MiniMax H3 Joint AV RF Restart / 联合音画流重启 (Advanced)

Some quality problems don't show up until you see the finished clip, and re-rolling the whole generation is a sledgehammer. This sampler is a different idea: run the normal dual-clock trajectory to a clean result, then re-noise the clean endpoint partway with the rectified-flow interpolation and take a second joint AV descent. It's the sampler-level equivalent of "do it again, but from where we already got to" - stochastic restart, not just an extra tail pass.

The distinction matters, because the pack already has cheaper tricks for "this ending is rough." Tail subdivision only refines the low-noise end of the same trajectory. This node is deliberately more aggressive: restart_video_sigma (default 0.15, range 0–0.5) picks how far back up the noise axis you jump before the second descent. The higher the sigma, the more of the image gets to re-decide itself. That's the whole point, and the whole risk - the README is blunt that this is "true stochastic restart and is higher risk than tail subdivision." You're not guaranteed an improvement; you're paying for a chance that a second coherent descent lands in a better spot.

How it wires in

It's a sampler factory, not a sampler itself: it emits model, sampler, and sigmas that you feed into a KSampler-style node alongside your av_latent (the H3 nested audio-video latent). Standard dual-clock values: steps 8, shift_video 12, shift_audio 3. Then the restart-specific knobs:

• restart_video_sigma - re-noise depth for the second descent (0.15 default).
• restart_steps - how many steps the second descent gets (3 default, 0 disables the restart entirely).
• restart_seed - an independent seed for the re-noising so your second chance isn't the same noise as your first.

Plus report_json describing what it actually did.

Installing and the honest take

Same pack as everything: T8mars MiniMax H3 Audio T8 (Manager → "MiniMax H3 Audio T8", or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into custom_nodes), restart, no pip deps. Bring your own H3 model/VAE/LoRA.

Where this fits: you've got a clip that's structurally fine but has a weak region or a wobble you can't fix with prompt edits, and you have VRAM headroom to burn a second descent. Start with restart_steps low (2–3) and restart_video_sigma modest (0.1–0.2), then judge by eye and ear. If the first pass was already good, don't touch this - every restart is extra sampling cost for a maybe. And don't stack it with the pack's other experimental patches (Enhance-A-Video, RF Restart, BlockCache) on the same run; the README explicitly warns against piling unvalidated experiment on experiment.