MiniMax H3 Motion Recovery Composer / 动作二采编排 (Advanced)

After MotionRetimingPrepareT8Advanced hands you a stretched latent, somebody has to decide how much of it gets re-sampled - and do it without breaking the chain's bookkeeping. That's MiniMaxH3MotionRecoveryComposerT8Advanced. It authenticates the prepared AV latent against the signed motion plan, then selects the final fraction of an existing H3 sigma schedule for the second V2V pass. In plain terms: it doesn't start the denoise from scratch, it takes the tail end of the sigma schedule so the re-sample only polishes rather than re-invents.

Two design details tell you this author is serious about not screwing up your pipeline. First, it installs no MODEL wrapper - the model stays stock, which means nothing is globally patched and the failure modes stay local. Second, it never auto-stacks Enhance-A-Video, STG, Restart or BlockCache on top of the repair pass. That's a trap the README explicitly warns about: layering global attention patches onto a first repair pass is how you get weird output and blame the wrong node. The composer refuses to do it silently.

Inputs

• av_latent - the stretched latent from Retiming Prepare.
• sigmas - the existing H3 sigma schedule it will take a fraction of.
• motion_plan - the signed plan, used to authenticate that this latent and schedule actually belong to this repair chain.
• mode - report_only by default: validate and report without committing to a real second pass.
• denoise_fraction (0.48) - the fraction of the schedule retained for pass 2. The tooltip is explicit: it's a fraction of the schedule, not a sigma value.
• minimum_second_pass_nfe (2) - a floor so the second pass doesn't get absurdly short.

Outputs

av_latent and sigmas (wired into the sampler), motion_plan (passed through), actual_second_pass_nfe (how many model evals the pass will really cost - worth watching, since every NFE is a full joint AV forward), and report_json.

Install and where it fits

Pack install once: ComfyUI Manager → search MiniMax H3 Audio T8, or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into ComfyUI/custom_nodes/, restart. No pip extras, no model downloads.

Chain position: Retiming Prepare → Recovery Composer → KSampler-style node → Recover AV. The report_only default is your friend - run it once to see actual_second_pass_nfe before you commit compute. And remember the honest framing: this whole Motion Recovery system is an Experimental chain, validated on specific clips by the author, not a universal quality guarantee. A mechanically clean second pass is a necessary condition for better fast-motion output, not a sufficient one - you still have to watch the result.