MiniMax H3 Two-Pass Sigma Plan / H3二采Sigma计划 (Advanced)

The two-pass idea - generate cheap and coarse, then refine with more detail - is the oldest trick in the diffusion book. H3 gets its own version here, and the clever part is in how the schedule is split. MiniMaxH3TwoPassSigmaPlanT8Advanced takes one native H3 rectified-flow trajectory, cuts it in base-flow time into a coarse half and a refine half, and then maps both halves through the actual video/audio shift values carried by your model. Default coarse_steps=4 + refine_steps=4 keeps the total at eight joint AV Transformer calls - no more compute than an 8-step single run, just arranged as two passes.

Why split at all

A single pass does everything at one resolution/noise posture. The two-pass shape lets the coarse pass settle structure and the refine pass spend its steps on detail - the flow-matching-era cousin of SD's hires fix, but done in latent space and respecting H3's joint AV clock. The restart_base_noise value (0.5 default) controls where the refine pass re-seeds its noise relative to the coarse result, which is the knob that decides how much "new" the refine pass is allowed to add.

The inputs are four and simple

• model - the tooltip tells you exactly where from: the MODEL output of the MiniMax H3 Dual-Clock Sampler. It needs that specific output because that's what carries the actual video/audio shift parameters it maps the sigma halves through.
• coarse_steps (4) and refine_steps (4) - the two halves' step counts. The author's documented sweet spot is 4+4 = 8 total forwards; you can push either half but you're paying per forward call.
• restart_base_noise (0.5) - refine-pass noise re-injection level.

Outputs:

coarse_sigmas and refine_sigmas (feed the first and second SamplerCustomAdvanced stages respectively) plus report_json. This is a planning node - it computes sigmas, it doesn't sample. Your SamplerCustomAdvanced nodes do the actual work.

Installing:

part of MiniMax H3 Audio T8. Manager → "MiniMax H3 Audio T8", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

Restart. No mandatory pip deps; H3 weights are on you.

The honest positioning

This is the schedule of the two-pass system; it works together with Latent Reconcile (which stitches the two passes' latents together) and the Detail Mixer (which tunes the refine pass). The README is careful not to overclaim: a mechanically clean two-pass run doesn't automatically mean better quality or safe VRAM - H3 is a big 33B model and the second pass is a second set of forwards on the same memory budget. Verify with your eyes and ears, and watch your VRAM headroom on 16GB cards.