image, Image.Image):
+            image = np.array(image)
+        elif isinstance(image, torch.Tensor):
+            image = image.detach().cpu().numpy()
+        elif isinstance(image, np.ndarray):
+            image = image.copy()
+        else:
+            raise f'Unsurpport datatype{type(image)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+
+        if mask is not None:
+            if mask_color:
+                image[np.array(mask) == 255] = mask_color
+            else:
+                image[np.array(mask) == 255] = 0
+        else:
+            img = np.transpose(image, (2, 0, 1))
+            mask = self.mask_generator(img)
+            mask = (np.transpose(mask,
+                                 (1, 2, 0)).squeeze(-1) * 255).astype(np.uint8)
+            if return_invert:
+                mask = invert_image(mask)
+            colored_mask = np.zeros_like(image)
+            if mask_color: colored_mask[:] = mask_color
+            image = np.where(mask[:, :, np.newaxis] == 255, colored_mask,
+                             image)
+
+        if return_mask:
+            ret_data = {'image': np.array(image), 'mask': np.array(mask)}
+        else:
+            ret_data = np.array(image)
+        return ret_data
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('ANNOTATORS',
+                            __class__.__name__,
+                            InpaintingAnnotator.para_dict,
+                            set_name=True)

scepter/modules/annotator/lama.py

@@ -0,0 +1,106 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+from abc import ABCMeta
+
+import numpy as np
+
+import cv2
+import torch
+from PIL import Image
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import dict_to_yaml
+from scepter.modules.utils.distribute import we
+from scepter.modules.utils.file_system import FS
+
+
+def dilate_mask(mask, dilate_factor=15):
+    mask = mask.astype(np.uint8)
+    mask = cv2.dilate(mask,
+                      np.ones((dilate_factor, dilate_factor), np.uint8),
+                      iterations=1)
+    return mask
+
+
+@ANNOTATORS.register_class()
+class LamaAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        from modelscope.pipelines.builder import PIPELINES
+        from modelscope.pipelines.cv import ImageInpaintingPipeline
+        from modelscope.pipelines import pipeline
+        from modelscope.utils.constant import Tasks
+        from modelscope.metainfo import Pipelines
+        from modelscope.models.cv.image_inpainting.refinement import refine_predict
+        from torch.utils.data._utils.collate import default_collate
+
+        @PIPELINES.register_module(Tasks.image_inpainting,
+                                   module_name=Pipelines.image_inpainting +
+                                   '-v2')
+        class ImageInpaintingPipelineV2(ImageInpaintingPipeline):
+            def perform_inference(self, data):
+                px_budget = 9000000
+                batch = default_collate([data])
+                if self.refine:
+                    assert 'unpad_to_size' in batch, 'Unpadded size is required for the refinement'
+                    assert 'cuda' in str(
+                        self.device), 'GPU is required for refinement'
+                    gpu_ids = str(self.device).split(':')[-1]
+                    cur_res = refine_predict(batch,
+                                             self.infer_model,
+                                             gpu_ids=gpu_ids,
+                                             modulo=self.pad_out_to_modulo,
+                                             n_iters=15,
+                                             lr=0.002,
+                                             min_side=512,
+                                             max_scales=3,
+                                             px_budget=px_budget)
+                    cur_res = cur_res[0].permute(1, 2,
+                                                 0).detach().cpu().numpy()
+                else:
+                    with torch.no_grad():
+                        batch = self.move_to_device(batch, self.device)
+                        batch['mask'] = (batch['mask'] > 0) * 1
+                        batch = self.infer_model(batch)
+                        cur_res = batch['inpainted'][0].permute(
+                            1, 2, 0).detach().cpu().numpy()
+                        unpad_to_size = batch.get('unpad_to_size', None)
+                        if unpad_to_size is not None:
+                            orig_height, orig_width = unpad_to_size
+                            cur_res = cur_res[:orig_height, :orig_width]
+
+                cur_res = np.clip(cur_res * 255, 0, 255).astype('uint8')
+                cur_res = cv2.cvtColor(cur_res, cv2.COLOR_RGB2BGR)
+                return cur_res
+
+        lama_model_dir = FS.get_dir_to_local_dir(cfg.PRETRAINED_MODEL)
+        self.lama_model = pipeline(Tasks.image_inpainting,
+                                   model=lama_model_dir,
+                                   pipeline_name=Pipelines.image_inpainting +
+                                   '-v2',
+                                   refine=True,
+                                   device='cuda:{}'.format(we.device_id))
+
+    def forward(self, image, mask):
+        mask = dilate_mask(mask, dilate_factor=19)
+        input_mask = Image.fromarray(mask)
+        mask_expanded = np.tile(np.expand_dims(mask, axis=-1), (1, 1, 3))
+        input_image_np = np.array(image)
+        input_image_np[mask_expanded == 255] = 0
+        input_image = Image.fromarray(input_image_np)
+        input = {
+            'img': input_image,
+            'mask': input_mask,
+        }
+        result = self.lama_model(input)
+        output_img = result['output_img']
+        return output_img[..., ::-1]
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('ANNOTATORS',
+                            __class__.__name__,
+                            LamaAnnotator.para_dict,
+                            set_name=True)

scepter/modules/annotator/mask_aug.py

@@ -0,0 +1,450 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import random
+from abc import ABCMeta
+from functools import partial
+
+import numpy as np
+import torch
+from PIL import Image, ImageDraw
+
+import cv2
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import Config
+from scepter.modules.utils.file_system import FS
+from scipy import ndimage
+from scipy.spatial import ConvexHull
+from skimage.draw import polygon
+
+
+@ANNOTATORS.register_class()
+class MaskDrawAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        self.task_type = cfg.get('TASK_TYPE', 'input_box')
+
+    def forward(self, mask=None, image=None, input_box=None, task_type=None):
+        task_type = task_type if task_type is not None else self.task_type
+
+        if mask is not None:
+            if isinstance(mask, Image.Image):
+                mask = np.array(mask)
+            elif isinstance(mask, torch.Tensor):
+                mask = mask.detach().cpu().numpy()
+            elif isinstance(mask, np.ndarray):
+                mask = mask.copy()
+            else:
+                raise f'Unsurpport datatype{type(mask)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+
+        if image is not None:
+            if isinstance(image, Image.Image):
+                image = np.array(image)
+            elif isinstance(image, torch.Tensor):
+                image = image.detach().cpu().numpy()
+            elif isinstance(image, np.ndarray):
+                image = image.copy()
+            else:
+                raise f'Unsurpport datatype{type(image)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+
+        mask_shape = mask.shape
+        if task_type == 'mask_point':
+            scribble = mask.transpose(1, 0)
+            labeled_array, num_features = ndimage.label(scribble >= 255)
+            centers = ndimage.center_of_mass(scribble, labeled_array,
+                                             range(1, num_features + 1))
+            centers = np.array(centers)
+            out_mask = np.zeros(mask_shape, dtype=np.uint8)
+            hull = ConvexHull(centers)
+            hull_vertices = centers[hull.vertices]
+            rr, cc = polygon(hull_vertices[:, 1], hull_vertices[:, 0],
+                             mask_shape)
+            out_mask[rr, cc] = 255
+        elif task_type == 'mask_box':
+            scribble = mask.transpose(1, 0)
+            labeled_array, num_features = ndimage.label(scribble >= 255)
+            centers = ndimage.center_of_mass(scribble, labeled_array,
+                                             range(1, num_features + 1))
+            centers = np.array(centers)
+            # (x1, y1, x2, y2)
+            x_min = centers[:, 0].min()
+            x_max = centers[:, 0].max()
+            y_min = centers[:, 1].min()
+            y_max = centers[:, 1].max()
+            out_mask = np.zeros(mask_shape, dtype=np.uint8)
+            out_mask[int(y_min):int(y_max) + 1,
+                     int(x_min):int(x_max) + 1] = 255
+            if image is not None:
+                out_image = image[int(y_min):int(y_max) + 1,
+                                  int(x_min):int(x_max) + 1]
+        elif task_type == 'input_box':
+            if isinstance(input_box, list):
+                input_box = np.array(input_box)
+            x_min, y_min, x_max, y_max = input_box
+            out_mask = np.zeros(mask_shape, dtype=np.uint8)
+            out_mask[int(y_min):int(y_max) + 1,
+                     int(x_min):int(x_max) + 1] = 255
+            if image is not None:
+                out_image = image[int(y_min):int(y_max) + 1,
+                                  int(x_min):int(x_max) + 1]
+        elif task_type == 'mask':
+            out_mask = mask
+        else:
+            raise NotImplementedError
+
+        if image is not None:
+            return out_image, out_mask
+        else:
+            return out_mask
+
+
+@ANNOTATORS.register_class()
+class MaskAugAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        # original / original_expand / hull / hull_expand / bbox / bbox_expand
+        self.mask_cfg = cfg.get('MASK_CFG', [{
+            'mode': 'original',
+            'proba': 0.1
+        }, {
+            'mode': 'original_expand',
+            'proba': 0.1
+        }, {
+            'mode': 'hull',
+            'proba': 0.1
+        }, {
+            'mode': 'hull_expand',
+            'proba': 0.1,
+            'kwargs': {
+                'expand_rate': 0.2
+            }
+        }, {
+            'mode': 'bbox',
+            'proba': 0.1
+        }, {
+            'mode': 'bbox_expand',
+            'proba': 0.1,
+            'kwargs': {
+                'min_expand_rate': 0.2,
+                'max_expand_rate': 0.5
+            }
+        }])
+        self.mask_cfg = Config.get_dict(self.mask_cfg) if isinstance(
+            self.mask_cfg, Config) else self.mask_cfg
+
+    def forward(self, mask, mask_cfg=None):
+        mask_cfg = mask_cfg if mask_cfg is not None else self.mask_cfg
+        if not isinstance(mask, list):
+            is_batch = False
+            masks = [mask]
+        else:
+            is_batch = True
+            masks = mask
+
+        mask_func = self.get_mask_func(mask_cfg)
+        # print(mask_func)
+        aug_masks = []
+        for submask in masks:
+            mask = self.get_mask(submask)
+            valid, large, h, w, bbox = self.get_mask_info(mask)
+            # print(valid, large, h, w, bbox)
+            if valid:
+                mask = mask_func(mask, bbox, h, w)
+            else:
+                mask = mask.astype(np.uint8)
+            aug_masks.append(mask)
+        return aug_masks if is_batch else aug_masks[0]
+
+    def get_mask(self, mask):
+        if isinstance(mask, Image.Image):
+            mask = np.array(mask)
+        elif isinstance(mask, torch.Tensor):
+            mask = mask.detach().cpu().numpy()
+        elif isinstance(mask, np.ndarray):
+            mask = mask.copy()
+        else:
+            raise f'Unsurpport datatype{type(mask)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+        return mask
+
+    def get_mask_info(self, mask):
+        h, w = mask.shape
+        locs = mask.nonzero()
+        valid = True
+        if len(locs) < 1 or locs[0].shape[0] < 1 or locs[1].shape[0] < 1:
+            valid = False
+            return valid, False, h, w, [0, 0, 0, 0]
+
+        left, right = np.min(locs[1]), np.max(locs[1])
+        top, bottom = np.min(locs[0]), np.max(locs[0])
+        bbox = [left, top, right, bottom]
+
+        large = False
+        if (right - left + 1) * (bottom - top + 1) > 0.9 * h * w:
+            large = True
+        return valid, large, h, w, bbox
+
+    def get_expand_params(self, mask_kwargs):
+        if 'expand_rate' in mask_kwargs:
+            expand_rate = mask_kwargs['expand_rate']
+        elif 'min_expand_rate' in mask_kwargs and 'max_expand_rate' in mask_kwargs:
+            expand_rate = random.uniform(mask_kwargs['min_expand_rate'],
+                                         mask_kwargs['max_expand_rate'])
+        else:
+            expand_rate = 0.3
+
+        if 'expand_iters' in mask_kwargs:
+            expand_iters = mask_kwargs['expand_iters']
+        else:
+            expand_iters = random.randint(1, 10)
+
+        if 'expand_lrtp' in mask_kwargs:
+            expand_lrtp = mask_kwargs['expand_lrtp']
+        else:
+            expand_lrtp = [
+                random.random(),
+                random.random(),
+                random.random(),
+                random.random()
+            ]
+
+        return expand_rate, expand_iters, expand_lrtp
+
+    def get_mask_func(self, mask_cfg):
+        if not isinstance(mask_cfg, list):
+            mask_cfg = [mask_cfg]
+        probas = [
+            item['proba'] if 'proba' in item else 1.0 / len(mask_cfg)
+            for item in mask_cfg
+        ]
+        sel_mask_cfg = random.choices(mask_cfg, weights=probas, k=1)[0]
+        mode = sel_mask_cfg['mode'] if 'mode' in sel_mask_cfg else 'original'
+        mask_kwargs = sel_mask_cfg[
+            'kwargs'] if 'kwargs' in sel_mask_cfg else {}
+
+        if mode == 'random':
+            mode = random.choice([
+                'original', 'original_expand', 'hull', 'hull_expand', 'bbox',
+                'bbox_expand'
+            ])
+        if mode == 'original':
+            mask_func = partial(self.generate_mask)
+        elif mode == 'original_expand':
+            expand_rate, expand_iters, expand_lrtp = self.get_expand_params(
+                mask_kwargs)
+            mask_func = partial(self.generate_mask,
+                                expand_rate=expand_rate,
+                                expand_iters=expand_iters,
+                                expand_lrtp=expand_lrtp)
+        elif mode == 'hull':
+            clockwise = random.choice([
+                True, False
+            ]) if 'clockwise' not in mask_kwargs else mask_kwargs['clockwise']
+            mask_func = partial(self.generate_hull_mask, clockwise=clockwise)
+        elif mode == 'hull_expand':
+            expand_rate, expand_iters, expand_lrtp = self.get_expand_params(
+                mask_kwargs)
+            clockwise = random.choice([
+                True, False
+            ]) if 'clockwise' not in mask_kwargs else mask_kwargs['clockwise']
+            mask_func = partial(self.generate_hull_mask,
+                                clockwise=clockwise,
+                                expand_rate=expand_rate,
+                                expand_iters=expand_iters,
+                                expand_lrtp=expand_lrtp)
+        elif mode == 'bbox':
+            mask_func = partial(self.generate_bbox_mask)
+        elif mode == 'bbox_expand':
+            expand_rate, expand_iters, expand_lrtp = self.get_expand_params(
+                mask_kwargs)
+            mask_func = partial(self.generate_bbox_mask,
+                                expand_rate=expand_rate,
+                                expand_iters=expand_iters,
+                                expand_lrtp=expand_lrtp)
+        else:
+            raise NotImplementedError
+        return mask_func
+
+    def generate_mask(self,
+                      mask,
+                      bbox,
+                      h,
+                      w,
+                      expand_rate=None,
+                      expand_iters=None,
+                      expand_lrtp=None):
+        bin_mask = mask.astype(np.uint8)
+        if expand_rate:
+            bin_mask = self.rand_expand_mask(bin_mask, bbox, h, w, expand_rate,
+                                             expand_iters, expand_lrtp)
+        return bin_mask
+
+    @staticmethod
+    def rand_expand_mask(mask,
+                         bbox,
+                         h,
+                         w,
+                         expand_rate=None,
+                         expand_iters=None,
+                         expand_lrtp=None):
+        expand_rate = 0.3 if expand_rate is None else expand_rate
+        expand_iters = random.randint(
+            1, 10) if expand_iters is None else expand_iters
+        expand_lrtp = [
+            random.random(),
+            random.random(),
+            random.random(),
+            random.random()
+        ] if expand_lrtp is None else expand_lrtp
+        # print('iters', expand_iters, 'expand_rate', expand_rate, 'expand_lrtp', expand_lrtp)
+        # mask = np.squeeze(mask)
+        left, top, right, bottom = bbox
+        # mask expansion
+        box_w = (right - left + 1) * expand_rate
+        box_h = (bottom - top + 1) * expand_rate
+        left_, right_ = int(
+            expand_lrtp[0] * min(box_w, left / 2) / expand_iters), int(
+                expand_lrtp[1] * min(box_w, (w - right) / 2) / expand_iters)
+        top_, bottom_ = int(
+            expand_lrtp[2] * min(box_h, top / 2) / expand_iters), int(
+                expand_lrtp[3] * min(box_h, (h - bottom) / 2) / expand_iters)
+        kernel_size = max(left_, right_, top_, bottom_)
+        if kernel_size > 0:
+            kernel = np.zeros((kernel_size * 2, kernel_size * 2),
+                              dtype=np.uint8)
+            new_left, new_right = kernel_size - right_, kernel_size + left_
+            new_top, new_bottom = kernel_size - bottom_, kernel_size + top_
+            kernel[new_top:new_bottom + 1, new_left:new_right + 1] = 1
+            mask = mask.astype(np.uint8)
+            mask = cv2.dilate(mask, kernel,
+                              iterations=expand_iters).astype(np.uint8)
+            # mask = new_mask - (mask / 2).astype(np.uint8)
+        # mask = np.expand_dims(mask, axis=-1)
+        return mask
+
+    @staticmethod
+    def _convexhull(image, clockwise):
+        # print('clockwise', clockwise)
+        contours, hierarchy = cv2.findContours(image, 2, 1)
+        cnt = np.concatenate(contours)  # merge all regions
+        hull = cv2.convexHull(cnt, clockwise=clockwise)
+        hull = np.squeeze(hull, axis=1).astype(np.float32).tolist()
+        hull = [tuple(x) for x in hull]
+        return hull  # b, 1, 2
+
+    def generate_hull_mask(self,
+                           mask,
+                           bbox,
+                           h,
+                           w,
+                           clockwise=None,
+                           expand_rate=None,
+                           expand_iters=None,
+                           expand_lrtp=None):
+        clockwise = random.choice([True, False
+                                   ]) if clockwise is None else clockwise
+        hull = self._convexhull(mask, clockwise)
+        mask_img = Image.new('L', (w, h), 0)
+        pt_list = hull
+        mask_img_draw = ImageDraw.Draw(mask_img)
+        mask_img_draw.polygon(pt_list, fill=255)
+        bin_mask = np.array(mask_img).astype(np.uint8)
+        if expand_rate:
+            bin_mask = self.rand_expand_mask(bin_mask, bbox, h, w, expand_rate,
+                                             expand_iters, expand_lrtp)
+        return bin_mask
+
+    def generate_bbox_mask(self,
+                           mask,
+                           bbox,
+                           h,
+                           w,
+                           expand_rate=None,
+                           expand_iters=None,
+                           expand_lrtp=None):
+        left, top, right, bottom = bbox
+        bin_mask = np.zeros((h, w), dtype=np.uint8)
+        bin_mask[top:bottom + 1, left:right + 1] = 255
+        if expand_rate:
+            bin_mask = self.rand_expand_mask(bin_mask, bbox, h, w, expand_rate,
+                                             expand_iters, expand_lrtp)
+        return bin_mask
+
+
+@ANNOTATORS.register_class()
+class MaskLayoutAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        ram_tag_color = cfg.get('RAM_TAG_COLOR', None)
+        default_color = cfg.get('DEFAULT_COLOR', [0, 0, 0])
+        self.use_aug = cfg.get('USE_AUG', False)
+        self.color_dict = {'default': tuple(default_color)}
+        if ram_tag_color is not None:
+            with FS.get_object(ram_tag_color) as object:
+                lines = object.decode('utf-8').strip().split('\n')
+            lines = [id_name_color.split('#;#') for id_name_color in lines]
+            self.color_dict.update({
+                id_name_color[1]: tuple(eval(id_name_color[2]))
+                for id_name_color in lines
+            })
+        if self.use_aug:
+            mask_aug_dict = {'NAME': 'MaskAugAnnotator'}
+            mask_aug_cfg = Config(cfg_dict=mask_aug_dict, load=False)
+            self.mask_aug_anno = ANNOTATORS.build(mask_aug_cfg)
+
+    def find_contours(self, mask):
+        # @mask: gray cv2 image
+        # contours, hier = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
+        contours, hier = cv2.findContours(mask, cv2.RETR_EXTERNAL,
+                                          cv2.CHAIN_APPROX_SIMPLE)
+        return contours
+
+    def draw_contours(self, canvas, contour, color):
+        canvas = np.ascontiguousarray(canvas, dtype=np.uint8)
+        canvas = cv2.drawContours(canvas, contour, -1, color, thickness=3)
+        return canvas
+
+    def get_mask(self, mask):
+        if isinstance(mask, Image.Image):
+            mask = np.array(mask)
+        elif isinstance(mask, torch.Tensor):
+            mask = mask.detach().cpu().numpy()
+        elif isinstance(mask, np.ndarray):
+            mask = mask.copy()
+        else:
+            raise f'Unsurpport datatype{type(mask)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+        return mask
+
+    def forward(self, mask=None, color=None, label=None, mask_cfg=None):
+        if not isinstance(mask, list):
+            is_batch = False
+            mask = [mask]
+        else:
+            is_batch = True
+
+        if label is not None and label in self.color_dict:
+            color = self.color_dict[label]
+        elif color is not None:
+            color = color
+        else:
+            color = self.color_dict['default']
+
+        ret_data = []
+        for sub_mask in mask:
+            sub_mask = self.get_mask(sub_mask)
+            if self.use_aug:
+                sub_mask = self.mask_aug_anno(sub_mask, mask_cfg)
+            canvas = np.ones((sub_mask.shape[0], sub_mask.shape[1], 3)) * 255
+            contour = self.find_contours(sub_mask)
+            frame = self.draw_contours(canvas, contour, color)
+            ret_data.append(frame)
+
+        if is_batch:
+            return ret_data
+        else:
+            return ret_data[0]

