avtc
01a324bb11
Extends tiling beyond UNet (Conv2d) models to support DiT architectures using toroidal attention patches and latent content wrapping for seamless infinite tiling. Adds rectangular tiling mode, WanVAE 5D tensor handling, and an AdvancedTilingRay node for distributed Raylight workers.
ComfyUI Advanced Tiling
ComfyUI tiling nodes inspired by spinagon/ComfyUI-seamless-tiling. This project enables the creation of tileable images in various shapes with customizable rotations. Example workflows are located in the workflows directory.
⚠️ There are some artifacts present. I'm not sure if this is because of a flawed implementation or simply because Stable Diffusion isn't intended for this purpose.
Implemented tiling modes
- Hexagon
- None (normal generation)
Supported models
- Stable Diffusion 1.5 (also 1.4)
- Stable Diffusion 2.1 (also 2.0)
- Stable Diffusion XL (SDXL)
TODO
- More tiling modes
- Optimize VAE decode (first pass is very slow)
- Support DiT based models (SD3, PixArt-Σ, FLUX.1)
Credits
spinagon/ComfyUI-seamless-tiling [GPL-3.0] - Used as a base for this project
Red Blob Games - Hexagonal Grids - Hexagonal grid math