SpotDiffusion Parameters

Straight from the author, no sugar-coating: this is the newest, least-baked tiling method in the pack, and it's labeled "experimental, in testing, and subject to change." If you're just trying to upscale something and want the best-looking result, use the regular method choices on the Tiled Diffusion node (Mixture of Diffusers or MultiDiffusion) instead - this node is for when raw speed matters more than final quality, and you already know that's the tradeoff you want.

What SpotDiffusion actually is

Rather than fixing tile positions and overlapping them (the approach every other method in this pack uses), SpotDiffusion - based on a 2024 paper - tries to eliminate seams by randomly shifting the denoise window on every timestep. Instead of the same tile boundaries sitting in the same place for the whole generation, they jitter around, so no single seam gets a chance to become visible and consistent across steps. The README is explicit about when this pays off: "it's mainly used for fast inferences by setting tile_overlap to 0; otherwise, it's better to stick with the other tiling strategies as they produce better outputs." In other words, its whole reason to exist is letting you run with zero overlap (which is normally what causes visible seams) and get away with it, because the shifting hides the seams instead of overlap-blending doing the work. That's a real speed win, since overlap means reprocessing the same pixels multiple times.

This node is genuinely obscure - it's new enough, and niche enough, that you won't find much discussion of it outside the pack's own docs. Treat that as confirmation of the author's own "experimental" label rather than a reason to avoid it; just go in knowing you're on the newest, least field-tested path in the pack.

How it fits with Tiled Diffusion

You still pick your tile sizing on the main Tiled Diffusion node - set its method to SpotDiffusion there (and set tile_overlap to 0 to actually get the speed benefit). This node chains onto that same patched model afterward to hand SpotDiffusion its own extra knobs before the result reaches your KSampler.

Inputs and outputs

• model (MODEL) - the model coming out of the Tiled Diffusion node, further patched here.
• shift_method (default random) - per the author's own tooltip: "Samples a shift size over a uniform distribution to shift tiles." The other two options, sorted and fibonacci, are alternative shift-ordering strategies; the author ships random as the default, so start there unless you have a specific reason to try the others.
• seed (default 0) - because the shifting is randomized, the seed is what makes a given SpotDiffusion run reproducible. Same seed, same prompt, same shift pattern.

Output is a MODEL, wired straight into your KSampler like any other model output in this pack.

How to install it

• ComfyUI Manager - search "Tiled Diffusion & VAE for ComfyUI", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/shiimizu/ComfyUI-TiledDiffusion, then restart.

Same install as every other node in this pack - no separate download, no extra dependency, it's all one repo. No model weights needed either; it patches whatever model you're already running.

Common issues & troubleshooting

Results look worse than Mixture of Diffusers or MultiDiffusion. That's expected, not a bug - the author's own README says the other strategies "produce better outputs" and frames this one as a speed play. If quality matters more than raw throughput for a given render, switch method back to one of the established options on the main Tiled Diffusion node.

You set tile_overlap above 0 and don't see much benefit. SpotDiffusion's speed case is specifically about running at tile_overlap = 0. If you're keeping overlap the way you'd set it for MultiDiffusion, you're paying the same compute cost for a method that was designed to skip it.

It's marked _for_testing and feels unfinished. It is, deliberately - the author says so outright. That's not a hidden landmine, it's the pack being upfront that this particular corner is still moving.