scepter/modules/annotator/midas/base_model.py

@@ -10,7 +10,7 @@ class BaseModel(torch.nn.Module):
         Args:
             path (str): file path
         """
-        parameters = torch.load(path, map_location=torch.device('cpu'))
+        parameters = torch.load(path, map_location=torch.device('cpu'), weights_only=True)
 
         if 'optimizer' in parameters:
             parameters = parameters['model']

scepter/modules/annotator/mlsd_op.py

@@ -29,7 +29,7 @@ class MLSDdetector(BaseAnnotator, metaclass=ABCMeta):
         pretrained_model = cfg.get('PRETRAINED_MODEL', None)
         if pretrained_model:
             with FS.get_from(pretrained_model, wait_finish=True) as local_path:
-                model.load_state_dict(torch.load(local_path), strict=True)
+                model.load_state_dict(torch.load(local_path, weights_only=True), strict=True)
         self.model = model.eval()
         self.thr_v = cfg.get('THR_V', 0.1)
         self.thr_d = cfg.get('THR_D', 0.1)

scepter/modules/annotator/openpose.py

@@ -9,9 +9,10 @@ import os
 from abc import ABCMeta
 from collections import OrderedDict
 
+import numpy as np
+
 import cv2
 import matplotlib
-import numpy as np
 import torch
 import torch.nn as nn
 from PIL import Image
@@ -422,7 +423,7 @@ class Hand(object):
         self.model = handpose_model()
         if torch.cuda.is_available():
             self.model = self.model.to(device)
-        model_dict = transfer(self.model, torch.load(model_path))
+        model_dict = transfer(self.model, torch.load(model_path, weights_only=True))
         self.model.load_state_dict(model_dict)
         self.model.eval()
         self.device = device
@@ -478,7 +479,7 @@ class Hand(object):
             map_ori = heatmap_avg[:, :, part]
             one_heatmap = gaussian_filter(map_ori, sigma=3)
             binary = np.ascontiguousarray(one_heatmap > thre, dtype=np.uint8)
-            # 全部小于阈值
+
             if np.sum(binary) == 0:
                 all_peaks.append([0, 0])
                 continue
@@ -502,7 +503,7 @@ class Body(object):
         self.model = bodypose_model()
         if torch.cuda.is_available():
             self.model = self.model.to(device)
-        model_dict = transfer(self.model, torch.load(model_path))
+        model_dict = transfer(self.model, torch.load(model_path, weights_only=True))
         self.model.load_state_dict(model_dict)
         self.model.eval()
         self.device = device
@@ -788,7 +789,7 @@ class OpenposeAnnotator(BaseAnnotator, metaclass=ABCMeta):
             raise f'Unsurpport datatype{type(image)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
         image = image[:, :, ::-1]
         candidate, subset = self.body_estimation(image)
-        canvas = np.zeros_like(image)
+        canvas = np.zeros_like(image, order='C')   # to check
         canvas = draw_bodypose(canvas, candidate, subset)
         if self.use_hand:
             hands_list = handDetect(candidate, subset, image)

scepter/modules/annotator/outpainting.py

@@ -0,0 +1,205 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import math
+import random
+from abc import ABCMeta
+
+import numpy as np
+import torch
+from PIL import Image, ImageDraw
+
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import dict_to_yaml
+
+
+@ANNOTATORS.register_class()
+class OutpaintingAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        self.mask_blur = cfg.get('MASK_BLUR', 0)
+        self.random_cfg = cfg.get('RANDOM_CFG', None)
+        self.return_mask = cfg.get('RETURN_MASK', False)
+        self.return_source = cfg.get('RETURN_SOURCE', True)
+        self.keep_padding_ratio = cfg.get('KEEP_PADDING_RATIO', 64)
+        self.mask_color = cfg.get('MASK_COLOR', 0)
+
+    def get_box(self, mask):
+        locs = np.where(mask == 255)
+        if len(locs) < 1 or locs[0].shape[0] < 1 or locs[1].shape[0] < 1:
+            return None
+        left, right = np.min(locs[1]), np.max(locs[1])
+        top, bottom = np.min(locs[0]), np.max(locs[0])
+        return [left, top, right, bottom]
+
+    def forward(self,
+                image,
+                ratio=0.3,
+                mask=None,
+                direction=['left', 'right', 'up', 'down'],
+                return_mask=None,
+                return_source=None,
+                mask_color=None):
+        return_mask = return_mask if return_mask is not None else self.return_mask
+        return_source = return_source if return_source is not None else self.return_source
+        mask_color = mask_color if mask_color is not None else self.mask_color
+        if isinstance(image, Image.Image):
+            image = image
+        elif isinstance(image, torch.Tensor):
+            image = Image.fromarray(image.detach().cpu().numpy())
+        elif isinstance(image, np.ndarray):
+            image = Image.fromarray(image.copy())
+        else:
+            raise f'Unsurpport datatype{type(image)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+
+        if self.random_cfg:
+            direction_range = self.random_cfg.get(
+                'DIRECTION_RANGE', ['left', 'right', 'up', 'down'])
+            ratio_range = self.random_cfg.get('RATIO_RANGE', [0.0, 1.0])
+            direction = random.sample(
+                direction_range,
+                random.choice(list(range(1,
+                                         len(direction_range) + 1))))
+            ratio = random.uniform(ratio_range[0], ratio_range[1])
+
+        if mask is None:
+            init_image = image
+            src_width, src_height = init_image.width, init_image.height
+            left = int(ratio * src_width) if 'left' in direction else 0
+            right = int(ratio * src_width) if 'right' in direction else 0
+            up = int(ratio * src_height) if 'up' in direction else 0
+            down = int(ratio * src_height) if 'down' in direction else 0
+            # print(direction, ratio, left, right, up, down)
+            tar_width = math.ceil(
+                (src_width + left + right) /
+                self.keep_padding_ratio) * self.keep_padding_ratio
+            tar_height = math.ceil(
+                (src_height + up + down) /
+                self.keep_padding_ratio) * self.keep_padding_ratio
+            if left > 0:
+                left = left * (tar_width - src_width) // (left + right)
+            if right > 0:
+                right = tar_width - src_width - left
+            if up > 0:
+                up = up * (tar_height - src_height) // (up + down)
+            if down > 0:
+                down = tar_height - src_height - up
+            if mask_color is not None:
+                img = Image.new('RGB', (tar_width, tar_height),
+                                color=mask_color)
+            else:
+                img = Image.new('RGB', (tar_width, tar_height))
+            img.paste(init_image, (left, up))
+            mask = Image.new('L', (img.width, img.height), 'white')
+            draw = ImageDraw.Draw(mask)
+
+            draw.rectangle(
+                (left + (self.mask_blur * 2 if left > 0 else 0), up +
+                 (self.mask_blur * 2 if up > 0 else 0), mask.width - right -
+                 (self.mask_blur * 2 if right > 0 else 0) - 1, mask.height - down -
+                 (self.mask_blur * 2 if down > 0 else 0) - 1),
+                fill='black')
+            # draw.rectangle(
+            #     (left + (self.mask_blur * 2 if left > 0 else 0), up +
+            #      (self.mask_blur * 2 if up > 0 else 0), left + src_width -
+            #      (self.mask_blur * 2 if right > 0 else 0), up + src_height -
+            #      (self.mask_blur * 2 if down > 0 else 0)),
+            #     fill='black')
+        else:
+            bbox = self.get_box(np.array(mask))
+            if bbox is None:
+                img = image
+                mask = mask
+                init_image = image
+            else:
+                mask = Image.new('L', (image.width, image.height), 'white')
+                mask_zero = Image.new('L',
+                                      (bbox[2] - bbox[0], bbox[3] - bbox[1]),
+                                      'black')
+                mask.paste(mask_zero, (bbox[0], bbox[1]))
+                crop_image = image.crop(bbox)
+                init_image = Image.new('RGB', (image.width, image.height),
+                                       'black')
+                init_image.paste(crop_image, (bbox[0], bbox[1]))
+                img = image
+        if return_mask:
+            if return_source:
+                ret_data = {
+                    'src_image': np.array(init_image),
+                    'image': np.array(img),
+                    'mask': np.array(mask)
+                }
+            else:
+                ret_data = {'image': np.array(img), 'mask': np.array(mask)}
+        else:
+            if return_source:
+                ret_data = {
+                    'src_image': np.array(init_image),
+                    'image': np.array(img)
+                }
+            else:
+                ret_data = np.array(img)
+        return ret_data
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('ANNOTATORS',
+                            __class__.__name__,
+                            OutpaintingAnnotator.para_dict,
+                            set_name=True)
+
+
+@ANNOTATORS.register_class()
+class OutpaintingResize(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+
+    def get_box(self, mask):
+        locs = np.where(mask == 0)
+        if len(locs) < 1 or locs[0].shape[0] < 1 or locs[1].shape[0] < 1:
+            return None
+        left, right = np.min(locs[1]), np.max(locs[1])
+        top, bottom = np.min(locs[0]), np.max(locs[0])
+        return [left, top, right, bottom]
+
+    def forward(self, image, target_image, mask=None):
+        if isinstance(image, Image.Image):
+            image = image
+        elif isinstance(image, torch.Tensor):
+            image = Image.fromarray(image.detach().cpu().numpy())
+        elif isinstance(image, np.ndarray):
+            image = Image.fromarray(image.copy())
+        else:
+            raise f'Unsurpport datatype{type(image)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+
+        if isinstance(target_image, Image.Image):
+            target_image = target_image
+        elif isinstance(target_image, torch.Tensor):
+            target_image = Image.fromarray(target_image.detach().cpu().numpy())
+        elif isinstance(target_image, np.ndarray):
+            target_image = Image.fromarray(target_image.copy())
+        else:
+            raise f'Unsurpport datatype{type(target_image)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+
+        bbox = self.get_box(np.array(mask))
+        if bbox is None:
+            init_image = image
+        else:
+            paste_img = image.resize((bbox[2] - bbox[0], bbox[3] - bbox[1]))
+            init_image = Image.new('RGB',
+                                   (target_image.width, target_image.height),
+                                   'black')
+            init_image.paste(paste_img, (bbox[0], bbox[1]))
+        ret_data = {'src_image': np.array(init_image)}
+        return ret_data
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('ANNOTATORS',
+                            __class__.__name__,
+                            OutpaintingResize.para_dict,
+                            set_name=True)

