update v0.0.4
This commit is contained in:
LouieStark
@@ -79,6 +79,7 @@ EXTENSION_PARAS:
|
||||
MANTRA_BOOK: scepter/methods/studio/extensions/mantra_book/mantra_book.yaml
|
||||
OFFICIAL_TUNERS: scepter/methods/studio/extensions/tuners/official_tuners.yaml
|
||||
OFFICIAL_CONTROLLERS: scepter/methods/studio/extensions/controllers/official_controllers.yaml
|
||||
TUNER_MANAGER: scepter/methods/studio/tuner_manager/tuner_manager.yaml
|
||||
CONTROLABLE_ANNOTATORS:
|
||||
-
|
||||
NAME: "CannyAnnotator"
|
||||
|
||||
@@ -0,0 +1,268 @@
|
||||
NAME: LARGEN
|
||||
IS_DEFAULT: True
|
||||
DEFAULT_PARAS:
|
||||
PARAS:
|
||||
RESOLUTIONS: [[1024, 1024]]
|
||||
INPUT:
|
||||
IMAGE:
|
||||
ORIGINAL_SIZE_AS_TUPLE: [1024, 1024]
|
||||
TARGET_SIZE_AS_TUPLE: [1024, 1024]
|
||||
AESTHETIC_SCORE: 6.0
|
||||
NEGATIVE_AESTHETIC_SCORE: 2.5
|
||||
PROMPT: ""
|
||||
NEGATIVE_PROMPT: ""
|
||||
PROMPT_PREFIX: ""
|
||||
CROP_COORDS_TOP_LEFT: [0, 0]
|
||||
SAMPLE: ddim
|
||||
SAMPLE_STEPS: 50
|
||||
GUIDE_SCALE: 7.5
|
||||
GUIDE_RESCALE: 0.5
|
||||
DISCRETIZATION: trailing
|
||||
REFINE_SAMPLE: ddim
|
||||
REFINE_GUIDE_SCALE: 7.5
|
||||
REFINE_GUIDE_RESCALE: 0.5
|
||||
REFINE_DISCRETIZATION: trailing
|
||||
OUTPUT:
|
||||
LATENT:
|
||||
BEFORE_REFINE_IMAGES:
|
||||
IMAGES:
|
||||
SEED:
|
||||
MODULES_PARAS:
|
||||
FIRST_STAGE_MODEL:
|
||||
FUNCTION:
|
||||
-
|
||||
NAME: encode
|
||||
DTYPE: float32
|
||||
INPUT: ["IMAGE"]
|
||||
-
|
||||
NAME: decode
|
||||
DTYPE: float32
|
||||
INPUT: ["LATENT"]
|
||||
PARAS:
|
||||
# SCALE_FACTOR DESCRIPTION: The vae embeding scale. TYPE: float default: 0.18215
|
||||
SCALE_FACTOR: 0.13025
|
||||
SIZE_FACTOR: 8
|
||||
DIFFUSION_MODEL:
|
||||
FUNCTION:
|
||||
-
|
||||
NAME: forward
|
||||
DTYPE: float16
|
||||
INPUT: ["SAMPLE_STEPS", "SAMPLE", "GUIDE_SCALE", "GUIDE_RESCALE", "DISCRETIZATION"]
|
||||
COND_STAGE_MODEL:
|
||||
FUNCTION:
|
||||
-
|
||||
NAME: encode
|
||||
DTYPE: float16
|
||||
INPUT: ["ORIGINAL_SIZE_AS_TUPLE", "CROP_COORDS_TOP_LEFT", "PROMPT", "NEGATIVE_PROMPT"]
|
||||
REFINER_MODEL:
|
||||
FUNCTION:
|
||||
-
|
||||
NAME: forward
|
||||
DTYPE: float16
|
||||
INPUT: ["SAMPLE_STEPS", "REFINE_SAMPLE", "REFINE_GUIDE_SCALE", "REFINE_GUIDE_RESCALE", "REFINE_DISCRETIZATION"]
|
||||
REFINER_COND_MODEL:
|
||||
FUNCTION:
|
||||
-
|
||||
NAME: encode
|
||||
DTYPE: float16
|
||||
INPUT: ["ORIGINAL_SIZE_AS_TUPLE", "AESTHETIC_SCORE", "NEGATIVE_AESTHETIC_SCORE", "CROP_COORDS_TOP_LEFT", "PROMPT", "NEGATIVE_PROMPT"]
|
||||
|
||||
MODEL:
|
||||
PRETRAINED_MODEL: ms://damo/LARGEN@models/largen_ckpt_s22k.pth
|
||||
# SCHEDULE_ARGS DESCRIPTION: TYPE: default: ''
|
||||
SCHEDULE:
|
||||
PARAMETERIZATION: "eps"
|
||||
TIMESTEPS: 1000
|
||||
ZERO_TERMINAL_SNR: False
|
||||
SCHEDULE_ARGS:
|
||||
# NAME DESCRIPTION: TYPE: default: ''
|
||||
NAME: "scaled_linear"
