MiniMax H3 Resolution-Aware Sigmas [Experimental]

Here's the theory behind MiniMax H3 Resolution-Aware Sigmas: when a video diffusion model denoises at a higher resolution than it was trained for, the noise schedule arguably isn't in the right place anymore - the model is walking a path calibrated for a smaller canvas. SD3's paper (their Eq. 23) played with a resolution-dependent relative shift to compensate. This node borrows that idea, composes it onto H3's native flow shift, and remaps your SIGMAS accordingly for the higher-res stage.

And here's the honest part, which is exactly the kind of thing you want to hear before you wire experimental nodes into a pipeline that takes hours per clip: the author's own matched decoded-media tests found no relevant quality improvement, and the current recommendation is mode=off. The mapping works structurally - the math does what it claims - but in the E0/E1 matched learned-refine media runs, the resolution-aware version showed no meaningful gain over the exact-identity control. So this node is a correctly-built experiment that didn't pay off, kept around for people who want to run the experiment themselves.

Placement matters more than any setting. This sits on the refine SIGMAS path of the integrated learned upscaler/refiner, not on Continuum's main schedule:

existing refine scheduler SIGMAS
        -> Resolution-Aware Sigmas
        -> MiniMax H3 Latent Upscaler + Refine (3D).sigmas

Feed its output into Continuum SIGMAS and you're doing it wrong - Continuum keeps ordinary MP/base geometry. That's also why Refine Target Geometry exists: it supplies the target metadata (base MP × scale) that this node's target_width/target_height inputs expect.

The inputs that matter. mode: off (exact SIGMAS parity, the recommendation), resolution_aware (the SD3-derived relative area map on H3's shared AV coordinate), calibrated (research-only direct factor). source_width/source_height of 0/0 (the default) auto-derive H3's native reference canvas from your target aspect ratio - 768px short edge, 768×1344 area cap, 32px alignment - and do not imply a low-res sampling pass. strength (default 1.0): 0 means no relative shift, 1 means analytic sqrt(target_area/reference_area). The author's tooltip-level advice is worth obeying: don't tune strength before you've run a matched off/1.0 media pair. Outputs: remapped sigmas plus H3_FLOW_DIAGNOSTICS with the resolved reference mode, areas, area ratio, and the effective shift factors.

Install

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

Restart. No extra deps, no models.

Bottom line

If you're chasing the same question the author chased - does resolution-aware sigma remapping help H3's learned refine? - this is the node that answers it with a controlled comparison instead of vibes. Set mode=off for the control, then resolution_aware at strength 1.0, and let the matched clips and the diagnostics decide. The author's evidence says it won't help; the node exists precisely so you don't have to take their word for it.