H3 Tiled Sampler [WIP]

The [WIP] in the name is not decoration. H3 Tiled Sampler [WIP] is the same idea as H3 Tiled Sampler - denoise a MiniMax H3 latent in overlapping tiles across the frame and overlapping windows along the clip, blended every step, so a long or upscaled clip refines in pieces that fit your card - but it is a different implementation with a different control set, and the pack ships both so you can compare them rather than being told which one is right.

The difference that matters most is a single widget: anchor.

What anchor does, and why you would want it

Where the tiles split the frame, each tile can be guided by its own region of the original latent_image instead of being left to the sampler's own idea of the picture. auto (the default) does that whenever tiles split the frame - exactly the case where unguided tiles drift apart - on always applies it, off never.

If you have ever tiled a refine and watched the seams creep because each tile sharpened slightly differently, guidance against the source latent is the lever that fixes it. On the stock tiled sampler you do not get this dial. If your outputs from the non-WIP node look faintly gridded, try the WIP one with anchor on and compare.

The other differences

strength replaces denoise, and it is described as the noise level the refine starts from, on any scheduler. The tooltip is unusually concrete and worth internalising: 0.3 keeps the clip and sharpens it, 0.5 adds detail, 0.8 redraws detail on the clip's own layout, 0.93 redraws nearly all of it, 1.0 from noise. That is a more useful ladder than a bare number.

tiling on auto here prefers to keep the whole clip in one time window and split the frame into the fewest tiles under the token budget, adding time windows only where no tiling of one window fits or the clip runs past 362 frames. Note the direction: frame-first, time-second. That is the opposite emphasis of the non-WIP node, which reaches for temporal windows first. Neither is wrong - one protects temporal consistency, the other keeps spatial detail in a single continuous shot.

seed, steps (8 default, for a turbo LoRA; ~25 for base), cfg (1 for distilled), sampler_name and scheduler are the usual KSampler settings. positive, negative and latent_image are the usual inputs, and audio takes the clip's audio latent so a video-only latent holds sound instead of coming back silent.

Output: one LATENT.

Install

ComfyUI Manager → WAS Node Suite v3 → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was-node-suite-comfyui.git

ComfyUI 0.14.0+, Python 3.10+. The pack installs nothing of its own: requirements.txt holds a comment and no packages, and it never runs pip. Expect the first start (and the first after an update, during recompile) to take an extra second. No weights to download for this node.

Honest advice on which one to use

Start with the non-WIP H3 Tiled Sampler. It has the simpler contract, its auto tiling leans temporal, and for a straightforward "upscale then refine at 0.4" pass it just works.

Reach for the WIP node when you have a specific problem: visible grid structure that anchoring would fix, or a shot you would rather keep in one spatial piece. And treat it as an experiment - a [WIP] label in this pack is the author telling you the node is still moving, so do not build a shared workflow on it and be surprised when widget names shift.

Everything else is the standard H3 caveat set: no upscaling of its own, partial denoise is the design case, it is written for the H3 patcher, and if you are in the US, EU, UK or South Korea the community licence does not cover running these weights locally.