scepter/modules/annotator/pidinet.py

@@ -0,0 +1,934 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import math
+from abc import ABCMeta
+
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import dict_to_yaml
+from scepter.modules.utils.distribute import we
+from scepter.modules.utils.file_system import FS
+
+CONFIGS = {
+    'baseline': {
+        'layer0': 'cv',
+        'layer1': 'cv',
+        'layer2': 'cv',
+        'layer3': 'cv',
+        'layer4': 'cv',
+        'layer5': 'cv',
+        'layer6': 'cv',
+        'layer7': 'cv',
+        'layer8': 'cv',
+        'layer9': 'cv',
+        'layer10': 'cv',
+        'layer11': 'cv',
+        'layer12': 'cv',
+        'layer13': 'cv',
+        'layer14': 'cv',
+        'layer15': 'cv',
+    },
+    'c-v15': {
+        'layer0': 'cd',
+        'layer1': 'cv',
+        'layer2': 'cv',
+        'layer3': 'cv',
+        'layer4': 'cv',
+        'layer5': 'cv',
+        'layer6': 'cv',
+        'layer7': 'cv',
+        'layer8': 'cv',
+        'layer9': 'cv',
+        'layer10': 'cv',
+        'layer11': 'cv',
+        'layer12': 'cv',
+        'layer13': 'cv',
+        'layer14': 'cv',
+        'layer15': 'cv',
+    },
+    'a-v15': {
+        'layer0': 'ad',
+        'layer1': 'cv',
+        'layer2': 'cv',
+        'layer3': 'cv',
+        'layer4': 'cv',
+        'layer5': 'cv',
+        'layer6': 'cv',
+        'layer7': 'cv',
+        'layer8': 'cv',
+        'layer9': 'cv',
+        'layer10': 'cv',
+        'layer11': 'cv',
+        'layer12': 'cv',
+        'layer13': 'cv',
+        'layer14': 'cv',
+        'layer15': 'cv',
+    },
+    'r-v15': {
+        'layer0': 'rd',
+        'layer1': 'cv',
+        'layer2': 'cv',
+        'layer3': 'cv',
+        'layer4': 'cv',
+        'layer5': 'cv',
+        'layer6': 'cv',
+        'layer7': 'cv',
+        'layer8': 'cv',
+        'layer9': 'cv',
+        'layer10': 'cv',
+        'layer11': 'cv',
+        'layer12': 'cv',
+        'layer13': 'cv',
+        'layer14': 'cv',
+        'layer15': 'cv',
+    },
+    'cvvv4': {
+        'layer0': 'cd',
+        'layer1': 'cv',
+        'layer2': 'cv',
+        'layer3': 'cv',
+        'layer4': 'cd',
+        'layer5': 'cv',
+        'layer6': 'cv',
+        'layer7': 'cv',
+        'layer8': 'cd',
+        'layer9': 'cv',
+        'layer10': 'cv',
+        'layer11': 'cv',
+        'layer12': 'cd',
+        'layer13': 'cv',
+        'layer14': 'cv',
+        'layer15': 'cv',
+    },
+    'avvv4': {
+        'layer0': 'ad',
+        'layer1': 'cv',
+        'layer2': 'cv',
+        'layer3': 'cv',
+        'layer4': 'ad',
+        'layer5': 'cv',
+        'layer6': 'cv',
+        'layer7': 'cv',
+        'layer8': 'ad',
+        'layer9': 'cv',
+        'layer10': 'cv',
+        'layer11': 'cv',
+        'layer12': 'ad',
+        'layer13': 'cv',
+        'layer14': 'cv',
+        'layer15': 'cv',
+    },
+    'rvvv4': {
+        'layer0': 'rd',
+        'layer1': 'cv',
+        'layer2': 'cv',
+        'layer3': 'cv',
+        'layer4': 'rd',
+        'layer5': 'cv',
+        'layer6': 'cv',
+        'layer7': 'cv',
+        'layer8': 'rd',
+        'layer9': 'cv',
+        'layer10': 'cv',
+        'layer11': 'cv',
+        'layer12': 'rd',
+        'layer13': 'cv',
+        'layer14': 'cv',
+        'layer15': 'cv',
+    },
+    'cccv4': {
+        'layer0': 'cd',
+        'layer1': 'cd',
+        'layer2': 'cd',
+        'layer3': 'cv',
+        'layer4': 'cd',
+        'layer5': 'cd',
+        'layer6': 'cd',
+        'layer7': 'cv',
+        'layer8': 'cd',
+        'layer9': 'cd',
+        'layer10': 'cd',
+        'layer11': 'cv',
+        'layer12': 'cd',
+        'layer13': 'cd',
+        'layer14': 'cd',
+        'layer15': 'cv',
+    },
+    'aaav4': {
+        'layer0': 'ad',
+        'layer1': 'ad',
+        'layer2': 'ad',
+        'layer3': 'cv',
+        'layer4': 'ad',
+        'layer5': 'ad',
+        'layer6': 'ad',
+        'layer7': 'cv',
+        'layer8': 'ad',
+        'layer9': 'ad',
+        'layer10': 'ad',
+        'layer11': 'cv',
+        'layer12': 'ad',
+        'layer13': 'ad',
+        'layer14': 'ad',
+        'layer15': 'cv',
+    },
+    'rrrv4': {
+        'layer0': 'rd',
+        'layer1': 'rd',
+        'layer2': 'rd',
+        'layer3': 'cv',
+        'layer4': 'rd',
+        'layer5': 'rd',
+        'layer6': 'rd',
+        'layer7': 'cv',
+        'layer8': 'rd',
+        'layer9': 'rd',
+        'layer10': 'rd',
+        'layer11': 'cv',
+        'layer12': 'rd',
+        'layer13': 'rd',
+        'layer14': 'rd',
+        'layer15': 'cv',
+    },
+    'c16': {
+        'layer0': 'cd',
+        'layer1': 'cd',
+        'layer2': 'cd',
+        'layer3': 'cd',
+        'layer4': 'cd',
+        'layer5': 'cd',
+        'layer6': 'cd',
+        'layer7': 'cd',
+        'layer8': 'cd',
+        'layer9': 'cd',
+        'layer10': 'cd',
+        'layer11': 'cd',
+        'layer12': 'cd',
+        'layer13': 'cd',
+        'layer14': 'cd',
+        'layer15': 'cd',
+    },
+    'a16': {
+        'layer0': 'ad',
+        'layer1': 'ad',
+        'layer2': 'ad',
+        'layer3': 'ad',
+        'layer4': 'ad',
+        'layer5': 'ad',
+        'layer6': 'ad',
+        'layer7': 'ad',
+        'layer8': 'ad',
+        'layer9': 'ad',
+        'layer10': 'ad',
+        'layer11': 'ad',
+        'layer12': 'ad',
+        'layer13': 'ad',
+        'layer14': 'ad',
+        'layer15': 'ad',
+    },
+    'r16': {
+        'layer0': 'rd',
+        'layer1': 'rd',
+        'layer2': 'rd',
+        'layer3': 'rd',
+        'layer4': 'rd',
+        'layer5': 'rd',
+        'layer6': 'rd',
+        'layer7': 'rd',
+        'layer8': 'rd',
+        'layer9': 'rd',
+        'layer10': 'rd',
+        'layer11': 'rd',
+        'layer12': 'rd',
+        'layer13': 'rd',
+        'layer14': 'rd',
+        'layer15': 'rd',
+    },
+    'carv4': {
+        'layer0': 'cd',
+        'layer1': 'ad',
+        'layer2': 'rd',
+        'layer3': 'cv',
+        'layer4': 'cd',
+        'layer5': 'ad',
+        'layer6': 'rd',
+        'layer7': 'cv',
+        'layer8': 'cd',
+        'layer9': 'ad',
+        'layer10': 'rd',
+        'layer11': 'cv',
+        'layer12': 'cd',
+        'layer13': 'ad',
+        'layer14': 'rd',
+        'layer15': 'cv'
+    }
+}
+
+
+def create_conv_func(op_type):
+    assert op_type in ['cv', 'cd', 'ad',
+                       'rd'], 'unknown op type: %s' % str(op_type)
+    if op_type == 'cv':
+        return F.conv2d
+    if op_type == 'cd':
+
+        def func(x,
+                 weights,
+                 bias=None,
+                 stride=1,
+                 padding=0,
+                 dilation=1,
+                 groups=1):
+            assert dilation in [1,
+                                2], 'dilation for cd_conv should be in 1 or 2'
+            assert weights.size(2) == 3 and weights.size(3) == 3, \
+                'kernel size for cd_conv should be 3x3'
+            assert padding == dilation, 'padding for cd_conv set wrong'
+
+            weights_c = weights.sum(dim=[2, 3], keepdim=True)
+            yc = F.conv2d(x,
+                          weights_c,
+                          stride=stride,
+                          padding=0,
+                          groups=groups)
+            y = F.conv2d(x,
+                         weights,
+                         bias,
+                         stride=stride,
+                         padding=padding,
+                         dilation=dilation,
+                         groups=groups)
+            return y - yc
+
+        return func
+    elif op_type == 'ad':
+
+        def func(x,
+                 weights,
+                 bias=None,
+                 stride=1,
+                 padding=0,
+                 dilation=1,
+                 groups=1):
+            assert dilation in [1,
+                                2], 'dilation for ad_conv should be in 1 or 2'
+            assert weights.size(2) == 3 and weights.size(3) == 3, \
+                'kernel size for ad_conv should be 3x3'
+            assert padding == dilation, 'padding for ad_conv set wrong'
+
+            shape = weights.shape
+            weights = weights.view(shape[0], shape[1], -1)
+            # clock-wise
+            weights_conv = (
+                weights -
+                weights[:, :, [3, 0, 1, 6, 4, 2, 7, 8, 5]]).view(shape)
+            y = F.conv2d(x,
+                         weights_conv,
+                         bias,
+                         stride=stride,
+                         padding=padding,
+                         dilation=dilation,
+                         groups=groups)
+            return y
+
+        return func
+    elif op_type == 'rd':
+
+        def func(x,
+                 weights,
+                 bias=None,
+                 stride=1,
+                 padding=0,
+                 dilation=1,
+                 groups=1):
+            assert dilation in [1,
+                                2], 'dilation for rd_conv should be in 1 or 2'
+            assert weights.size(2) == 3 and weights.size(3) == 3, \
+                'kernel size for rd_conv should be 3x3'
+            padding = 2 * dilation
+
+            shape = weights.shape
+            if weights.is_cuda:
+                buffer = torch.cuda.FloatTensor(shape[0], shape[1],
+                                                5 * 5).fill_(0)
+            else:
+                buffer = torch.zeros(shape[0], shape[1], 5 * 5)
+            weights = weights.view(shape[0], shape[1], -1)
+            buffer[:, :, [0, 2, 4, 10, 14, 20, 22, 24]] = weights[:, :, 1:]
+            buffer[:, :, [6, 7, 8, 11, 13, 16, 17, 18]] = -weights[:, :, 1:]
+            buffer[:, :, 12] = 0
+            buffer = buffer.view(shape[0], shape[1], 5, 5)
+            y = F.conv2d(x,
+                         buffer,
+                         bias,
+                         stride=stride,
+                         padding=padding,
+                         dilation=dilation,
+                         groups=groups)
+            return y
+
+        return func
+    else:
+        print('impossible to be here unless you force that', flush=True)
+        return None
+
+
+def config_model(model):
+    model_options = list(CONFIGS.keys())
+    assert model in model_options, \
+        'unrecognized model, please choose from %s' % str(model_options)
+
+    pdcs = []
+    for i in range(16):
+        layer_name = 'layer%d' % i
+        op = CONFIGS[model][layer_name]
+        pdcs.append(create_conv_func(op))
+    return pdcs
+
+
+def config_model_converted(model):
+    model_options = list(CONFIGS.keys())
+    assert model in model_options, \
+        'unrecognized model, please choose from %s' % str(model_options)
+
+    pdcs = []
+    for i in range(16):
+        layer_name = 'layer%d' % i
+        op = CONFIGS[model][layer_name]
+        pdcs.append(op)
+    return pdcs
+
+
+def convert_pdc(op, weight):
+    if op == 'cv':
+        return weight
+    elif op == 'cd':
+        shape = weight.shape
+        weight_c = weight.sum(dim=[2, 3])
+        weight = weight.view(shape[0], shape[1], -1)
+        weight[:, :, 4] = weight[:, :, 4] - weight_c
+        weight = weight.view(shape)
+        return weight
+    elif op == 'ad':
+        shape = weight.shape
+        weight = weight.view(shape[0], shape[1], -1)
+        weight_conv = (weight -
+                       weight[:, :, [3, 0, 1, 6, 4, 2, 7, 8, 5]]).view(shape)
+        return weight_conv
+    elif op == 'rd':
+        shape = weight.shape
+        buffer = torch.zeros(shape[0], shape[1], 5 * 5, device=weight.device)
+        weight = weight.view(shape[0], shape[1], -1)
+        buffer[:, :, [0, 2, 4, 10, 14, 20, 22, 24]] = weight[:, :, 1:]
+        buffer[:, :, [6, 7, 8, 11, 13, 16, 17, 18]] = -weight[:, :, 1:]
+        buffer = buffer.view(shape[0], shape[1], 5, 5)
+        return buffer
+    raise ValueError('wrong op {}'.format(str(op)))
+
+
+def convert_pidinet(state_dict, config):
+    pdcs = config_model_converted(config)
+    new_dict = {}
+    for pname, p in state_dict.items():
+        if 'init_block.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[0], p)
+        elif 'block1_1.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[1], p)
+        elif 'block1_2.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[2], p)
+        elif 'block1_3.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[3], p)
+        elif 'block2_1.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[4], p)
+        elif 'block2_2.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[5], p)
+        elif 'block2_3.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[6], p)
+        elif 'block2_4.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[7], p)
+        elif 'block3_1.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[8], p)
+        elif 'block3_2.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[9], p)
+        elif 'block3_3.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[10], p)
+        elif 'block3_4.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[11], p)
+        elif 'block4_1.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[12], p)
+        elif 'block4_2.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[13], p)
+        elif 'block4_3.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[14], p)
+        elif 'block4_4.conv1.weight' in pname:
+            new_dict[pname] = convert_pdc(pdcs[15], p)
+        else:
+            new_dict[pname] = p
+    return new_dict
+
+
+class Conv2d(nn.Module):
+    def __init__(self,
+                 pdc,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride=1,
+                 padding=0,
+                 dilation=1,
+                 groups=1,
+                 bias=False):
+        super().__init__()
+        if in_channels % groups != 0:
+            raise ValueError('in_channels must be divisible by groups')
+        if out_channels % groups != 0:
+            raise ValueError('out_channels must be divisible by groups')
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.stride = stride
+        self.padding = padding
+        self.dilation = dilation
+        self.groups = groups
+        self.weight = nn.Parameter(
+            torch.Tensor(out_channels, in_channels // groups, kernel_size,
+                         kernel_size))
+        if bias:
+            self.bias = nn.Parameter(torch.Tensor(out_channels))
+        else:
+            self.register_parameter('bias', None)
+        self.reset_parameters()
+        self.pdc = pdc
+
+    def reset_parameters(self):
+        nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5))
+        if self.bias is not None:
+            fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.weight)
+            bound = 1 / math.sqrt(fan_in)
+            nn.init.uniform_(self.bias, -bound, bound)
+
+    def forward(self, input):
+        return self.pdc(input, self.weight, self.bias, self.stride,
+                        self.padding, self.dilation, self.groups)
+
+
+class CSAM(nn.Module):
+    """
+    Compact Spatial Attention Module
+    """
