H3 Tiles

H3 Tiles is the model-side answer to "my H3 clip is too big to refine". It takes a MODEL and a tiling mode and hands back a MODEL. That is it. But what it hands back is a model whose every call is split into overlapping tiles across the frame and overlapping windows along the clip, blended where they meet at each step, so the tiles behave like one clip instead of a grid of pictures that have to be stitched afterwards.

The difference between this and H3 Tiled Sampler is worth getting straight before you wire anything. Tiles wraps the model, so any sampler you already use - KSampler, a custom sampler, whatever your workflow is built around - refines a long H3 latent in pieces that fit the card. Tiled Sampler is a sampler node that does its own tiling internally. If you are happy with your sampler, take Tiles.

The workflow it belongs to

The shape is: upscale the latent first (or decode, upscale, re-encode), then sample at a partial denoise so the model adds detail rather than inventing a new clip.

H3 latent → Latent Upscale → H3 Tiles → KSampler (denoise ~0.4) → VAE Decode
                    ↘ model ────────────↗

Wrap the model after any LoRA, the same way you would with H3 Low VRAM. And this node does no upscaling itself - that is the point of it being a model wrapper. It only decides how the work gets divided.

How the tiling decides

tiling is a dynamic combo with a manual branch, but the two routes you will actually use are auto and manual. On auto the node splits the clip into windows along time under a token budget, and splits across the frame only where a window will not fit on its own. That ordering matters: temporal windows are cheaper to blend than spatial tiles, and it keeps cross-frame motion intact for as long as possible. On manual you set tile and window sizes yourself, which is what you want when you already know your card's limit and would rather not let the node guess.

Because the blend happens at every step rather than once at the end, seams do not accumulate into faint rectangular ghosting down a long clip. That is the whole reason a tiled sampler is more than a clever crop.

Install

ComfyUI Manager, search WAS Node Suite v3, install, restart. Or:

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

Requires ComfyUI 0.14.0+ and Python 3.10+. Nothing gets pip-installed - the pack's requirements.txt is a comment, and the README is explicit that install is the whole story, now and on later starts. No weights needed. First start (and the first start after each update, while it recompiles) takes a second or two longer.

Where people get burned

Expecting it to upscale. It does not. If your refinement is mushy, that is your upscale step or your denoise, not the tiling.

Full denoise. Tiling a latent from noise on a long clip gives you tiles that agree less and less as the denoise rises. Partial denoise is the design case; if you need to redraw the clip, go back to the sampler and do it in one pass or expect to fight seams.

Wrong tiling mode for a 16 GB card. auto is tuned to preserve time continuity, which can leave a single time window too wide for a small card and force spatial tiles anyway. If you know your ceiling, manual is more predictable.

Sitting it next to H3 Low VRAM and expecting both to help. They overlap; the tiled model already runs its calls the constrained way. Pick one - Tiles for bigger effective resolution, Low VRAM for long token sequences on a plain model.

Licensing, not VRAM. If you are in the US, EU, UK or South Korea, the MiniMax H3 community licence does not grant you the right to run the local weights, whatever your GPU can hold.