ComfyUI-MooshieTiledDiffusion

Standalone export of the tiled diffusion node from MooshieUI, published by request (issue #338). Provided as is, with no promise of maintenance or support.

What it does

Adds one node, Apply Tiled Diffusion, that patches a model to denoise in tiles and blend the tiles back together. This lets you generate and upscale above a model's native resolution without the seams and artifacts you get from a single large pass.

Two methods:

• MultiDiffusion: overlapping tiles with cosine feathered blending. Best quality.
• SpotDiffusion: random tile shift per denoising step, no overlap. Fastest.

It supports 4D and 5D latents, so it works with Anima (COSMOS based) models as well as standard diffusion models. It works in txt2img, img2img, and inpainting, and crops spatial conditioning (inpaint masks and ControlNet) per tile.

The only dependencies are torch and the ComfyUI API, both of which ComfyUI already provides. There are no extra packages to install.

Install

Pick one.

• ComfyUI-Manager: search for Mooshie Tiled Diffusion and install.
• Comfy Registry: install comfyui-mooshietileddiffusion.
• Manual: clone this repo into ComfyUI/custom_nodes and restart ComfyUI.

cd ComfyUI/custom_nodes
git clone https://github.com/Mooshieblob1/ComfyUI-MooshieTiledDiffusion

Usage

1. Add the Apply Tiled Diffusion node (category: model_patches/unet).
2. Connect your model into the node and the node MODEL output into your KSampler.
3. Set tile_width and tile_height to the model native resolution (for example 1024 for SDXL or Flux, 512 for SD1.5).
4. Set tile_overlap for MultiDiffusion. Higher overlap means smoother seams and slower generation. Overlap is ignored by SpotDiffusion.

Use MultiDiffusion when you want the cleanest result. Use SpotDiffusion when you want speed and can accept an experimental method.

This is useful above about 2.4 MP, where a single pass tends to show artifacts.

License

GPL-3.0-or-later. The node is derived from ComfyUI, which is GPL-3.0.

Credits

• ComfyUI by Comfy Org.
• MultiDiffusion: Bar-Tal et al., MultiDiffusion: Fusing Diffusion Paths for Controlled Image Generation, ICML 2023. arxiv.org/abs/2302.08113
• SpotDiffusion: Ding et al., 2024. arxiv.org/abs/2407.15507