+    def __init__(self, channels):
+        super().__init__()
+
+        mid_channels = 4
+        self.relu1 = nn.ReLU()
+        self.conv1 = nn.Conv2d(channels,
+                               mid_channels,
+                               kernel_size=1,
+                               padding=0)
+        self.conv2 = nn.Conv2d(mid_channels,
+                               1,
+                               kernel_size=3,
+                               padding=1,
+                               bias=False)
+        self.sigmoid = nn.Sigmoid()
+        nn.init.constant_(self.conv1.bias, 0)
+
+    def forward(self, x):
+        y = self.relu1(x)
+        y = self.conv1(y)
+        y = self.conv2(y)
+        y = self.sigmoid(y)
+
+        return x * y
+
+
+class CDCM(nn.Module):
+    """
+    Compact Dilation Convolution based Module
+    """
+    def __init__(self, in_channels, out_channels):
+        super().__init__()
+
+        self.relu1 = nn.ReLU()
+        self.conv1 = nn.Conv2d(in_channels,
+                               out_channels,
+                               kernel_size=1,
+                               padding=0)
+        self.conv2_1 = nn.Conv2d(out_channels,
+                                 out_channels,
+                                 kernel_size=3,
+                                 dilation=5,
+                                 padding=5,
+                                 bias=False)
+        self.conv2_2 = nn.Conv2d(out_channels,
+                                 out_channels,
+                                 kernel_size=3,
+                                 dilation=7,
+                                 padding=7,
+                                 bias=False)
+        self.conv2_3 = nn.Conv2d(out_channels,
+                                 out_channels,
+                                 kernel_size=3,
+                                 dilation=9,
+                                 padding=9,
+                                 bias=False)
+        self.conv2_4 = nn.Conv2d(out_channels,
+                                 out_channels,
+                                 kernel_size=3,
+                                 dilation=11,
+                                 padding=11,
+                                 bias=False)
+        nn.init.constant_(self.conv1.bias, 0)
+
+    def forward(self, x):
+        x = self.relu1(x)
+        x = self.conv1(x)
+        x1 = self.conv2_1(x)
+        x2 = self.conv2_2(x)
+        x3 = self.conv2_3(x)
+        x4 = self.conv2_4(x)
+        return x1 + x2 + x3 + x4
+
+
+class MapReduce(nn.Module):
+    """
+    Reduce feature maps into a single edge map
+    """
+    def __init__(self, channels):
+        super().__init__()
+        self.conv = nn.Conv2d(channels, 1, kernel_size=1, padding=0)
+        nn.init.constant_(self.conv.bias, 0)
+
+    def forward(self, x):
+        return self.conv(x)
+
+
+class PDCBlock(nn.Module):
+    def __init__(self, pdc, inplane, ouplane, stride=1):
+        super().__init__()
+        self.stride = stride
+
+        self.stride = stride
+        if self.stride > 1:
+            self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
+            self.shortcut = nn.Conv2d(inplane,
+                                      ouplane,
+                                      kernel_size=1,
+                                      padding=0)
+        self.conv1 = Conv2d(pdc,
+                            inplane,
+                            inplane,
+                            kernel_size=3,
+                            padding=1,
+                            groups=inplane,
+                            bias=False)
+        self.relu2 = nn.ReLU()
+        self.conv2 = nn.Conv2d(inplane,
+                               ouplane,
+                               kernel_size=1,
+                               padding=0,
+                               bias=False)
+
+    def forward(self, x):
+        if self.stride > 1:
+            x = self.pool(x)
+        y = self.conv1(x)
+        y = self.relu2(y)
+        y = self.conv2(y)
+        if self.stride > 1:
+            x = self.shortcut(x)
+        y = y + x
+        return y
+
+
+class PDCBlock_converted(nn.Module):
+    """
+    CPDC, APDC can be converted to vanilla 3x3 convolution
+    RPDC can be converted to vanilla 5x5 convolution
+    """
+    def __init__(self, pdc, inplane, ouplane, stride=1):
+        super().__init__()
+        self.stride = stride
+
+        if self.stride > 1:
+            self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
+            self.shortcut = nn.Conv2d(inplane,
+                                      ouplane,
+                                      kernel_size=1,
+                                      padding=0)
+        if pdc == 'rd':
+            self.conv1 = nn.Conv2d(inplane,
+                                   inplane,
+                                   kernel_size=5,
+                                   padding=2,
+                                   groups=inplane,
+                                   bias=False)
+        else:
+            self.conv1 = nn.Conv2d(inplane,
+                                   inplane,
+                                   kernel_size=3,
+                                   padding=1,
+                                   groups=inplane,
+                                   bias=False)
+        self.relu2 = nn.ReLU()
+        self.conv2 = nn.Conv2d(inplane,
+                               ouplane,
+                               kernel_size=1,
+                               padding=0,
+                               bias=False)
+
+    def forward(self, x):
+        if self.stride > 1:
+            x = self.pool(x)
+        y = self.conv1(x)
+        y = self.relu2(y)
+        y = self.conv2(y)
+        if self.stride > 1:
+            x = self.shortcut(x)
+        y = y + x
+        return y
+
+
+class PiDiNet(nn.Module):
+    def __init__(self,
+                 inplane,
+                 pdcs,
+                 dil=None,
+                 sa=False,
+                 convert=False,
+                 mean=[0.485, 0.456, 0.406],
+                 std=[0.229, 0.224, 0.225]):
+        super().__init__()
+        self.sa = sa
+        if dil is not None:
+            assert isinstance(dil, int), 'dil should be an int'
+        self.dil = dil
+        self.mean = mean
+        self.std = std
+
+        self.fuseplanes = []
+
+        self.inplane = inplane
+        if convert:
+            if pdcs[0] == 'rd':
+                init_kernel_size = 5
+                init_padding = 2
+            else:
+                init_kernel_size = 3
+                init_padding = 1
+            self.init_block = nn.Conv2d(3,
+                                        self.inplane,
+                                        kernel_size=init_kernel_size,
+                                        padding=init_padding,
+                                        bias=False)
+            block_class = PDCBlock_converted
+        else:
+            self.init_block = Conv2d(pdcs[0],
+                                     3,
+                                     self.inplane,
+                                     kernel_size=3,
+                                     padding=1)
+            block_class = PDCBlock
+
+        self.block1_1 = block_class(pdcs[1], self.inplane, self.inplane)
+        self.block1_2 = block_class(pdcs[2], self.inplane, self.inplane)
+        self.block1_3 = block_class(pdcs[3], self.inplane, self.inplane)
+        self.fuseplanes.append(self.inplane)  # C
+
+        inplane = self.inplane
+        self.inplane = self.inplane * 2
+        self.block2_1 = block_class(pdcs[4], inplane, self.inplane, stride=2)
+        self.block2_2 = block_class(pdcs[5], self.inplane, self.inplane)
+        self.block2_3 = block_class(pdcs[6], self.inplane, self.inplane)
+        self.block2_4 = block_class(pdcs[7], self.inplane, self.inplane)
+        self.fuseplanes.append(self.inplane)  # 2C
+
+        inplane = self.inplane
+        self.inplane = self.inplane * 2
+        self.block3_1 = block_class(pdcs[8], inplane, self.inplane, stride=2)
+        self.block3_2 = block_class(pdcs[9], self.inplane, self.inplane)
+        self.block3_3 = block_class(pdcs[10], self.inplane, self.inplane)
+        self.block3_4 = block_class(pdcs[11], self.inplane, self.inplane)
+        self.fuseplanes.append(self.inplane)  # 4C
+
+        self.block4_1 = block_class(pdcs[12],
+                                    self.inplane,
+                                    self.inplane,
+                                    stride=2)
+        self.block4_2 = block_class(pdcs[13], self.inplane, self.inplane)
+        self.block4_3 = block_class(pdcs[14], self.inplane, self.inplane)
+        self.block4_4 = block_class(pdcs[15], self.inplane, self.inplane)
+        self.fuseplanes.append(self.inplane)  # 4C
+
+        self.conv_reduces = nn.ModuleList()
+        if self.sa and self.dil is not None:
+            self.attentions = nn.ModuleList()
+            self.dilations = nn.ModuleList()
+            for i in range(4):
+                self.dilations.append(CDCM(self.fuseplanes[i], self.dil))
+                self.attentions.append(CSAM(self.dil))
+                self.conv_reduces.append(MapReduce(self.dil))
+        elif self.sa:
+            self.attentions = nn.ModuleList()
+            for i in range(4):
+                self.attentions.append(CSAM(self.fuseplanes[i]))
+                self.conv_reduces.append(MapReduce(self.fuseplanes[i]))
+        elif self.dil is not None:
+            self.dilations = nn.ModuleList()
+            for i in range(4):
+                self.dilations.append(CDCM(self.fuseplanes[i], self.dil))
+                self.conv_reduces.append(MapReduce(self.dil))
+        else:
+            for i in range(4):
+                self.conv_reduces.append(MapReduce(self.fuseplanes[i]))
+
+        self.classifier = nn.Conv2d(4, 1, kernel_size=1)  # has bias
+        nn.init.constant_(self.classifier.weight, 0.25)
+        nn.init.constant_(self.classifier.bias, 0)
+
+    def get_weights(self):
+        conv_weights = []
+        bn_weights = []
+        relu_weights = []
+        for pname, p in self.named_parameters():
+            if 'bn' in pname:
+                bn_weights.append(p)
+            elif 'relu' in pname:
+                relu_weights.append(p)
+            else:
+                conv_weights.append(p)
+
+        return conv_weights, bn_weights, relu_weights
+
+    def forward(self, x):
+        """x: [B, 3, H, W] within range [0, 1].
+        """
+        x = (x - x.new_tensor(self.mean).view(1, -1, 1, 1)) / \
+            x.new_tensor(self.std).view(1, -1, 1, 1)
+        h, w = x.size()[2:]
+
+        x = self.init_block(x)
+
+        x1 = self.block1_1(x)
+        x1 = self.block1_2(x1)
+        x1 = self.block1_3(x1)
+
+        x2 = self.block2_1(x1)
+        x2 = self.block2_2(x2)
+        x2 = self.block2_3(x2)
+        x2 = self.block2_4(x2)
+
+        x3 = self.block3_1(x2)
+        x3 = self.block3_2(x3)
+        x3 = self.block3_3(x3)
+        x3 = self.block3_4(x3)
+
+        x4 = self.block4_1(x3)
+        x4 = self.block4_2(x4)
+        x4 = self.block4_3(x4)
+        x4 = self.block4_4(x4)
+
+        x_fuses = []
+        if self.sa and self.dil is not None:
+            for i, xi in enumerate([x1, x2, x3, x4]):
+                x_fuses.append(self.attentions[i](self.dilations[i](xi)))
+        elif self.sa:
+            for i, xi in enumerate([x1, x2, x3, x4]):
+                x_fuses.append(self.attentions[i](xi))
+        elif self.dil is not None:
+            for i, xi in enumerate([x1, x2, x3, x4]):
+                x_fuses.append(self.dilations[i](xi))
+        else:
+            x_fuses = [x1, x2, x3, x4]
+
+        e1 = self.conv_reduces[0](x_fuses[0])
+        e1 = F.interpolate(e1, (h, w), mode='bilinear', align_corners=False)
+
+        e2 = self.conv_reduces[1](x_fuses[1])
+        e2 = F.interpolate(e2, (h, w), mode='bilinear', align_corners=False)
+
+        e3 = self.conv_reduces[2](x_fuses[2])
+        e3 = F.interpolate(e3, (h, w), mode='bilinear', align_corners=False)
+
+        e4 = self.conv_reduces[3](x_fuses[3])
+        e4 = F.interpolate(e4, (h, w), mode='bilinear', align_corners=False)
+
+        outputs = [e1, e2, e3, e4]
+        output = self.classifier(torch.cat(outputs, dim=1))
+
+        outputs.append(output)
+        outputs = [torch.sigmoid(r) for r in outputs]
+        return outputs[-1]
+
+
+@ANNOTATORS.register_class()
+class PiDiAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        pretrained_model = cfg.get('PRETRAINED_MODEL', None)
+        vanilla_cnn = cfg.get('VANILLA_CNN', True)
+        pdcs = config_model_converted(
+            'carv4') if vanilla_cnn else config_model('carv4')
+        self.model = PiDiNet(60, pdcs, dil=24, sa=True,
+                             convert=vanilla_cnn).eval()
+        if pretrained_model:
+            with FS.get_from(pretrained_model, wait_finish=True) as local_path:
+                state = torch.load(local_path,
+                                   map_location='cpu', weights_only=True)['state_dict']
+                if vanilla_cnn:
+                    state = convert_pidinet(state, 'carv4')
+                state = {
+                    k[len('module.'):] if k.startswith('module.') else k: v
+                    for k, v in state.items()
+                }
+                self.model.load_state_dict(state)
+
+    @torch.no_grad()
+    @torch.inference_mode()
+    @torch.autocast('cuda', enabled=False)
+    def forward(self, image, return_grayscale=False):
+        is_batch = False if len(image.shape) == 3 else True
+        if isinstance(image, torch.Tensor):
+            if len(image.shape) == 3:
+                image = rearrange(image, 'h w c -> 1 c h w')
+            elif len(image.shape) == 4:
+                image = rearrange(image, 'b h w c -> b c h w')
+            else:
+                raise "Unsurpport input image's shape"
+        elif isinstance(image, np.ndarray):
+            image = torch.from_numpy(image.copy()).float()
+            if len(image.shape) == 3:
+                image = rearrange(image, 'h w c -> 1 c h w')
+            elif len(image.shape) == 4:
+                image = rearrange(image, 'b h w c -> b c h w')
+            else:
+                raise "Unsurpport input image's shape"
+        else:
+            raise "Unsurpport input image's type"
+        image = image.float().div(255)
+        image = image.to(we.device_id)
+        edge = self.model(image)
+        edge = edge.squeeze(dim=1)
+        edge = 255 - (edge * 255.0).clip(0, 255)  # return white background
+        edge = edge.cpu().numpy()
+        edge = edge.astype(np.uint8)
+        if not is_batch:
+            edge = edge.squeeze()
+        if not return_grayscale:
+            edge = edge[..., None].repeat(3, -1)
+        return edge
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('ANNOTATORS',
+                            __class__.__name__,
+                            PiDiAnnotator.para_dict,
+                            set_name=True)