|
||||
BETA_MIN: 0.00085
|
||||
BETA_MAX: 0.0120
|
||||
# DIFFUSION_MODEL DESCRIPTION: TYPE: default: ''
|
||||
DIFFUSION_MODEL:
|
||||
# NAME DESCRIPTION: TYPE: default: 'DiffusionUNetXL'
|
||||
NAME: LargenUNetXL
|
||||
# PRETRAINED_MODEL DESCRIPTION: Whole model's pretrained model path. TYPE: NoneType default: None
|
||||
PRETRAINED_MODEL:
|
||||
# IN_CHANNELS DESCRIPTION: Unet channels for input, considering the input image's channels. TYPE: int default: 4
|
||||
IN_CHANNELS: 9
|
||||
# OUT_CHANNELS DESCRIPTION: Unet channels for output, considering the input image's channels. TYPE: int default: 4
|
||||
OUT_CHANNELS: 4
|
||||
# NUM_RES_BLOCKS DESCRIPTION: The blocks's number of res. TYPE: int default: 2
|
||||
NUM_RES_BLOCKS: 2
|
||||
# MODEL_CHANNELS DESCRIPTION: base channel count for the model. TYPE: int default: 320
|
||||
MODEL_CHANNELS: 320
|
||||
# ATTENTION_RESOLUTIONS DESCRIPTION: A collection of downsample rates at which attention will take place. May be a set, list, or tuple. For example, if this contains 4, then at 4x downsampling, attentio will be used. TYPE: list default: [4, 2]
|
||||
ATTENTION_RESOLUTIONS: [4, 2]
|
||||
# DROPOUT DESCRIPTION: The dropout rate. TYPE: int default: 0
|
||||
DROPOUT: 0
|
||||
# CHANNEL_MULT DESCRIPTION: channel multiplier for each level of the UNet. TYPE: list default: [1, 2, 4]
|
||||
CHANNEL_MULT: [1, 2, 4]
|
||||
# CONV_RESAMPLE DESCRIPTION: Use conv to resample when downsample. TYPE: bool default: True
|
||||
CONV_RESAMPLE: True
|
||||
# DIMS DESCRIPTION: The Conv dims which 2 represent Conv2D. TYPE: int default: 2
|
||||
DIMS: 2
|
||||
# NUM_CLASSES DESCRIPTION: The class num for class guided setting, also can be set as continuous. TYPE: str default: 'sequential'
|
||||
NUM_CLASSES: sequential
|
||||
# USE_CHECKPOINT DESCRIPTION: Use gradient checkpointing to reduce memory usage. TYPE: bool default: False
|
||||
USE_CHECKPOINT: False
|
||||
# NUM_HEADS DESCRIPTION: The number of attention heads in each attention layer. TYPE: int default: -1
|
||||
NUM_HEADS: -1
|
||||
# NUM_HEADS_CHANNELS DESCRIPTION: If specified, ignore num_heads and instead use a fixed channel width per attention head. TYPE: int default: 64
|
||||
NUM_HEADS_CHANNELS: 64
|
||||
# USE_SCALE_SHIFT_NORM DESCRIPTION: The scale and shift for the outnorm of RESBLOCK, use a FiLM-like conditioning mechanism. TYPE: bool default: False
|
||||
USE_SCALE_SHIFT_NORM: False
|
||||
# RESBLOCK_UPDOWN DESCRIPTION: Use residual blocks for up/downsampling, if False use Conv. TYPE: bool default: False
|
||||
RESBLOCK_UPDOWN: False
|
||||
# USE_NEW_ATTENTION_ORDER DESCRIPTION: Whether use new attention(qkv before split heads or not) or not. TYPE: bool default: True
|
||||
USE_NEW_ATTENTION_ORDER: True
|
||||
