MiniMax H3 Learned Two-Pass Parity Plan / H3学习放大原版二采计划 (Advanced)

The learned upscale in this pack works, but it produces a low-res latent that still needs a final high-res pass - and H3's joint audio-video sampler means you can't just feed it two different sigma schedules and hope the clocks line up. This node is the glue that makes the two-pass flow reproducible: it hands you the exact sigma sequences for the low-res and high-res passes, built to match the published reference schedule. If you've ever watched a two-pass upscale drift out of sync because the audio clock disagreed with the video clock, this is the node that exists to stop that.

It sits in T8/MiniMax H3/Latent/Experimental and is the plumbing between "I upscaled in latent space" and "I sampled the result properly."

How it works

You connect an H3 model - the tooltip is specific: it must carry the same video/audio shifts that both passes will use, because the native H3 sampler derives the matching audio clock from the active video/audio shifts. No shifts, no matching audio clock, no correct two-pass.

The node then reproduces two things: the published LBH low-resolution simple-schedule split (that's base_steps, default 8, and coarse_steps, default 4 - the low-res pass), and the current 3/4/5-step high-resolution raw video-sigma sequences (refine_steps, default 4). The defaults form the pack's "4+4 graph": 8 total joint AV forwards, which the README flags as the author-aligned two-pass profile. If you saved a workflow that explicitly stored 3 or 5 refine steps, it stays compatible - the tooltip is clear that 4 is the current default, not a hard requirement.

Worth the warning: this is a parity plan, not a quality promise. It reproduces the published schedule so the community can A/B against the reference. The README is emphatic that matching the reference's numbers doesn't prove the result looks good - only that it ran the same steps.

The inputs that matter

• model - must carry matching video/audio shifts. This is the one that breaks silently if wrong.
• coarse_steps (4) and base_steps (8) - the low-res pass profile.
• refine_steps (4, range 3–5) - the high-res pass. 4 is the current reference default.

Outputs

coarse_sigmas and refine_sigmas (both SIGMAS - wire them into the two sampler stages in order), plus report_json documenting the schedule that was built.

Installing it

Pack-level: ComfyUI Manager → "MiniMax H3 Audio T8", or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into ComfyUI/custom_nodes, restart. No extra pip deps, no downloads - this is a pure schedule-computation node. The model you connect is your own H3.

Common issues

The high-res pass sounds wrong even though the video looks right. Check that the model you fed in actually carries the same video/audio shifts as the sampler used. The audio clock is derived from those shifts; feed a model with different shifts and the audio clock lands in the wrong place. This is the exact failure the tooltip warns about.

"Only 3/4/5 refine steps are valid." The high-res profile is bounded at 3–5 by the published schedule. If you need more steps, that's a different experiment - this node reproduces the reference, it doesn't extend it.