scepter/modules/annotator/raft.py

@@ -0,0 +1,62 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import torch
+import random
+import numpy as np
+import argparse
+
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import Config
+from scepter.modules.utils.distribute import we
+from scepter.modules.utils.file_system import FS
+
+try:
+    from raft import RAFT
+    from raft.utils.utils import InputPadder
+    from raft.utils import flow_viz
+except:
+    import warnings
+    warnings.warn("ignore raft import, please pip install raft.")
+
+
+@ANNOTATORS.register_class()
+class RAFTAnnotator(BaseAnnotator):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        params = {
+            "small": False,
+            "mixed_precision": False,
+            "alternate_corr": False
+        }
+        params = argparse.Namespace(**params)
+        model = RAFT(params)
+        if cfg.PRETRAINED_MODEL is not None:
+            with FS.get_from(cfg.PRETRAINED_MODEL,
+                             wait_finish=True) as local_path:
+                model.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(local_path, map_location="cpu", weights_only=True).items()})
+        self.model = model.to(we.device_id).eval()
+
+    def forward(self, frames):
+        # frames / RGB
+        frames = [torch.from_numpy(frame.astype(np.uint8)).permute(2, 0, 1).float()[None].to(we.device_id) for frame in frames]
+        flow_up_list, flow_up_vis_list = [], []
+        with torch.no_grad():
+            for i, (image1, image2) in enumerate(zip(frames[:-1], frames[1:])):
+                padder = InputPadder(image1.shape)
+                image1, image2 = padder.pad(image1, image2)
+                flow_low, flow_up = self.model(image1, image2, iters=20, test_mode=True)
+                flow_up = flow_up[0].permute(1, 2, 0).cpu().numpy()
+                flow_up_vis = flow_viz.flow_to_image(flow_up)
+                flow_up_list.append(flow_up)
+                flow_up_vis_list.append(flow_up_vis)
+        return flow_up_list, flow_up_vis_list  # RGB
+
+
+@ANNOTATORS.register_class()
+class RAFTVisAnnotator(RAFTAnnotator):
+    def forward(self, frames):
+        flow_up_list, flow_up_vis_list = super().forward(frames)
+        return flow_up_vis_list[:1] + flow_up_vis_list

scepter/modules/annotator/region_canvas.py

@@ -0,0 +1,95 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import random
+from abc import ABCMeta
+
+import cv2
+import numpy as np
+import torch
+from PIL import Image
+
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import Config, dict_to_yaml
+
+
+@ANNOTATORS.register_class()
+class RegionCanvasAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        self.scale_range = cfg.get('SCALE_RANGE', [0.75, 1.0])
+        self.canvas_value = cfg.get('CANVAS_VALUE', 255)
+        self.use_resize = cfg.get('USE_RESIZE', True)
+        self.use_canvas = cfg.get('USE_CANVAS', True)
+        self.use_aug = cfg.get('USE_AUG', False)
+        if self.use_aug:
+            mask_aug_dict = {'NAME': 'MaskAugAnnotator'}
+            mask_aug_cfg = Config(cfg_dict=mask_aug_dict, load=False)
+            self.mask_aug_anno = ANNOTATORS.build(mask_aug_cfg)
+
+
+    def forward(self,
+                image,
+                mask,
+                mask_cfg=None):
+        if isinstance(image, Image.Image):
+            image = np.array(image)
+        elif isinstance(image, torch.Tensor):
+            image = image.detach().cpu().numpy()
+        elif isinstance(image, np.ndarray):
+            image = image
+        else:
+            raise f'Unsurpport datatype{type(image)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+
+        mask = np.array(mask).astype(np.uint8)
+        image_h, image_w = image.shape[:2]
+
+        if self.use_aug:
+            mask = self.mask_aug_anno(mask, mask_cfg)
+
+        # get region with white bg
+        image[np.array(mask) == 0] = self.canvas_value
+        x, y, w, h = cv2.boundingRect(mask)
+        region_crop = image[y:y + h, x:x + w]
+
+        if self.use_resize:
+            # resize region
+            scale_min, scale_max = self.scale_range
+            scale_factor = random.uniform(scale_min, scale_max)
+            new_w, new_h = int(image_w * scale_factor), int(image_h * scale_factor)
+            obj_scale_factor = min(new_w/w, new_h/h)
+
+            new_w = int(w * obj_scale_factor)
+            new_h = int(h * obj_scale_factor)
+            region_crop_resized = cv2.resize(region_crop, (new_w, new_h), interpolation=cv2.INTER_AREA)
+        else:
+            region_crop_resized = region_crop
+
+        if self.use_canvas:
+            # plot region into canvas
+            new_canvas = np.ones_like(image) * self.canvas_value
+            max_x = max(0, image_w - new_w)
+            max_y = max(0, image_h - new_h)
+            new_x = random.randint(0, max_x)
+            new_y = random.randint(0, max_y)
+
+            new_canvas[new_y:new_y + new_h, new_x:new_x + new_w] = region_crop_resized
+        else:
+            new_canvas = region_crop_resized
+        return new_canvas
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('ANNOTATORS',
+                            __class__.__name__,
+                            RegionCanvasAnnotator.para_dict,
+                            set_name=True)
+
+
+@ANNOTATORS.register_class()
+class RegionCanvasCropAnnotator(RegionCanvasAnnotator):
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        self.use_resize, self.use_canvas = False, False

scepter/modules/annotator/registry.py

@@ -15,7 +15,7 @@ def build_annotator(cfg, registry, logger=None, *args, **kwargs):
         raise TypeError(f'Config must be type dict, got {type(cfg)}')
     if cfg.have('PRETRAINED_MODEL'):
         pretrain_cfg = cfg.PRETRAINED_MODEL
-        if pretrain_cfg is not None and not isinstance(pretrain_cfg, (str)):
+        if pretrain_cfg is not None and not isinstance(pretrain_cfg, (str, list)):
             raise TypeError('Pretrain parameter must be a string')
     else:
         pretrain_cfg = None

scepter/modules/annotator/segmentation.py

@@ -0,0 +1,388 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import random
+from abc import ABCMeta
+
+import cv2
+import numpy as np
+import torch
+import torchvision.transforms as T
+from PIL import Image
+from pycocotools import mask as mask_utils
+from scipy import ndimage
+try:
+    from sklearn.cluster import KMeans
+except:
+    import warnings
+    warnings.warn("ignore sklearn import, please pip install scikit-learn.")
+from torchvision.ops.boxes import batched_nms
+
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import dict_to_yaml
+from scepter.modules.utils.distribute import we
+from scepter.modules.utils.file_system import FS
+
+
+def find_dominant_color(image, k=1):
+    pixels = image.reshape((-1, 3))
+    mask = (pixels != [0, 0, 0]).all(axis=1)
+    pixels = pixels[mask]
+    try:
+        kmeans = KMeans(n_clusters=k, n_init='auto')
+        kmeans.fit(pixels)
+        dominant_color = kmeans.cluster_centers_.astype(int)[0]
+    except Exception:
+        dominant_color = np.array([255, 255, 255])
+    return dominant_color
+
+
+def cv2_resize_crop(image, resize_size, crop_size):
+    resize_height, resize_width = resize_size
+    crop_height, crop_width = crop_size
+
+    resized_image = cv2.resize(image, (resize_width, resize_height))
+
+    center_x, center_y = resize_width // 2, resize_height // 2
+    crop_start_x = max(center_x - crop_width // 2, 0)
+    crop_start_y = max(center_y - crop_height // 2, 0)
+    crop_end_x = crop_start_x + crop_width
+    crop_end_y = crop_start_y + crop_height
+
+    crop_end_x = min(crop_end_x, resize_width)
+    crop_end_y = min(crop_end_y, resize_height)
+
+    center_cropped_image = resized_image[crop_start_y:crop_end_y,
+                                         crop_start_x:crop_end_x]
+
+    return center_cropped_image
+
+
+@ANNOTATORS.register_class()
+class ESAMAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        try:
+            from efficient_sam.efficient_sam import build_efficient_sam
+        except Exception:
+            raise NotImplementedError(
+                'Please install efficient_sam and segment_anything modules.')
+
+        pretrained_model = cfg.get('PRETRAINED_MODEL', None)
+        if pretrained_model:
+            with FS.get_from(pretrained_model, wait_finish=True) as local_path:
+                self.efficient_sam_module = build_efficient_sam(
+                    encoder_patch_embed_dim=384,
+                    encoder_num_heads=6,
+                    checkpoint=local_path).eval().to(we.device_id)
+        self.GRID_SIZE = cfg.get('GRID_SIZE', 16)
+        self.save_mode = cfg.get('SAVE_MODE', 'P')
+        self.use_dominant_color = cfg.get('USE_DOMINANT_COLOR', False)
+        self.return_mask = cfg.get('RETURN_MASK', False)
+
+    @torch.no_grad()
+    def get_predictions_given_embeddings_and_queries(self, img, points,
+                                                     point_labels, model):
+        from segment_anything.utils.amg import calculate_stability_score
+        predicted_masks, predicted_iou = [], []
+        bs = 128
+        num = int(float(self.GRID_SIZE * self.GRID_SIZE) /
+                  bs) if self.GRID_SIZE * self.GRID_SIZE % bs == 0 else int(
+                      float(self.GRID_SIZE * self.GRID_SIZE) / bs) + 1
+        for i in range(num):
+            predicted_mask_item, predicted_iou_item = model(
+                img[None, ...], points[:, i * bs:(i + 1) * bs, ...],
+                point_labels[:, i * bs:(i + 1) * bs, :])
+            predicted_masks.append(predicted_mask_item)
+            predicted_iou.append(predicted_iou_item)
+            torch.cuda.empty_cache()
+        # # predicted:  torch.Size([1, 1024, 3]) torch.Size([1, 1024, 3, 512, 512])
+        # print('predicted: ', predicted_iou.size(), predicted_masks.size())
+        predicted_masks = torch.cat(predicted_masks, dim=1)
+        predicted_iou = torch.cat(predicted_iou, dim=1)
+        sorted_ids = torch.argsort(predicted_iou, dim=-1, descending=True)
+        predicted_iou_scores = torch.take_along_dim(predicted_iou,
+                                                    sorted_ids,
+                                                    dim=2)
+        predicted_masks = torch.take_along_dim(predicted_masks,
+                                               sorted_ids[..., None, None],
+                                               dim=2)
+        predicted_masks = predicted_masks[0]
+        iou = predicted_iou_scores[0, :, 0]
+        index_iou = iou > 0.7
+        iou_ = iou[index_iou]
+        masks = predicted_masks[index_iou]
+        score = calculate_stability_score(masks, 0.0, 1.0)
+        score = score[:, 0]
+        index = score > 0.9
+        masks = masks[index]
+        iou_ = iou_[index]
+        masks = torch.ge(masks, 0.0)
+        return masks, iou_
+
+    def singel_mask_to_rle(self, mask):
+        rle = mask_utils.encode(
+            np.array(mask[:, :, None], order='F', dtype='uint8'))[0]
+        rle['counts'] = rle['counts'].decode('utf-8')
+        return rle
+
+    def process_small_region(self, rles):
+        from segment_anything.utils.amg import rle_to_mask, remove_small_regions, \
+            batched_mask_to_box, mask_to_rle_pytorch
+        new_masks = []
+        scores = []
+        min_area = 100
+        nms_thresh = 0.7
+        for rle in rles:
+            mask = rle_to_mask(rle[0])
+
+            mask, changed = remove_small_regions(mask, min_area, mode='holes')
+            unchanged = not changed
+            mask, changed = remove_small_regions(mask,
+                                                 min_area,
+                                                 mode='islands')
+            unchanged = unchanged and not changed
+
+            new_masks.append(torch.as_tensor(mask).unsqueeze(0))
+            # Give score=0 to changed masks and score=1 to unchanged masks
+            # so NMS will prefer ones that didn't need postprocessing
+            scores.append(float(unchanged))
+
+        # Recalculate boxes and remove any new duplicates
+        masks = torch.cat(new_masks, dim=0).to(we.device_id)
+        boxes = batched_mask_to_box(masks)
+        keep_by_nms = batched_nms(
+            boxes.float(),
+            torch.as_tensor(scores).to(we.device_id),
+            torch.zeros_like(boxes[:, 0]),  # categories
+            iou_threshold=nms_thresh,
+        )
+
+        # Only recalculate RLEs for masks that have changed
+        for i_mask in keep_by_nms:
+            if scores[i_mask] == 0.0:
+                mask_torch = masks[i_mask].unsqueeze(0)
+                rles[i_mask] = mask_to_rle_pytorch(mask_torch)
+        masks = [rle_to_mask(rles[i][0]) for i in keep_by_nms]
+        return masks
+
+    def run_everything_ours(self, img_tensor, model):
+        from segment_anything.utils.amg import mask_to_rle_pytorch
+        img_tensor = img_tensor.squeeze(0)
+        _, original_image_h, original_image_w = img_tensor.shape
+        xy = []
+        for i in range(self.GRID_SIZE):
+            curr_x = 0.5 + i / self.GRID_SIZE * original_image_w
+            for j in range(self.GRID_SIZE):
+                curr_y = 0.5 + j / self.GRID_SIZE * original_image_h
+                xy.append([curr_x, curr_y])
+
+        xy = torch.from_numpy(np.array(xy))
+        points = xy
+        num_pts = xy.shape[0]
+        point_labels = torch.ones(num_pts, 1)
+        with torch.no_grad():
+            predicted_masks, predicted_iou = self.get_predictions_given_embeddings_and_queries(
+                img_tensor,
+                points.reshape(1, num_pts, 1, 2).to(we.device_id),
+                point_labels.reshape(1, num_pts, 1).to(we.device_id),
+                model,
+            )
+        # print('predicted_masks: ', predicted_masks[0][0:1].dtype, predicted_masks[0][0:1].device)
+        rle = [mask_to_rle_pytorch(m[0:1]) for m in predicted_masks]
+        # transform to numpy
+        size, counts = [], []
+        for rle_item in rle:
+            size.append(rle_item[0]['size'])
+            counts += rle_item[0]['counts']
+            counts += '#'
+        predicted_masks = self.process_small_region(rle)
+        return predicted_masks
+
+    def forward(self, image, return_mask=None):
+        return_mask = return_mask if return_mask is not None else self.return_mask
+        if isinstance(image, Image.Image):
+            image = np.array(image)
+        elif isinstance(image, torch.Tensor):
+            image = image.detach().cpu().numpy()
+        elif isinstance(image, np.ndarray):
+            image = image.copy()
+        else:
+            raise f'Unsurpport datatype{type(image)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+        h, w = image.shape[:2]
+        max_rate = max(float(w) / 1024.0, float(h) / 1024.0)
+        w_ori = int(float(w) / max_rate)
+        h_ori = int(float(h) / max_rate)
+        # image = T.ToTensor()(T.Resize((h_ori, w_ori))(Image.fromarray(image)))
+        # image_pad = T.Pad((0, 0, 1024 - w_ori, 1024 - h_ori))(image)
+        image_pad = T.Pad((0, 0, 1024 - w_ori, 1024 - h_ori))(T.Resize(
+            (h_ori, w_ori))(Image.fromarray(image)))
+        input_image = T.ToTensor()(image_pad)
+        input_image = input_image.unsqueeze(0).to(we.device_id)
+
+        mask_efficient_sam_vits = self.run_everything_ours(
+            input_image, self.efficient_sam_module)
+        annos = []
+        mask_efficient_sam_vits = sorted(list(mask_efficient_sam_vits),
+                                         key=lambda m: int(m.sum()),
+                                         reverse=True)
+        mask_efficient_sam_vits = mask_efficient_sam_vits[:256]
+        for mask in mask_efficient_sam_vits:
+            mask_item = mask_utils.encode(
+                np.array(mask[:, :, None], order='F', dtype='uint8'))[0]
+            mask_item['counts'] = mask_item['counts'].decode('utf-8')
+            mask_area = int(mask.sum())
+            annos.append({'mask': mask_item, 'mask_area': mask_area})
+
+        annos = sorted(annos, key=lambda x: x['mask_area'], reverse=True)
+        seg_img = None
+        dominant_palette = []
+        image_pad_np = np.array(image_pad)
+        for idx, anno in enumerate(annos):
+            color = idx
+            if idx > 255:
+                break
+            mask = np.array(mask_utils.decode(anno['mask'])).astype(np.uint8)
+            h, w = mask.shape[:2]
+            if seg_img is None:
+                seg_img = np.ones((h, w, 3)) * 255
+            if self.use_dominant_color:
+                masked_image = cv2.bitwise_and(image_pad_np,
+                                               image_pad_np,
+                                               mask=mask)
+                dominant_color = find_dominant_color(masked_image).tolist()
+                dominant_palette.append(dominant_color)
+            seg_img[mask.astype(bool)] = [color, color, color]
+        seg_img = Image.fromarray(seg_img.astype(np.uint8)).convert('L')
+
+        resize_rate = max(float(h_ori) / 1024.0, float(w_ori) / 1024.0)
+        h_new = int(float(h_ori) / resize_rate)
+        w_new = int(float(w_ori) / resize_rate)
+        seg_img = seg_img.crop((0, 0, w_new, h_new))
+        if self.save_mode == 'P':
+            palette = []
+            for i in range(256):
+                if not self.use_dominant_color:
+                    palette_item = [random.randint(0, 255) for _ in range(3)]
+                else:
+                    palette_item = dominant_palette[i] if i < len(
+                        dominant_palette) else [255, 255, 255]
+                palette += palette_item
+            seg_img = seg_img.convert('P')
+            seg_img.putpalette(palette)
+            seg_rgb_img = seg_img.convert('RGB')
+            if return_mask:
+                return {
+                    'image': np.array(seg_rgb_img),
+                    'mask': np.array(seg_img)
+                }
+            else:
+                return np.array(seg_rgb_img)
+        else:
+            return np.array(seg_img)
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('ANNOTATORS',
+                            __class__.__name__,
+                            ESAMAnnotator.para_dict,
+                            set_name=True)
+
+
+@ANNOTATORS.register_class()
+class SAMAnnotatorDraw(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        from segment_anything import sam_model_registry, SamPredictor
+        from segment_anything.utils.transforms import ResizeLongestSide
+
+        self.transform = ResizeLongestSide(1024)
+        self.task_type = cfg.get('TASK_TYPE', 'input_box')
+        self.sam_model = cfg.get('SAM_MODEL', 'vit_b')
+        pretrained_model = cfg.get('PRETRAINED_MODEL', 'sam_vit_b_01ec64.pth')
+
+        if pretrained_model:
+            with FS.get_from(pretrained_model, wait_finish=True) as local_path:
+                seg_model = sam_model_registry[self.sam_model](
+                    checkpoint=local_path).eval().to(we.device_id)
+                self.sam_predictor = SamPredictor(seg_model)
+
+    def forward(self,
+                image,
+                input_box=None,
+                mask=None,
+                task_type=None,
+                multimask_output=False):
+        task_type = task_type if task_type is not None else self.task_type
+
+        if isinstance(image, Image.Image):
+            image = np.array(image)
+        elif isinstance(image, torch.Tensor):
+            image = image.detach().cpu().numpy()
+        elif isinstance(image, np.ndarray):
+            image = image.copy()
+        else:
+            raise f'Unsurpport datatype{type(image)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+
+        if mask is not None:
+            if isinstance(mask, Image.Image):
+                mask = np.array(mask)
+            elif isinstance(mask, torch.Tensor):
+                mask = mask.detach().cpu().numpy()
+            elif isinstance(mask, np.ndarray):
+                mask = mask.copy()
+            else:
+                raise f'Unsurpport datatype{type(mask)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+
+        if task_type == 'mask_point':
+            scribble = mask.transpose(2, 1, 0)[0]
+            labeled_array, num_features = ndimage.label(scribble >= 255)
+            centers = ndimage.center_of_mass(scribble, labeled_array,
+                                             range(1, num_features + 1))
+            point_coords = np.array(centers)
+            point_labels = np.array([1] * len(centers))
+            sample = {
+                'point_coords': point_coords,
+                'point_labels': point_labels
+            }
+
+        elif task_type == 'mask_box':
+            scribble = mask.transpose(2, 1, 0)[0]
+            labeled_array, num_features = ndimage.label(scribble >= 255)
+            centers = ndimage.center_of_mass(scribble, labeled_array,
+                                             range(1, num_features + 1))
+            centers = np.array(centers)
+            # (x1, y1, x2, y2)
+            x_min = centers[:, 0].min()
+            x_max = centers[:, 0].max()
+            y_min = centers[:, 1].min()
+            y_max = centers[:, 1].max()
+            bbox = np.array([x_min, y_min, x_max, y_max])
+            sample = {'box': bbox}
+
+        elif task_type == 'input_box':
+            if isinstance(input_box, list):
+                input_box = np.array(input_box)
+            sample = {'box': input_box}
+
+        self.sam_predictor.set_image(image)
+        masks, scores, logits = self.sam_predictor.predict(
+            **sample, multimask_output=True)
+        index = np.argmax(scores)
+
+        ret_data = {
+            'mask': (masks[index] * 255).astype(np.uint8),
+            'score': scores[index]
+        }
+        return ret_data
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('ANNOTATORS',
+                            __class__.__name__,
+                            SAMAnnotatorDraw.para_dict,
+                            set_name=True)