# USE_SPATIAL_TRANSFORMER DESCRIPTION: Custom transformer which support the context, if context_dim is not None, the parameter must set True TYPE: bool default: True
|
||||
USE_SPATIAL_TRANSFORMER: True
|
||||
# TRANSFORMER_DEPTH DESCRIPTION: Custom transformer's depth, valid when USE_SPATIAL_TRANSFORMER is True. TYPE: list default: [1, 2, 10]
|
||||
TRANSFORMER_DEPTH: [1, 2, 10]
|
||||
# TRANSFORMER_DEPTH_MIDDLE DESCRIPTION: Custom transformer's depth of middle block, If set None, use TRANSFORMER_DEPTH last value. TYPE: NoneType default: None
|
||||
# TRANSFORMER_DEPTH_MIDDLE: None
|
||||
# CONTEXT_DIM DESCRIPTION: Custom context info, if set, USE_SPATIAL_TRANSFORMER also set True. TYPE: int default: 2048
|
||||
CONTEXT_DIM: 2048
|
||||
# DISABLE_SELF_ATTENTIONS DESCRIPTION: Whether disable the self-attentions on some level, should be a list, [False, True, ...] TYPE: NoneType default: None
|
||||
# DISABLE_SELF_ATTENTIONS: None
|
||||
# NUM_ATTENTION_BLOCKS DESCRIPTION: The number of attention blocks for attention layer. TYPE: NoneType default: None
|
||||
# NUM_ATTENTION_BLOCKS: None
|
||||
# DISABLE_MIDDLE_SELF_ATTN DESCRIPTION: Whether disable the self-attentions in middle blocks. TYPE: bool default: False
|
||||
DISABLE_MIDDLE_SELF_ATTN: False
|
||||
# USE_LINEAR_IN_TRANSFORMER DESCRIPTION: Custom transformer's parameter, valid when USE_SPATIAL_TRANSFORMER is True. TYPE: bool default: True
|
||||
USE_LINEAR_IN_TRANSFORMER: True
|
||||
# ADM_IN_CHANNELS DESCRIPTION: Used when num_classes == 'sequential' or 'timestep'. TYPE: int default: 2816
|
||||
ADM_IN_CHANNELS: 2816
|
||||
# USE_SENTENCE_EMB DESCRIPTION: Used sentence emb or not, default False. TYPE: bool default: False
|
||||
USE_SENTENCE_EMB: False
|
||||
# USE_WORD_MAPPING DESCRIPTION: Used word mapping or not, default False. TYPE: bool default: False
|
||||
USE_WORD_MAPPING: False
|
||||
TRANSFORMER_BLOCK_TYPE: att_v2
|
||||
IMAGE_SCALE: 1.0
|
||||
USE_REFINE: False
|
||||
# FIRST_STAGE_MODEL DESCRIPTION: TYPE: default: ''
|
||||
FIRST_STAGE_MODEL:
|
||||
NAME: AutoencoderKL
|
||||
EMBED_DIM: 4
|
||||
IGNORE_KEYS: [ ]
|
||||
BATCH_SIZE: 1
|
||||
#
|
||||
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 DESCRIPTION: TYPE: default: ''
|
||||
COND_STAGE_MODEL:
|
||||
# NAME DESCRIPTION: TYPE: default: 'GeneralConditioner'
|
||||
NAME: GeneralConditioner
|
||||
USE_GRAD: False
|
||||
# EMBEDDERS DESCRIPTION: TYPE: default: ''
|
||||
EMBEDDERS:
|
||||
-
|
||||
# NAME DESCRIPTION: TYPE: default: 'FrozenCLIPEmbedder'
|
||||
NAME: FrozenCLIPEmbedder
|
||||
# PRETRAINED_MODEL DESCRIPTION: TYPE: str default: ''
|
||||
PRETRAINED_MODEL: ms://AI-ModelScope/clip-vit-large-patch14
|
||||
TOKENIZER_PATH: ms://AI-ModelScope/clip-vit-large-patch14
|
||||
# MAX_LENGTH DESCRIPTION: TYPE: int default: 77
|
||||
MAX_LENGTH: 77
|
||||
