tiled-diffusion-ng
Mirror of https://git.colorized.life/tiled-diffusion-ng/
Nodes (5)
Tiled Diffusion NG
The canonical home of this repository is at https://git.colorized.life/tiled-diffusion-ng/
Five ComfyUI nodes for tiled sampling, blending four overlapping views into one full-resolution latent through a single KSampler trajectory.
cd /path/to/ComfyUI/custom_nodes
git clone https://git.colorized.life/tiled-diffusion-ng.git tiled-diffusion-ng
# Restart ComfyUI.
Supported models
| Model | Supported | ControlNets | | --- | :---: | :---: | | SDXL | ✅ | ✅ native Apply Controlnet | | Anima | ✅ | ✅ TileAnimaLLLiteApply | | Krea2 | ✅ | ✅ TileKrea2Conditioning |
SDXL support covers base models and ordinary RGB SDXL ControlNet.
Nodes
| Display name | Node ID | Inputs → output |
| --- | --- | --- |
| TilePlan | TiledDiffusionNG_TilePlan | model, latent, tile_overlap → TILE_PLAN |
| TileView | TiledDiffusionNG_TileView | image, tile_plan → IMAGE batch |
| TileSampler | TiledDiffusionNG_TileSampler | KSampler inputs, tile_plan, optional local_positive → LATENT |
| TiledAnimaLLLiteApply | TiledDiffusionNG_TiledAnimaLLLiteApply | model, model_patch, tile_plan, reference_tiles, strength and schedule → MODEL |
| TileKrea2Conditioning | TiledDiffusionNG_TileKrea2Conditioning | clip, reference_tiles, optional prompts and baseline, strength, schedule and downsizing → CONDITIONING list |
Workflow
Connect the same model and target latent to the planner and sampler. The plan creates four overlapping views in clockwise order: TL, TR, BR, BL. Overlap is the shared width in pixels, default 64. Gaussian weights blend predictions at every model evaluation.
flowchart LR
A[Model and Latent] --> B[TilePlan] --> C["TileView (optional)"] --> D[TileSampler]
Vision views are optional and require a reference image matching the plan's full pixel dimensions. External nodes can turn those views into an execution list of four complete positive conditionings in tile order. Each replaces the global positive for its tile; the negative stays shared. Without local positives, all four tiles use the global positive.
Upscale the output latent externally and create a fresh plan to refine again, or decode it to an image.
License
Copyright © 2026 Lany Atwood [email protected]. The project's Python source and tests are licensed under AGPL-3.0-only.
Based on the Mixture of Diffusers equations by Álvaro Barbero Jiménez, with an independently written Gaussian and fusion implementation. See COPYRIGHT for algorithm references and attribution.