scepter/modules/annotator/sketch.py

@@ -0,0 +1,157 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+from abc import ABCMeta
+
+import numpy as np
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import dict_to_yaml
+from scepter.modules.utils.distribute import we
+from scepter.modules.utils.file_system import FS
+
+
+class SketchNet(nn.Module):
+    def __init__(self, mean, std):
+        assert isinstance(mean, float) and isinstance(std, float)
+        super().__init__()
+        self.mean = mean
+        self.std = std
+
+        # layers
+        self.layers = nn.Sequential(nn.Conv2d(1, 48, 5, 2, 2),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(48, 128, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(128, 128, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(128, 128, 3, 2, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(128, 256, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(256, 256, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(256, 256, 3, 2, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(256, 512, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(512, 1024, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(1024, 1024, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(1024, 1024, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(1024, 1024, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(1024, 512, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(512, 256, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.ConvTranspose2d(256, 256, 4, 2, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(256, 256, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(256, 128, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.ConvTranspose2d(128, 128, 4, 2, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(128, 128, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(128, 48, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.ConvTranspose2d(48, 48, 4, 2, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(48, 24, 3, 1, 1),
+                                    nn.ReLU(inplace=True),
+                                    nn.Conv2d(24, 1, 3, 1, 1), nn.Sigmoid())
+
+    def forward(self, x):
+        """x: [B, 1, H, W] within range [0, 1]. Sketch pixels in dark color.
+        """
+        x = (x - self.mean) / self.std
+        return self.layers(x)
+
+
+@ANNOTATORS.register_class()
+class SketchAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        pretrained_model = cfg.get('PRETRAINED_MODEL', None)
+        self.model = SketchNet(mean=0.9664114577640158,
+                               std=0.0858381272736797).eval()
+        if pretrained_model:
+            with FS.get_from(pretrained_model, wait_finish=True) as local_path:
+                state = torch.load(local_path, map_location='cpu', weights_only=True)
+                self.model.load_state_dict(state)
+
+    @torch.no_grad()
+    @torch.inference_mode()
+    @torch.autocast('cuda', enabled=False)
+    def forward(self, image):
+        is_batch = False if len(image.shape) == 3 else True
+        if isinstance(image, torch.Tensor):
+            if len(image.shape) == 3:
+                if torch.equal(image[:, :, 0], image[:, :, 1]) and torch.equal(
+                        image[:, :, 1], image[:, :, 2]):
+                    image = image[:, :, 0]
+                else:
+                    raise "Unsurpport input image's shape and each channel is different"
+            elif len(image.shape) == 4:
+                if (torch.equal(image[:, :, :, 0], image[:, :, :, 1])
+                        and torch.equal(image[:, :, :, 1], image[:, :, :, 2])):
+                    image = image[:, :, :, 0]
+                else:
+                    raise "Unsurpport input image's shape and each channel is different"
+            if len(image.shape) == 2:
+                image = rearrange(image, 'h w -> 1 h w')
+                B, H, W = image.shape
+            elif len(image.shape) == 3:
+                B, H, W = image.shape
+            else:
+                raise "Unsurpport input image's shape"
+        elif isinstance(image, np.ndarray):
+            image = image.copy()
+            if len(image.shape) == 3:
+                if np.array_equal(image[:, :, 0],
+                                  image[:, :, 1]) and np.array_equal(
+                                      image[:, :, 1], image[:, :, 2]):
+                    image = image[:, :, 0]
+                else:
+                    raise "Unsurpport input image's shape and each channel is different"
+            elif len(image.shape) == 4:
+                if (np.array_equal(image[:, :, :, 0], image[:, :, :, 1]) and
+                        np.array_equal(image[:, :, :, 1], image[:, :, :, 2])):
+                    image = image[:, :, :, 0]
+                else:
+                    raise "Unsurpport input image's shape and each channel is different"
+            image = torch.from_numpy(image).float()
+            if len(image.shape) == 2:
+                image = rearrange(image, 'h w -> 1 1 h w')
+            elif len(image.shape) == 3:
+                image = rearrange(image, 'b h w -> b 1 h w')
+            else:
+                raise "Unsurpport input image's shape"
+        else:
+            raise "Unsurpport input image's type"
+        image = image.float().div(255)
+        image = image.to(we.device_id)
+        edge = self.model(image)
+        edge = edge.squeeze(dim=1)
+        edge = (edge * 255.0).clip(0, 255)
+        edge = edge.cpu().numpy()
+        edge = edge.astype(np.uint8)
+        if not is_batch:
+            edge = edge.squeeze()
+        return edge[..., None].repeat(3, -1)
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('ANNOTATORS',
+                            __class__.__name__,
+                            SketchAnnotator.para_dict,
+                            set_name=True)

scepter/modules/annotator/video_segmentation.py

@@ -0,0 +1,153 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+from abc import ABCMeta
+
+import numpy as np
+import torch
+from PIL import Image
+from scipy import ndimage
+try:
+    from sklearn.cluster import KMeans
+except:
+    import warnings
+    warnings.warn("ignore sklearn import, please pip install scikit-learn.")
+
+from scepter.modules.annotator.base_annotator import BaseAnnotator
+from scepter.modules.annotator.registry import ANNOTATORS
+from scepter.modules.utils.config import dict_to_yaml
+from scepter.modules.utils.file_system import FS
+import pycocotools.mask as mask_utils
+
+
+def single_mask_to_rle(mask):
+    rle = mask_utils.encode(np.array(mask[:, :, None], order="F", dtype="uint8"))[0]
+    rle["counts"] = rle["counts"].decode("utf-8")
+    return rle
+
+def single_rle_to_mask(rle):
+    mask = np.array(mask_utils.decode(rle)).astype(np.uint8)
+    return mask
+
+def single_mask_to_xyxy(mask):
+    bbox = np.zeros((4), dtype=int)
+    rows, cols = np.where(np.array(mask))
+    if len(rows) > 0 and len(cols) > 0:
+        x_min, x_max = np.min(cols), np.max(cols)
+        y_min, y_max = np.min(rows), np.max(rows)
+        bbox[:] = [x_min, y_min, x_max, y_max]
+    return bbox.tolist()
+
+@ANNOTATORS.register_class()
+class SAM2DrawVideoAnnotator(BaseAnnotator, metaclass=ABCMeta):
+    para_dict = {}
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        self.task_type = cfg.get('TASK_TYPE', 'input_box')
+        from sam2.build_sam import build_sam2_video_predictor
+        config_path = FS.get_from(cfg.CONFIG_PATH, local_path=cfg.CONFIG_LOCAL_PATH, wait_finish=True)
+        pretrained_model = FS.get_from(cfg.PRETRAINED_MODEL, wait_finish=True)
+        self.video_predictor = build_sam2_video_predictor(config_path, pretrained_model, fill_hole_area=0)
+
+    def forward(self,
+                video,
+                input_box=None,
+                mask=None,
+                task_type=None):
+        task_type = task_type if task_type is not None else self.task_type
+
+        if mask is not None:
+            if isinstance(mask, Image.Image):
+                mask = np.array(mask)
+            elif isinstance(mask, torch.Tensor):
+                mask = mask.detach().cpu().numpy()
+            elif isinstance(mask, np.ndarray):
+                mask = mask.copy()
+            else:
+                raise f'Unsurpport datatype{type(mask)}, only surpport np.ndarray, torch.Tensor, Pillow Image.'
+
+        if task_type == 'mask_point':
+            if len(mask.shape) == 3:
+                scribble = mask.transpose(2, 1, 0)[0]
+            else:
+                scribble = mask.transpose(1, 0)   # (H, W) -> (W, H)
+            labeled_array, num_features = ndimage.label(scribble >= 255)
+            centers = ndimage.center_of_mass(scribble, labeled_array,
+                                             range(1, num_features + 1))
+            point_coords = np.array(centers)
+            point_labels = np.array([1] * len(centers))
+            sample = {
+                'points': point_coords,
+                'labels': point_labels
+            }
+        elif task_type == 'mask_box':
+            if len(mask.shape) == 3:
+                scribble = mask.transpose(2, 1, 0)[0]
+            else:
+                scribble = mask.transpose(1, 0)  # (H, W) -> (W, H)
+            labeled_array, num_features = ndimage.label(scribble >= 255)
+            centers = ndimage.center_of_mass(scribble, labeled_array,
+                                             range(1, num_features + 1))
+            centers = np.array(centers)
+            # (x1, y1, x2, y2)
+            x_min = centers[:, 0].min()
+            x_max = centers[:, 0].max()
+            y_min = centers[:, 1].min()
+            y_max = centers[:, 1].max()
+            bbox = np.array([x_min, y_min, x_max, y_max])
+            sample = {'box': bbox}
+        elif task_type == 'input_box':
+            if isinstance(input_box, list):
+                input_box = np.array(input_box)
+            sample = {'box': input_box}
+        elif task_type == 'mask':
+            sample = {'mask': mask}
+        else:
+            raise NotImplementedError
+
+        ann_frame_idx = 0
+        object_id = 0
+        with (torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16)):
+
+            inference_state = self.video_predictor.init_state(video_path=video)
+
+            if task_type in ['mask_point', 'mask_box', 'input_box']:
+                _, out_obj_ids, out_mask_logits = self.video_predictor.add_new_points_or_box(
+                    inference_state=inference_state,
+                    frame_idx=ann_frame_idx,
+                    obj_id=object_id,
+                    **sample
+                )
+            elif task_type in ['mask']:
+                _, out_obj_ids, out_mask_logits = self.video_predictor.add_new_mask(
+                    inference_state=inference_state,
+                    frame_idx=ann_frame_idx,
+                    obj_id=object_id,
+                    **sample
+                )
+            else:
+                raise NotImplementedError
+
+            video_segments = {}  # video_segments contains the per-frame segmentation results
+            for out_frame_idx, out_obj_ids, out_mask_logits in self.video_predictor.propagate_in_video(inference_state):
+                frame_segments = {}
+                for i, out_obj_id in enumerate(out_obj_ids):
+                    mask = (out_mask_logits[i] > 0.0).cpu().numpy().squeeze(0)
+                    frame_segments[out_obj_id] = {
+                        "mask": single_mask_to_rle(mask),
+                        "mask_area": int(mask.sum()),
+                        "mask_box": single_mask_to_xyxy(mask),
+                    }
+                video_segments[out_frame_idx] = frame_segments
+
+        ret_data = {
+            "annotations": video_segments
+        }
+        return ret_data
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('ANNOTATORS',
+                            __class__.__name__,
+                            SAM2DrawVideoAnnotator.para_dict,
+                            set_name=True)