# FREEZE DESCRIPTION: TYPE: bool default: True
|
||||
FREEZE: True
|
||||
# LAYER DESCRIPTION: TYPE: str default: 'last'
|
||||
LAYER: hidden
|
||||
# LAYER_IDX DESCRIPTION: TYPE: NoneType default: None
|
||||
LAYER_IDX: 11
|
||||
# USE_FINAL_LAYER_NORM DESCRIPTION: TYPE: bool default: False
|
||||
USE_FINAL_LAYER_NORM: False
|
||||
UCG_RATE: 0.0
|
||||
INPUT_KEYS: ["prompt"]
|
||||
LEGACY_UCG_VALUE:
|
||||
-
|
||||
# NAME DESCRIPTION: TYPE: default: 'FrozenOpenCLIPEmbedder2'
|
||||
NAME: FrozenOpenCLIPEmbedder2
|
||||
# ARCH DESCRIPTION: TYPE: str default: 'ViT-H-14'
|
||||
ARCH: ViT-bigG-14
|
||||
# MAX_LENGTH DESCRIPTION: TYPE: int default: 77
|
||||
MAX_LENGTH: 77
|
||||
# FREEZE DESCRIPTION: TYPE: bool default: True
|
||||
FREEZE: True
|
||||
# ALWAYS_RETURN_POOLED DESCRIPTION: Whether always return pooled results or not ,default False. TYPE: bool default: False
|
||||
ALWAYS_RETURN_POOLED: True
|
||||
# LEGACY DESCRIPTION: Whether use legacy returnd feature or not ,default True. TYPE: bool default: True
|
||||
LEGACY: False
|
||||
# LAYER DESCRIPTION: TYPE: str default: 'last'
|
||||
LAYER: penultimate
|
||||
UCG_RATE: 0.0
|
||||
INPUT_KEYS: ["prompt"]
|
||||
LEGACY_UCG_VALUE:
|
||||
-
|
||||
# NAME DESCRIPTION: TYPE: default: 'ConcatTimestepEmbedderND'
|
||||
NAME: ConcatTimestepEmbedderND
|
||||
# OUT_DIM DESCRIPTION: Output dim TYPE: int default: 256
|
||||
OUT_DIM: 256
|
||||
UCG_RATE: 0.0
|
||||
INPUT_KEYS: ["original_size_as_tuple"]
|
||||
LEGACY_UCG_VALUE:
|
||||
-
|
||||
# NAME DESCRIPTION: TYPE: default: 'ConcatTimestepEmbedderND'
|
||||
NAME: ConcatTimestepEmbedderND
|
||||
# OUT_DIM DESCRIPTION: Output dim TYPE: int default: 256
|
||||
OUT_DIM: 256
|
||||
UCG_RATE: 0.0
|
||||
INPUT_KEYS: ["crop_coords_top_left"]
|
||||
LEGACY_UCG_VALUE:
|
||||
-
|
||||
# NAME DESCRIPTION: TYPE: default: 'ConcatTimestepEmbedderND'
|
||||
NAME: ConcatTimestepEmbedderND
|
||||
# OUT_DIM DESCRIPTION: Output dim TYPE: int default: 256
|
||||
OUT_DIM: 256
|
||||
UCG_RATE: 0.0
|
||||
INPUT_KEYS: ["target_size_as_tuple"]
|
||||
LEGACY_UCG_VALUE:
|
||||
-
|
||||
NAME: IPAdapterPlusEmbedder
|
||||
CLIP_DIR: ms://damo/LARGEN@models/clip_encoder/
|
||||
PRETRAINED_MODEL: ms://damo/LARGEN@models/ip-adapter-plus_sdxl_vit-h.bin
|
||||
INPUT_KEYS: [ "ref_ip", "ref_detail" ]
|
||||
IN_DIM: 1280
|
||||
HEADS: 20
|
||||
CROSSATTN_DIM: 2048
|
||||
-
|
||||
NAME: TransparentEmbedder
|
||||
INPUT_KEYS: [ "tar_x0", "tar_mask_latent" ]
|
||||
-
|
||||
NAME: NoiseConcatEmbedder
|
||||
INPUT_KEYS: [ "tar_mask_latent", "masked_x0" ]
|
||||
-
|
||||
NAME: TransparentEmbedder
|
||||
INPUT_KEYS: [ "ref_x0" ]
|
||||
-
|
||||
NAME: TransparentEmbedder
|
||||
INPUT_KEYS: [ "task" ]
|
||||
-
|
||||
NAME: TransparentEmbedder
|
||||
INPUT_KEYS: [ "image_scale" ]
|
||||
Reference in New Issue
Block a user