MiniMax H3 Two-Pass Detail Mixer / 二采混合细节设置 (Advanced)

This is the "make the refine pass smarter" node - a box of four separate detail techniques that bolt onto an existing high-resolution refine schedule without replacing it. MiniMaxH3TwoPassDetailMixerT8Advanced gives you Tail subdivision, Model-Time Bias, H3 STG, and joint AV RF Restart as independent toggles, so you can stack the ones that help your footage and leave the rest off. The description is firm about where it belongs: after learned latent upscale/reconcile, and Temporal Detail (if you use it) stays after AV Decode. It's a refinement layer, not a replacement for the base schedule.

The four techniques, decoded

• Tail subdivision (enable_tail, off by default) - adds extra_tail_steps at the quiet end of the schedule (the "tail" of sigma space, where the fine detail forms) with tail_spacing controlling how they're laid out. Cheap way to buy more care at the end without touching the coarse pass.
• Model-Time Bias (enable_model_time_bias) - applies a bias (default −0.025) to the model's time conditioning between bias_start_progress and bias_end_progress (0.7→0.95 of the run, in video_sigma domain by default). It nudges when the model thinks it is during the refine, which is a knob some find adds crispness in the late refinement window.
• H3 STG (enable_stg) - spatial-temporal guidance, H3's analog of classifier-free guidance applied across both space and time. stg_scale (0.35) is the guidance strength and stg_double_blocks (default "25") lists which transformer blocks get doubled for the guidance branch. Start at low scale; STG is where things can get weird fast.
• Joint AV RF Restart (enable_restart) - rectified-flow restart: re-inject noise at restart_video_sigma (0.15) and take restart_steps (3) more steps from there, with a fixed restart_seed. The classic "add a bit of noise back at the end and re-denoise to gain detail" maneuver, adapted to the joint AV latent.

Every toggle defaults off

That's the design intent - the node is inert until you explicitly reach for a technique, so it can't change old workflows. The inputs you'll touch first are shift_video/shift_audio (6/3), which set the sigma-shift posture the whole refine runs under.

Outputs:

model, sampler and sigmas (the modified triple for your SamplerCustomAdvanced), plus actual_nfe, planned_joint_av_forwards, and report_json. The last two matter because compute is the currency here - planned_joint_av_forwards tells you how many Transformer forwards this configuration will cost before you run it.

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.

The honest advice

Enable one technique at a time on a fixed seed - STG and Restart both have real failure modes (overshoot, detail soup) and you won't know which knob helped if you flip four at once. The author's own notes treat this whole family as "not proven to be a stable quality win," so treat it as a tool you audition, not a preset you trust.