scepter/modules/data/__init__.py

@@ -1,4 +1,21 @@
 # -*- coding: utf-8 -*-
 # Copyright (c) Alibaba, Inc. and its affiliates.
+from typing import TYPE_CHECKING
+from scepter.modules.utils.import_utils import LazyImportModule
 
-from scepter.modules.data import dataset, sampler
+
+if TYPE_CHECKING:
+    from scepter.modules.data import dataset, sampler
+else:
+    _import_structure = {
+        'data': ['dataset', 'sampler']
+    }
+
+    import sys
+    sys.modules[__name__] = LazyImportModule(
+        __name__,
+        globals()['__file__'],
+        _import_structure,
+        module_spec=__spec__,
+        extra_objects={},
+    )

scepter/modules/data/dataset/__init__.py

@@ -1,11 +1,35 @@
 # -*- coding: utf-8 -*-
 # Copyright (c) Alibaba, Inc. and its affiliates.
+from typing import TYPE_CHECKING
+from scepter.modules.utils.import_utils import LazyImportModule
 
-from scepter.modules.data.dataset.base_dataset import BaseDataset
-from scepter.modules.data.dataset.dataset import (Image2ImageDataset,
-                                                  ImageClassifyPublicDataset,
-                                                  ImageTextPairDataset,
-                                                  Text2ImageDataset)
-from scepter.modules.data.dataset.ms_dataset import (
-    ImageTextPairFolderDataset, ImageTextPairMSDataset)
-from scepter.modules.data.dataset.registry import DATASETS
+
+if TYPE_CHECKING:
+    from scepter.modules.data.dataset.base_dataset import BaseDataset
+    from scepter.modules.data.dataset.dataset import (Image2ImageDataset,
+                                                      ImageClassifyPublicDataset,
+                                                      ImageTextPairDataset,
+                                                      Text2ImageDataset)
+    from scepter.modules.data.dataset.ms_dataset import (
+        ImageTextPairFolderDataset, ImageTextPairMSDataset)
+    from scepter.modules.data.dataset.registry import DATASETS
+    from scepter.modules.data.dataset.video_gen_dataset import VideoGenDataset
+else:
+    _import_structure = {
+        'base_dataset': ['BaseDataset'],
+        'dataset': ['Image2ImageDataset', 'ImageClassifyPublicDataset',
+                    'ImageTextPairDataset', 'Text2ImageDataset'],
+        'ms_dataset': ['ImageTextPairFolderDataset',
+                       'ImageTextPairMSDataset'],
+        'registry': ['DATASETS'],
+        'video_gen_dataset': ['VideoGenDataset']
+    }
+
+    import sys
+    sys.modules[__name__] = LazyImportModule(
+        __name__,
+        globals()['__file__'],
+        _import_structure,
+        module_spec=__spec__,
+        extra_objects={},
+    )

scepter/modules/data/dataset/base_dataset.py

@@ -82,6 +82,8 @@ class BaseDataset(Dataset, metaclass=ABCMeta):
                                                overwrite=False)
         self.worker_id = worker_id
         self.logger = self.worker_logger
+        self.local_we["seed"] += (worker_id + self.local_we['rank'] * 1234)
+        self.seed = self.local_we["seed"]
         we.set_env(self.local_we)
 
     @abstractmethod

scepter/modules/data/dataset/dataset.py

@@ -4,6 +4,7 @@
 import numbers
 import os
 import sys
+import copy
 from collections.abc import Iterable
 
 import numpy as np
@@ -242,6 +243,8 @@ class Text2ImageDataset(BaseDataset):
         prompt_prefix = cfg.get('PROMPT_PREFIX', '')
         path_prefix = cfg.get('PATH_PREFIX', '')
         use_num = cfg.get('USE_NUM', -1)
+        meta_cfg = cfg.get('META_CFG', None)
+        meta_cfg = meta_cfg.get_lowercase_dict() if meta_cfg is not None else None
 
         image_size = cfg.get('IMAGE_SIZE', 1024)
         if isinstance(image_size, numbers.Number):
@@ -264,7 +267,12 @@ class Text2ImageDataset(BaseDataset):
 
         self.items = list()
         for i, row in enumerate(rows):
-            item = {'index': i, 'meta': {'image_size': image_size}}
+            if meta_cfg is not None:
+                meta_cfg_copy = copy.deepcopy(meta_cfg)
+                meta_cfg_copy['image_size'] = image_size
+                item = {'index': i, 'meta': meta_cfg_copy}
+            else:
+                item = {'index': i, 'meta': {'image_size': image_size}}
             for key, value in zip(fields, row):
                 if key in ['prompt', 'caption', 'text']:
                     item['ori_prompt'] = value

scepter/modules/data/dataset/ms_dataset.py

@@ -1,16 +1,28 @@
 # -*- coding: utf-8 -*-
 # Copyright (c) Alibaba, Inc. and its affiliates.
 
+import io
+import math
 import numbers
 import os
 import sys
+from collections import defaultdict
+
+import numpy as np
+import torch
+import torchvision.transforms as T
+from PIL import Image
+from torchvision.transforms.functional import InterpolationMode
 
 from scepter.modules.data.dataset.base_dataset import BaseDataset
 from scepter.modules.data.dataset.registry import DATASETS
+from scepter.modules.transform.io import pillow_convert
 from scepter.modules.utils.config import dict_to_yaml
 from scepter.modules.utils.distribute import we
 from scepter.modules.utils.file_system import FS
 
+Image.MAX_IMAGE_PIXELS = None
+
 
 @DATASETS.register_class()
 class ImageTextPairMSDataset(BaseDataset):
@@ -102,7 +114,7 @@ class ImageTextPairMSDataset(BaseDataset):
                 self.output_size = [self.output_size, self.output_size]
         # Use modelscope dataset
         if not ms_dataset_name:
-            raise (
+            raise ValueError(
                 'Your must set MS_DATASET_NAME as modelscope dataset or your local dataset orignized '
                 'as modelscope dataset.')
         if FS.exists(ms_dataset_name):
@@ -125,7 +137,7 @@ class ImageTextPairMSDataset(BaseDataset):
                     split=ms_dataset_split,
                     download_mode=DownloadMode.FORCE_REDOWNLOAD)
             except Exception as sec_e:
-                raise f'Load Modelscope dataset failed {sec_e}.'
+                raise ValueError(f'Load Modelscope dataset failed {sec_e}.')
         if ms_remap_keys:
             self.data = self.data.remap_columns(ms_remap_keys.get_dict())
 
@@ -245,7 +257,7 @@ class ImageTextPairFolderDataset(BaseDataset):
                 self.output_size = [self.output_size, self.output_size]
         # Use modelscope dataset
         if not data_folder or not FS.exists(data_folder):
-            raise ('Your must set datafolder for local dataset.')
+            raise ValueError('Your must set datafolder for local dataset.')
         data_folder = FS.get_dir_to_local_dir(data_folder)
         all_lines = open(os.path.join(data_folder, 'train.csv'),
                          'r').read().split('\n')
@@ -311,3 +323,254 @@ class ImageTextPairFolderDataset(BaseDataset):
                             __class__.__name__,
                             ImageTextPairMSDataset.para_dict,
                             set_name=True)
+
+
+@DATASETS.register_class()
+class ImageTextPairMSDatasetForACE(BaseDataset):
+    para_dict = {
+        'MS_DATASET_NAME': {
+            'value': '',
+            'description': 'Modelscope dataset name.'
+        },
+        'MS_DATASET_NAMESPACE': {
+            'value': '',
+            'description': 'Modelscope dataset namespace.'
+        },
+        'MS_DATASET_SUBNAME': {
+            'value': '',
+            'description': 'Modelscope dataset subname.'
+        },
+        'MS_DATASET_SPLIT': {
+            'value': '',
+            'description':
+            'Modelscope dataset split set name, default is train.'
+        },
+        'MS_REMAP_KEYS': {
+            'value':
+            None,
+            'description':
+            'Modelscope dataset header of list file, the default is Target:FILE; '
+            'If your file is not this header, please set this field, which is a map dict.'
+            "For example, { 'Image:FILE': 'Target:FILE' } will replace the filed Image:FILE to Target:FILE"
+        },
+        'MS_REMAP_PATH': {
+            'value':
+            None,
+            'description':
+            'When modelscope dataset name is not None, that means you use the dataset from modelscope,'
+            ' default is None. But if you want to use the datalist from modelscope and the file from '
+            'local device, you can use this field to set the root path of your images. '
+        },
+        'TRIGGER_WORDS': {
+            'value':
+            '',
+            'description':
+            'The words used to describe the common features of your data, especially when you customize a '
+            'tuner. Use these words you can get what you want.'
+        },
+        'REPLACE_STYLE': {
+            'value':
+            False,
+            'description':
+            'Whether use the MS_DATASET_SUBNAME to replace the word in your description, default is False.'
+        },
+        'HIGHLIGHT_KEYWORDS': {
+            'value':
+            '',
+            'description':
+            'The keywords you want to highlight in prompt, which will be replace by <HIGHLIGHT_KEYWORDS>.'
+        },
+        'KEYWORDS_SIGN': {
+            'value':
+            '',
+            'description':
+            'The keywords sign you want to add, which is like <{HIGHLIGHT_KEYWORDS}{KEYWORDS_SIGN}>'
+        },
+        'ALIGN_SIZE': {
+            'value': False,
+            'description':
+                'Whether ensure the size align between the source image and target image.'
+        },
+        'OUTPUT_SIZE': {
+            'value':
+            None,
+            'description':
+            'If you use the FlexibleResize transforms, this filed will output the image_size as [h, w],'
+            'which will be used to set the output size of images used to train the model.'
+        },
+    }
+
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg=cfg, logger=logger)
+        from modelscope import MsDataset
+        from modelscope.utils.constant import DownloadMode
+        ms_dataset_name = cfg.get('MS_DATASET_NAME', None)
+        ms_dataset_namespace = cfg.get('MS_DATASET_NAMESPACE', None)
+        ms_dataset_subname = cfg.get('MS_DATASET_SUBNAME', None)
+        ms_dataset_split = cfg.get('MS_DATASET_SPLIT', 'train')
+        ms_remap_keys = cfg.get('MS_REMAP_KEYS', None)
+        ms_remap_path = cfg.get('MS_REMAP_PATH', None)
+
+        self.max_seq_len = cfg.get('MAX_SEQ_LEN', 1024)
+        self.max_aspect_ratio = cfg.get('MAX_ASPECT_RATIO', 4)
+        self.d = cfg.get('DOWNSAMPLE_RATIO', 16)
+        self.replace_style = cfg.get('REPLACE_STYLE', False)
+        self.trigger_words = cfg.get('TRIGGER_WORDS', '')
+        self.replace_keywords = cfg.get('HIGHLIGHT_KEYWORDS', '')
+        self.keywords_sign = cfg.get('KEYWORDS_SIGN', '')
+        self.add_indicator = cfg.get('ADD_INDICATOR', False)
+
+        self.align_size = cfg.get('ALIGN_SIZE', False)
+        # Use modelscope dataset
+        if not ms_dataset_name:
+            raise ValueError(
+                'Your must set MS_DATASET_NAME as modelscope dataset or your local dataset orignized '
+                'as modelscope dataset.')
+        if FS.exists(ms_dataset_name):
+            ms_dataset_name = FS.get_dir_to_local_dir(ms_dataset_name)
+            self.ms_dataset_name = ms_dataset_name
+            # ms_remap_path = ms_dataset_name
+        try:
+            self.data = MsDataset.load(str(ms_dataset_name),
+                                       namespace=ms_dataset_namespace,
+                                       subset_name=ms_dataset_subname,
+                                       split=ms_dataset_split)
+        except Exception:
+            self.logger.info(
+                "Load Modelscope dataset failed, retry with download_mode='force_redownload'."
+            )
+            try:
+                self.data = MsDataset.load(
+                    str(ms_dataset_name),
+                    namespace=ms_dataset_namespace,
+                    subset_name=ms_dataset_subname,
+                    split=ms_dataset_split,
+                    download_mode=DownloadMode.FORCE_REDOWNLOAD)
+            except Exception as sec_e:
+                raise ValueError(f'Load Modelscope dataset failed {sec_e}.')
+        if ms_remap_keys:
+            self.data = self.data.remap_columns(ms_remap_keys.get_dict())
+
+        if ms_remap_path:
+
+            def map_func(example):
+                return {
+                    k: os.path.join(ms_remap_path, v)
+                    if k.endswith(':FILE') else v
+                    for k, v in example.items()
+                }
+
+            self.data = self.data.ds_instance.map(map_func)
+
+        self.transforms = T.Compose([
+            T.ToTensor(),
+            T.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+        ])
+
+    def __len__(self):
+        if self.mode == 'train':
+            return sys.maxsize
+        else:
+            return len(self.data)
+
+    def _get(self, index: int):
+        current_data = self.data[index % len(self.data)]
+
+        tar_image_path = current_data.get('Target:FILE', '')
+        src_image_path = current_data.get('Source:FILE', '')
+
+        style = current_data.get('Style', '')
+        prompt = current_data.get('Prompt', current_data.get('prompt', ''))
+        if self.replace_style and not style == '':
+            prompt = prompt.replace(style, f'<{self.keywords_sign}>')
+
+        elif not self.replace_keywords.strip() == '':
+            prompt = prompt.replace(
+                self.replace_keywords,
+                '<' + self.replace_keywords + f'{self.keywords_sign}>')
+
+        if not self.trigger_words == '':
+            prompt = self.trigger_words.strip() + ' ' + prompt
+
+        src_image = self.load_image(self.ms_dataset_name,
+                                    src_image_path,
+                                    cvt_type='RGB')
+        tar_image = self.load_image(self.ms_dataset_name,
+                                    tar_image_path,
+                                    cvt_type='RGB')
+        src_image = self.image_preprocess(src_image)
+        tar_image = self.image_preprocess(tar_image, size = src_image.shape[:2]  if self.align_size else None)
+
+        tar_image = self.transforms(tar_image)
+        src_image = self.transforms(src_image)
+        src_mask = torch.ones_like(src_image[[0]])
+        tar_mask = torch.ones_like(tar_image[[0]])
+        if self.add_indicator:
+            if '{image}' not in prompt:
+                prompt = '{image}, ' + prompt
+        return {
+            'src_image_list': [src_image],
+            'src_mask_list': [src_mask],
+            'image': tar_image,
+            'image_mask': tar_mask,
+            'prompt': [prompt],
+            'edit_id': [0]
+        }
+
+    def load_image(self, prefix, img_path, cvt_type=None):
+        if img_path is None or img_path == '':
+            return None
+        img_path = os.path.join(prefix, img_path)
+        with FS.get_object(img_path) as image_bytes:
+            image = Image.open(io.BytesIO(image_bytes))
+            if cvt_type is not None:
+                image = pillow_convert(image, cvt_type)
+        return image
+
+    def image_preprocess(self,
+                         img,
+                         size=None,
+                         interpolation=InterpolationMode.BILINEAR):
+        H, W = img.height, img.width
+        if H / W > self.max_aspect_ratio:
+            img = T.CenterCrop((self.max_aspect_ratio * W, W))(img)
+        elif W / H > self.max_aspect_ratio:
+            img = T.CenterCrop((H, self.max_aspect_ratio * H))(img)
+
+        if size is None:
+            # resize image for max_seq_len, while keep the aspect ratio
+            H, W = img.height, img.width
+            scale = min(
+                1.0,
+                math.sqrt(self.max_seq_len / ((H / self.d) * (W / self.d))))
+            rH = int(
+                H * scale) // self.d * self.d  # ensure divisible by self.d
+            rW = int(W * scale) // self.d * self.d
+        else:
+            rH, rW = size
+        img = T.Resize((rH, rW), interpolation=interpolation,
+                       antialias=True)(img)
+        return np.array(img, dtype=np.uint8)
+
+    @staticmethod
+    def get_config_template():
+        return dict_to_yaml('DATASet',
+                            __class__.__name__,
+                            ImageTextPairMSDatasetForACE.para_dict,
+                            set_name=True)
+
+    @staticmethod
+    def collate_fn(batch):
+        collect = defaultdict(list)
+        for sample in batch:
+            for k, v in sample.items():
+                collect[k].append(v)
+
+        new_batch = dict()
+        for k, v in collect.items():
+            if all([i is None for i in v]):
+                new_batch[k] = None
+            else:
+                new_batch[k] = v
+
+        return new_batch

