2 Step Latent Upscale for Minimax H3

[gardonkulous]

The current refiner method for Minimax H3 works, but is inefficient because all steps need to be sampled again. The community has developed workflows for LTX-style latent upscaling in Minimax that work great and allow fewer steps on the upscale pass:
https://www.youtube.com/watch?v=jzLnoVBuU6I

It would be great to have an option like this in addition to the ReDetailer.

[roadmaus]

That looks very interesting. Thanks for the heads up. Will look into it if time grants it

[roadmaus]

@gardonkulous Currently working on it and then testing it. Will be on main soon if it checks out

[roadmaus]

Looked into it properly, and it's on main now (b1a98fa).

What I found first. The recipe in that video re-noises to the same sigma our refine already started from (their ManualSigmas start at 0.90; our 8-step / denoise 0.5 refine starts at 0.92). So the speed in the video comes from running 3–4 steps at the target instead of the full count again, plus a cheap first pass — not from the upscaler. And the README's four-way comparison decodes the interpolated latent raw, with no refine, which isn't the case we're in.

What I measured (same first pass, same seed, turbo 8 steps, 512 → 768, LBH-123-AI's 3D net vs the bicubic we had):

• At refine denoise 0.5 the net changes nothing on a face — you can't tell the two apart.
• At denoise 0.3, bicubic smears the latent grid across the face; the net holds it, at 4 and at 3 steps. That's where it earns its place: a lighter refine that keeps the first pass's identity, at about a third of the refine's cost (net + 3 steps ≈ 105 s vs ≈ 180 s for the 8-step bicubic refine here).
• It softens non-face texture slightly against bicubic-then-redraw (a fern behind the subject reads a bit smoother). Worth knowing.

What shipped. Drop minimax_h3_latent_upscaler_3d_fp16.safetensors into models/latent_upscale_models, pick it under the node's weights control (new Latent upscaler row), then in the resolution popover, past native, switch drawn up by from bicubic to trained. That applies the measured recipe — refine 0.30, 3 steps — and both dials stay yours to change; a steps dial is there on the bicubic road too now (all is the old behaviour). The network is re-implemented inside the pack, so no second custom node needed. Nothing changes for pieces that don't throw the switch.

Thanks for the pointer — it turned into a real improvement, just not the one the video sells.

[gardonkulous]

Cool I can't wait to try it!