MiniMax H3 Detail Mixer Sampler / 混合细节采样 (Advanced)

H3's detail problems - slightly mushy faces, tails that look unfinished, shots that wobble - each have a community fix, but bolting four fixes onto one sampler by hand is a recipe for nonsense. MiniMaxH3DetailMixerSamplerT8Advanced bundles four of T8mars' experimental techniques into a single fail-closed sampler setup, and here's the crucial part: all four default to off. It's a sampler composer, not a magic "make it better" button. You opt in, one experiment at a time.

The four knobs

• Tail subdivision (enable_tail): adds extra_tail_steps at the low-sigma end with your choice of tail_spacing. The idea is the same one the KB's sampler notes describe for flow models - the last steps are where fine detail actually settles, and a few extra cheap steps there can buy texture without redoing the whole schedule.
• Smooth Model-Time Bias (enable_model_time_bias): applies a bias (default −0.025, so it only goes negative) over bias_start_progress→bias_end_progress on the video sigma domain. It's a gentle nudge to the model's time input late in sampling, not a CFG change.
• H3 STG (enable_stg): self-guidance for H3. stg_scale (0.35 default) with stg_double_blocks (default "25", the block index list) and a start/end progress window. This is the attention-level trick, so it costs extra compute where it runs.
• Joint AV RF Restart (enable_restart): re-noises to restart_video_sigma and takes restart_steps more passes. It's the "one do-over near the end" idea, on a fixed restart_seed.

The outputs nobody reads (but should)

Alongside the usual model/sampler/sigmas you wire into the sampler, this node reports actual_nfe and planned_joint_av_forwards. That second one is the honest bill: every enabled effect adds joint AV forwards, and H3 forwards are expensive (the model shares compute between video and audio). Check it before you queue a run, not after.

The discipline

One effect, one run, fixed seed, compare. The pack's own README is explicit that none of these are claimed as stable quality uplifts yet - they're experiments with mechanical checks, not endorsements. If you flip all four on and the clip comes out different, you won't know which knob did it. The default profile is turbo_standard8; the composer expects you to treat that as the baseline contract and stay on it while you experiment.

Installing it

From the MiniMax H3 Audio T8 pack. Manager search MiniMax H3 Audio T8, or:

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

restart. No forced pip deps; H3 weights/CLIP/VAE/LoRAs are yours to place in models/. H3 weights' Community License excludes the US, EU, UK and South Korea.

Where it bites

The trap is assuming enable_tail etc. are safe defaults because they're present. They default off for a reason - several of these effects only have single-material mechanical verification, and the README warns against stacking them with Enhance-A-Video, STG, BlockCache or other global patches on a first pass. If your run explodes or quality drops, check actual_nfe first: you may have accidentally paid for a 15-forward schedule on an 8-step budget.