scepter/modules/data/dataset/registry.py

@@ -258,6 +258,8 @@ class DataObject(object):
             if sampler_name == 'MixtureOfSamplers':
                 subsampler_configs = self.data_sampler_config.get(
                     'SUB_SAMPLERS', [])
+                keep_order = self.data_sampler_config.get(
+                    'KEEP_ORDER', False)
                 subsamplers = list()
                 subsampler_probs = list()
                 for ssconfig in subsampler_configs:
@@ -277,7 +279,7 @@ class DataObject(object):
                     subsampler_probs.append(prob)
                 self.batch_sampler = MixtureOfSamplers(subsamplers,
                                                        subsampler_probs, rank,
-                                                       seed)
+                                                       seed, keep_order = keep_order)
             elif sampler_name == 'MultiLevelBatchSampler':
                 self.batch_sampler = self._instantiate_multi_level_batch_sampler(
                     self.data_sampler_config, self.batch_size, rank, seed)
@@ -302,9 +304,10 @@ class DataObject(object):
         delimiter = sampler_config.get('DELIMITER', ',')
         path_prefix = sampler_config.get('PATH_PREFIX', '')
         prompt_prefix = sampler_config.get('PROMPT_PREFIX', '')
+        oss_prefix = sampler_config.get('OSS_PREFIX', '')
         return MultiLevelBatchSampler(batch_size, index_file, image_size,
                                       fields, delimiter, path_prefix,
-                                      prompt_prefix, rank, seed)
+                                      prompt_prefix, oss_prefix, rank, seed)
 
 
 def build_dataset_config(cfg, registry, logger=None, *args, **kwargs):
@@ -335,8 +338,12 @@ def build_dataset_config(cfg, registry, logger=None, *args, **kwargs):
             f'registry must be type Registry, got {type(registry)}')
 
     cfg = deep_copy(cfg)
-
     req_type = cfg.get('NAME')
+
+    from scepter.modules.utils.import_utils import LazyImportModule
+    sig = (registry.name.upper(), req_type)
+    LazyImportModule.import_module(sig)
+
     if isinstance(req_type, str):
         req_type_entry = registry.get(req_type)
         if req_type_entry is None:

scepter/modules/data/dataset/video_gen_dataset.py

@@ -0,0 +1,202 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+import io
+import os
+import random
+import sys
+import warnings
+
+import numpy as np
+import torch
+from tqdm import tqdm
+
+from scepter.modules.data.dataset import DATASETS, BaseDataset
+from scepter.modules.utils.distribute import we
+from scepter.modules.utils.file_system import FS
+
+try:
+    import decord
+    decord.bridge.set_bridge('torch')
+except ImportError:
+    warnings.warn(
+        'The `decord` package is required for loading the video dataset. Install with `pip install decord`'
+    )
+
+
+@DATASETS.register_class()
+class VideoGenDataset(BaseDataset):
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger=logger)
+        self.prompt_prefix = cfg.get('PROMPT_PREFIX', '')
+        self.path_prefix = cfg.get('PATH_PREFIX', '')
+        self.p_zero = cfg.get('P_ZERO', 0.0)
+        self.max_num_frames = cfg.get('NUM_FRAMES', 49)
+        self.fps = cfg.get('FPS', 8)
+        self.height = cfg.get('HEIGHT', 480)
+        self.width = cfg.get('WIDTH', 720)
+        self.skip_frames_start = cfg.get('SKIP_FRAMES_START', 0)
+        self.skip_frames_end = cfg.get('SKIP_FRAMES_END', 0)
+        self.data_type = cfg.get('DATA_TYPE', 't2v')
+
+    def worker_init_fn(self, worker_id, num_workers=1):
+        super().worker_init_fn(worker_id, num_workers=num_workers)
+        randseed = np.random.randint(0, 2**32 - num_workers - 1)
+        workerseed = randseed + worker_id
+        random.seed(workerseed)
+        np.random.seed(workerseed)
+
+    def _preprocess_video_data(self, video_path):
+
+        with FS.get_object(video_path) as video_data:
+            video_reader = decord.VideoReader(io.BytesIO(video_data),
+                                              width=self.width,
+                                              height=self.height)
+        video_num_frames = len(video_reader)
+
+        start_frame = min(self.skip_frames_start, video_num_frames)
+        end_frame = max(0, video_num_frames - self.skip_frames_end)
+        if end_frame <= start_frame:
+            frames = video_reader.get_batch([start_frame])
+        elif end_frame - start_frame <= self.max_num_frames:
+            frames = video_reader.get_batch(list(range(start_frame,
+                                                       end_frame)))
+        else:
+            indices = list(
+                range(start_frame, end_frame,
+                      (end_frame - start_frame) // self.max_num_frames))
+            frames = video_reader.get_batch(indices)
+
+        # Ensure that we don't go over the limit
+        frames = frames[:self.max_num_frames]
+        selected_num_frames = frames.shape[0]
+
+        # Choose first (4k + 1) frames as this is how many is required by the VAE
+        remainder = (3 + (selected_num_frames % 4)) % 4
+        if remainder != 0:
+            frames = frames[:-remainder]
+        selected_num_frames = frames.shape[0]
+
+        assert (selected_num_frames - 1) % 4 == 0
+
+        # Training transforms
+        frames = frames.float().div_(127.5).sub_(1.)
+        frames = frames.permute(3, 0, 1, 2).contiguous()  # [C, F, H, W]
+        return frames
+
+    def _parse_index(self, index):
+        meta = dict()
+        for key, value in zip(index[-1], index[:-1]):
+            if key in ['oss_key', 'path', 'video_path', 'target_video_path']:
+                meta['video_path'] = value
+            elif key in ['source_video_path', 'src_video_path']:
+                meta['src_video_path'] = value
+            elif key in ['prompt', 'caption', 'text']:
+                meta['prompt'] = value
+            elif key in ['width', 'height']:
+                meta[key] = int(value)
+            else:
+                meta[key] = value
+        return meta
+
+    def _get(self, index):
+        meta = self._parse_index(index)
+
+        video_path = os.path.join(self.path_prefix, meta.get('video_path', ''))
+        video = self._preprocess_video_data(video_path)
+
+        prompt = self.prompt_prefix + meta.get('prompt', '')
+        if self.mode == 'train' and np.random.uniform() < self.p_zero:
+            prompt = ''
+
+        item = {
+            'video': video,
+            'prompt': prompt,
+            'meta': meta,
+        }
+        if 'i2v' in self.data_type:
+            item['image'] = item['video'][:, :1, :, :]
+        if 'v2v' in self.data_type:
+            src_video_path = os.path.join(self.path_prefix,
+                                          meta.get('src_video_path', ''))
+            src_video = self._preprocess_video_data(src_video_path)
+            item['src_video'] = src_video
+        return item
+
+    def __len__(self):
+        return sys.maxsize
+
+    @staticmethod
+    def collate_fn(batch):
+        collect = {}
+        for sample in batch:
+            for k, v in sample.items():
+                if k not in collect:
+                    collect[k] = []
+                collect[k].append(v)
+        return collect
+
+
+@DATASETS.register_class()
+class VideoGenDatasetOTF(VideoGenDataset):
+    def __init__(self, cfg, logger=None):
+        super().__init__(cfg, logger)
+        self.data_file = cfg.DATA_FILE
+        self.delimiter = cfg.get('DELIMITER', '#;#')
+        self.fields = cfg.get('FIELDS', ['video_path', 'prompt'])
+        self.use_num = cfg.get('USE_NUM', -1)
+
+        from scepter.modules.model.registry import MODELS
+        model_cfg = cfg.get('MODEL', None)
+        if model_cfg is not None:
+            self.model = MODELS.build(
+                cfg.MODEL,
+                logger=logger).eval().requires_grad_(False).to(we.device_id)
+        self.items = self.parse_data(self.data_file, self.delimiter,
+                                     self.fields)
+        if self.use_num and self.use_num > 0:
+            self.items = self.items[:self.use_num]
+        self.data = self.encode(self.items)
+        self.real_number = len(self.data)
+        if model_cfg is not None:
+            self.model.to('cpu')
+            del self.model
+            torch.cuda.empty_cache()
+
+    def parse_data(self, data_file, delimiter, fields):
+        items = list()
+        with FS.get_object(data_file) as local_data:
+            rows = [
+                i.split(delimiter,
+                        len(fields) - 1)
+                for i in local_data.decode('utf-8').strip().split('\n')
+            ]
+            for i, row in enumerate(rows):
+                item = {}
+                for key, value in zip(self.fields, row):
+                    if key in ['oss_key', 'path', 'video_path']:
+                        item['video_path'] = value
+                    elif key in ['prompt', 'caption', 'text']:
+                        item['prompt'] = value
+                    elif key in ['width', 'height']:
+                        item[key] = int(value)
+                    else:
+                        item[key] = value
+                items.append(item)
+        return items
+
+    def encode(self, items):
+        self.logger.info('Start to encode video data [{}]!'.format(len(items)))
+        for item in tqdm(items):
+            video_path = os.path.join(self.path_prefix,
+                                      item.get('video_path', ''))
+            video = self._preprocess_video_data(video_path)
+            latent = self.model.encode_first_stage(
+                video.unsqueeze(0).to(we.device_id)).squeeze(0)
+            item['video_latent'] = latent.detach().cpu()
+            item['video'] = video
+            if self.data_type == 'i2v':
+                item['image'] = item['video'][:, :1, :, :]
+        return items
+
+    def _get(self, index):
+        return self.data[index % self.real_number]

scepter/modules/data/sampler/__init__.py

@@ -1,9 +1,31 @@
 # -*- coding: utf-8 -*-
 # Copyright (c) Alibaba, Inc. and its affiliates.
+from typing import TYPE_CHECKING
+from scepter.modules.utils.import_utils import LazyImportModule
 
-from scepter.modules.data.sampler.base_sampler import BaseSampler
-from scepter.modules.data.sampler.registry import SAMPLERS
-from scepter.modules.data.sampler.sampler import (
-    EvalDistributedSampler, LoopSampler, MixtureOfSamplers,
-    MultiFoldDistributedSampler, MultiLevelBatchSampler,
-    MultiLevelBatchSamplerMultiSource, ResolutionBatchSampler)
+
+if TYPE_CHECKING:
+    from scepter.modules.data.sampler.base_sampler import BaseSampler
+    from scepter.modules.data.sampler.registry import SAMPLERS
+    from scepter.modules.data.sampler.sampler import (
+        EvalDistributedSampler, LoopSampler, MixtureOfSamplers,
+        MultiFoldDistributedSampler, MultiLevelBatchSampler,
+        MultiLevelBatchSamplerMultiSource, ResolutionBatchSampler)
+else:
+    _import_structure = {
+        'base_sampler': ['BaseSampler'],
+        'registry': ['SAMPLERS'],
+        'sampler': ['EvalDistributedSampler', 'LoopSampler',
+                    'MixtureOfSamplers', 'MultiFoldDistributedSampler',
+                    'MultiLevelBatchSampler', 'MultiLevelBatchSamplerMultiSource',
+                    'ResolutionBatchSampler']
+    }
+
+    import sys
+    sys.modules[__name__] = LazyImportModule(
+        __name__,
+        globals()['__file__'],
+        _import_structure,
+        module_spec=__spec__,
+        extra_objects={},
+    )

scepter/modules/data/sampler/registry.py

@@ -35,8 +35,12 @@ def build_sampler_config(cfg, registry, logger=None, **kwargs):
             f'registry must be type Registry, got {type(registry)}')
 
     cfg = deep_copy(cfg)
-
     req_type = cfg.get('NAME')
+
+    from scepter.modules.utils.import_utils import LazyImportModule
+    sig = (registry.name.upper(), req_type)
+    LazyImportModule.import_module(sig)
+
     if isinstance(req_type, str):
         req_type_entry = registry.get(req_type)
         if req_type_entry is None: