Amazing Upscale Node Workflow for DIT Model

This workflow is designed for simple logic amazing upscale nodes in the DIT model. It supports common applications for Flux, Hunyuan, and SD3. The workflow tiles the initial image into smaller pieces, uses an image-interrogator to extract prompts for each tile, and performs an accurate upscale process. This approach minimizes hallucinations and ensures proper condition handling.

We hope you enjoy using it!

What's New

TeaCache Sampler Integration for Hunyuan Video

Thanks to the contributions from the TeaCache code repository (ali-vilab/TeaCache) and code references from facok/ComfyUI-TeaCacheHunyuanVideo, we’ve added support for the TeaCache sampler.

• How to Use: Replace the samplercustomadvanced node in the official workflow with the TeaCache sampler node. Adjust the acceleration rate as needed to start using it.

• Performance: In testing with an NVIDIA 4090, rendering a 720×480 resolution video with 65 frames took only 55 seconds using a speedup factor of x2.1. This is approximately twice as fast as the original method.

• Caution: While the TeaCache sampler significantly accelerates processing, it may reduce image quality and dynamic effects. Use with discretion.

• Precision Support: Supports bf16 and fp8.

https://github.com/user-attachments/assets/af06b9d3-9c84-4a83-ba90-eb4ec4bb2e99

---

Instructions

Smart Tile Workflow

The Smart Tile nodes provide an interactive tile workflow for layouts where you want to split an image by visual structure, such as face, hands, subject, clothes, foreground, and background.

Current workflow nodes:

• TTP Smart Tile Interactive Crop: loads or receives a source image, provides the visual tile editor, supports manual tiles, painted masks, and SAM/QwenVL auto tile requests, and outputs a variable-size tile_set.
• TTP Smart Tile Set Preview: previews a tile set as a contact sheet or a single selected tile.
• TTP QwenVL3 Local Loader: loads a local QwenVL tagging model from ComfyUI models.
• TTP Smart Tile QwenVL Prompt Set Builder: prepares per-tile prompts before loop processing.
• TTP Smart Tile Loop Source: outputs one tile at a time for sampler/img2img processing.
• TTP Smart Tile Loop Collect: collects processed tiles back into the tile set.
• TTP Smart Tile Image Upscale Prep: optionally upscales one tile before sampling.
• TTP Smart Tile Assemble: assembles processed tiles back into the final image with feathered blending, mask support, priority/layer handling, color correction, optional CPU/GPU pixel alignment, and optional GPU paste/weight accumulation.
• TTP Smart Tile Save Final Image: saves only the final loop result and embeds workflow metadata.

Interactive loop workflow:

TTP Smart Tile Interactive Crop
  -> TTP Smart Tile QwenVL Prompt Set Builder (optional)
  -> TTP Smart Tile Loop Source
  -> VAE Encode / Sampler / VAE Decode
  -> TTP Smart Tile Loop Collect
  -> TTP Smart Tile Assemble
  -> TTP Smart Tile Save Final Image

By default, TTP Smart Tile Assemble uses assemble_mode=final_only. Connect TTP Smart Tile Loop Collect.done to TTP Smart Tile Assemble.done so loop runs return a lightweight preview while done is false, then perform the full assemble once after the last tile. Switch assemble_mode to always only when you want a full recomposite after every tile; if pixel alignment is enabled, unfinished loop runs are automatically treated as final_only to avoid repeated expensive alignment passes. assemble_device controls the paste/weight accumulation device (auto, cpu, or gpu), and pixel alignment can use the GPU canvas when GPU assemble is active. Use base_canvas_mode=black when you want connected source/base images to remain available as references without being pasted underneath the tiles. Enable small_tile_on_top when small detail tiles should automatically stack above larger body/background/context tiles in overlap areas. Official Transfer Color methods and PIL resize/crop handling remain on their existing paths for compatibility.

TTP Smart Tile Image Upscale Prep prepares each loop tile before img2img sampling. It can use a connected ComfyUI UPSCALE_MODEL through the same tiled upscale-model path as the built-in upscale node, or fall back to lanczos, bicubic, bilinear, area, or nearest resize when no model is connected or use_upscale_model is off. scale sets the requested enlargement, max_megapixels caps the final tile pixel count, and round_to snaps the final width/height after the cap. When the cap is active, the node rounds down so the rounded tile stays under the megapixel budget. Tile coordinates are not changed; they remain in original-image space so assemble can map the processed tile back by sample_box and output_scale.

TTP Smart Tile Interactive Crop is the recommended starting point when you want to manually or automatically split a still image by visual regions. Its image input follows the official Load Image pattern, so uploads go to ComfyUI's input folder and the workflow stores the selected filename instead of embedding the whole image. The editor can generate a standard grid from column/row numbers, replace the full layout with that grid, subdivide the currently selected tile, add painted-mask tiles, and fill uncovered gaps. It stores the tile layout in a hidden widget so the workflow keeps the current plan.

Example layout:

{
  "defaults": {
    "pad": 128,
    "blend": 48,
    "priority": 50,
    "importance": 1.0
  },
  "tiles": [
    {
      "name": "full_image",
      "x": 0,
      "y": 0,
      "w": 1.0,
      "h": 1.0,
      "pad": 0,
      "blend": 96,
      "priority": 10,
      "importance": 0.5
    },
    {
      "name": "face",
      "x": 0.35,
      "y": 0.08,
      "w": 0.30,
      "h": 0.28,
      "pad": 192,
      "blend": 64,
      "priority": 100,
      "importance": 1.0
    }
  ]
}

Coordinates can be pixel values or normalized values from 0.0 to 1.0. A rectangle whose coordinates are all in 0..1 is treated as normalized, including browser-serialized 0 and 1 edges. pad is seam overlap: it expands only the tile edges that touch another tile. Outer canvas edges and non-adjacent gap edges are not expanded. blend, priority, and importance control how the sampled tile is pasted back.

For standard grid layouts, edge tiles are expanded inward with real source pixels when needed so the ComfyUI IMAGE batch has a consistent size without fake outer padding. Irregular manual layouts with uncovered gaps may still need transport padding because a single IMAGE batch cannot contain mixed image sizes.

1. Image Tile Batch Node

This node cuts an image into pieces automatically based on your specified width and height. It also records the necessary information for further processing.

Parameter	Description
Width	The width of each tile.
Height	The height of each tile.
Image	The image to be tiled.

Node View:

---

2. Image Assembly Node

This node reassembles image tiles back into a complete image while preventing visible lines between the tiles. It operates in pixel mode.

Parameter	Description
Tiles	Input the tiled image batch. Replace individual tiles if needed.
Position	Paired with the Image Tile Batch Node.
Original Size	Paired with the Image Tile Batch Node.
Grid Size	Paired with the Image Tile Batch Node.
Padding	The padding value used to merge the image pieces.

Node View:

---

3. Tile Image Size Node

This node calculates the resolution of each tile based on the original image dimensions and your specified width/height factors.

Parameter	Description
Width Factor	Divides the image width into equal parts.
Height Factor	Divides the image height into equal parts.

For example: A width factor of 2 and a height factor of 3 will divide the image into 6 equal tiles.

Node View:

---

4. Coordinate Splitter Node

This node converts position information into coordinates and connects them to the corresponding positions.

Node View:

---

5. Cond to Batch Node

This node converts condition lists into batches. It is reserved for future functionality expansion and connects to the conditions.

Node View:

---

6. Condition Merge Node

This node merges all tiled conditions into one and prepares them for building the final image. It connects to the Coordinate Splitter Node and Cond to Batch Node.

Node View:

---

Examples

Pixel Example (Recommended)

Latent Example

---

ControlNet Tile Integration

This workflow supports ControlNet Tile for enhanced upscaling. Here's an example of using tiles with the Hunyuan DIT model:

Resource	Link
Tile Example	Hugging Face Tile
Hunyuan 1.2	Download Hunyuan 1.2

Workflow Example:

---

Star History