MiniMax H3 Tiled Sampler

MiniMax H3 Tiled Sampler

Performance

Upscaling a 15-second video to about 2 megapixels (1080p) with this tiled sampler used roughly 75% of a 16 GB RTX 4060 Ti (~12 GB VRAM) and took 70 minutes.

A single ComfyUI node, MiniMax H3 Tiled Sampler (model/sampling/minimax), that denoises a MiniMax H3 audio-video latent in overlapping spatial tiles.

H3 packs video and audio into one attention sequence, so its cost is driven by the video token count (latent_t * h/2 * w/2) and its trained canvas stops at 768x1344 pixels. Tiling the video stream keeps every tile inside that native field of view and ties peak VRAM to the tile size rather than to the canvas, which is what makes larger canvases practical.

Inputs

Same controls as KSampler (seed, steps, cfg, sampler_name, scheduler, denoise), plus:

| Input | Meaning | | --- | --- | | tile_width / tile_height | Tile size in pixels, rounded down to 32. Defaults are the model's native canvas, 1344x768. | | tile_overlap | Overlap between neighbouring tiles in pixels. Larger values hide seams and add tiles. |

latent must be an H3 AV latent, from Empty MiniMax H3 AV Latent, MiniMax H3 Image to Video or MiniMax H3 Reference to Video. When the canvas fits in a single tile the node samples it whole, so it is a drop-in replacement for KSampler in an H3 workflow.

How it works

Sampling itself is unchanged: the node clones the model and adds a DIFFUSION_MODEL wrapper, then hands the latent to the stock sampler. Per model call, the wrapper:

1. Cuts the video latent into overlapping tiles aligned to the DiT's 2x2 patch grid, pulling the last tile of each axis back so all tiles are the same size.
2. Runs one forward per tile. Text, audio, keyframe and reference rows ride along in every tile, so audio stays global and guides keep working.
3. Gives each tile's rows the position ids they hold in the full packed layout, so RoPE coordinates stay absolute. Rebuilding the layout at tile size would renormalize the spatial grid and make the model read a tile as a whole frame.
4. Blends tile video velocities with linear ramps over the overlap and averages the audio velocities.

A PREPARE_SAMPLING wrapper shrinks the VRAM estimate to one tile's worth of video cells, so the model is not offloaded for a sequence length that is never built.

Notes

• Batch size is 1, as with H3 itself.
• Latent width and height must be multiples of 32 pixels, which every H3 latent node already guarantees.
• Attention cannot cross tile borders, so global spatial coherence weakens as tiles get smaller. Keep tiles at the native canvas size and raise tile_overlap if you see seams.
• Keyframe conditions are cropped per tile, so the 0.1% condition noise augmentation H3 applies by default is drawn per tile instead of once for the whole canvas.