MiniMax H3 Refine Target Geometry [Experimental]

MiniMax H3 Refine Target Geometry is the most honest node in this pack, because it does exactly one thing and says so on the tin: it mirrors the scale-by-multiplier sizing that the companion MiniMax H3 Latent Upscaler + Refine (3D) node uses, and hands you the resulting target dimensions as plain integers. No latent upscaling. No sampling. No model. Pure arithmetic.

Why would you want arithmetic? This node exists to feed a specific downstream experiment - Resolution-Aware Sigmas. That experimental node wants to know the target geometry of the learned-refine stage so it can derive a resolution-aware sigma map, but the refine sizing logic lives in the companion upscaler's node. Rather than duplicate that logic or hardcode numbers, this node reimplements the same calculation so your graph can compute target width/height and pass them into the sigma-map experiment as honest, traceable inputs.

Inputs are the geometry parameters: base_width and base_height (feed the same MP dimensions that already drive Continuum - that's the point, one source of truth), scale (multiplier, default 1.2), align (rounding step, default 32, matching H3's geometry-snapping world), and keep_proportion (default true). Outputs: target_width and target_height as INTs plus an H3_FLOW_DIAGNOSTICS blob recording all inputs, the resolution, and a semantics note reminding you this is refine metadata and never a Continuum resize.

The author's placement warning is worth quoting in spirit because it's the easiest way to break your graph: feed the same base MP into this node and into Continuum, use this node's outputs only to parameterize Resolution-Aware Sigmas on the downstream refine SIGMAS path, and do not feed them back into Continuum. Continuum keeps its ordinary MP/base geometry. This node is a metadata mirror, not an authority.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/MiniMax-H3-Flow-Aligned-Regenerate.git

Restart. No dependencies, no models.

Should you even use it?

Only if you're running the resolution-aware refine SIGMAS experiment, which the author's own matched media testing currently recommends keeping off (mode=off) because it produced no relevant improvement. If that sounds like a very specific door into a very narrow room - it is. This node is a supporting utility for an experimental branch of a single-author research pack. File it under "good to know exists," skip it for real work, and if you do open the experiment, wire it exactly as the README's placement diagram shows: base sizing fans out to Continuum and to this node, and the resulting integers go only to the sigma map.