MiniMax H3 Resolution

MiniMax H3 doesn't letterbox. That one sentence explains why so many people fight with its resolutions - and why this node exists. With most video models you pick a resolution, and the model makes it work or letterboxes the difference. H3 doesn't: the sampled frame is whatever width/height you give it, and every reference image or video carries its own spatial grid into the same packed sequence. Type a resolution H3 didn't pick and you get a reframe - a crop or a zoom nobody asked for. This node is a canvas calculator that makes the numbers come out right.

The rule that drives everything

H3's grids are area-normalized: each axis covers (dim / sqrt(w·h)) · 32 units of RoPE space, centered on 16. Two consequences fall out, and they're worth internalizing because they're the whole model's behavior:

• Aspect ratio decides framing. The target and reference only register when their canvases have the same ratio. A wider target reaches past the reference's grid; a narrower one sees a center crop. That's the zoom people report.
• Absolute resolution moves nothing. The extent stays ratio · 32 whatever the pixel count - a bigger canvas just samples the same field of view more densely. So the canvas is free to scale, as long as the ratio is held.

So the node's philosophy: ratio from the reference, size from a multiplier - and never confuse the two. 1.0x is the native canvas H3 was trained on (768 short edge, capped at 768×1344, rounded to multiples of 32). Anything else is a deliberate step away from it.

Inputs and outputs

• aspect_ratio - "From reference", presets (16:9, 9:16, 1:1, 4:3, 3:4, 2.39:1), or "Custom". Default "From reference" takes the ratio from reference.
• reference (optional) - the clay pass / draft image whose encoded canvas sets the ratio. Note it's the encoded canvas, not the file size - that distinction matters (below).
• scale - 0.5x draft, 0.75x preview, 1.0x native, up to 2.0x, or Custom with custom_scale.
• custom_aspect_width / custom_aspect_height - used when ratio is Custom or no reference is connected.

Outputs are width, height, and an info string that tells you the native 1.0x canvas, the final canvas, the ratio, and - when a reference is attached - the percentage of reframe vs. the reference, plus a WARNING if it's over 1%.

The gotcha that makes the info output worth reading

The ratio comes from the reference's encoded canvas, and a pass below H3's pixel budget isn't upscaled - it's only rounded to 32. So 640×360 encodes as 640×352, ratio 1.818, where the native canvas would be 1.75. That's a 4% reframe nobody ordered. The node prints an explicit note when this happens, telling you to feed at least ~1,032k pixels to avoid it. Feed it a proper clay pass and the number just works.

Two more honest notes: at 1.5x–2.0x, attention grows with the square of the multiplier, so the VRAM hit is steeper than the canvas looks. And the pack's README frames this node exactly as described here - aspect from the reference clip, size scaled from H3's own canvas.

Install

It's part of Mickmumpitz-Nodes (MIT, deps: numpy, Pillow, opencv-python - nothing extra for this node). ComfyUI Manager, search "Mickmumpitz", or:

cd ComfyUI/custom_nodes && git clone https://github.com/mickmumpitz/ComfyUI-Mickmumpitz-Nodes.git

Restart, find "MiniMax H3 Resolution" under Mickmumpitz/utils. Wire the outputs into your H3 width/height inputs and stop hand-computing ratios - the node exists so the math is done once, correctly, and you never have to think about RoPE units again.