MiniMax H3 Selective Segment Repair / 选择性分段重做 (Advanced)

This is the entry point of the Selective Repair chain - the node that decides what needs fixing. It takes your long-video timeline and produces a non-destructive repair plan: a list of segments to redo, each with a compiled prompt, a repair seed, and context padding, all ready for the rest of the chain (RepairSegmentSelect → Bind → regenerate → Stage → Accept → Compose) to execute one by one. Its own job is planning only. It never overwrites, never auto-accepts, and never queues a generation.

You pick segments two ways, via selection_policy:

• manual (default) - you type the indices in manual_indices (e.g. "0,3,5"). Simple, explicit, you're in charge.
• evidence-based - feed a quality_report_json plus thresholds_json, and segments failing your thresholds get added automatically. The default threshold shape is {"identity":{"min":0.75}}, so this is built for quality reports that score identity/similarity per segment. If you have a downstream quality-scoring node, this is the auto-scheduling route.

Either way, the plan records a repair_mode (auto default), an optional prompt_addendum you can attach to a redo ("fix the left hand"), and a seed_stride (1009) used to derive a fresh, deterministic repair seed per segment so re-runs are reproducible.

Inputs worth touching

• timeline - a H3_T8_STUDIO_TIMELINE, required, from the pack's Studio/Long Video machinery.
• manual_indices - a comma/space-separated string when policy is manual.
• context_before_frames / context_after_frames - 22/22 defaults. These are the H3 grid-context paddings each redo carries so a regenerated segment blends into its neighbors instead of starting cold. 22 is the standard H3 context frame count; don't set them to 0 unless you specifically want zero-context redos.
• prompt_addendum - extra instructions for the redo prompt.
• quality_report_json / thresholds_json - only relevant in evidence mode.

Outputs: repair_plan (typed handle) and repair_plan_json for review.

Installing

In the T8mars MiniMax H3 Audio T8 pack:

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

or ComfyUI Manager → "MiniMax H3 Audio T8", restart. No pip deps; planning is pure computation.

The take

Start with manual - type the indices, review repair_plan_json, and check the context padding before you let it ride. The evidence route is seductive ("just fix whatever's below 0.75!") but it inherits the quality of whatever produced quality_report_json, and a bad score can pull in segments you never meant to touch. The context values are the real craftsmanship here: too little and your redo visibly clunks into its neighbors, too much and you're re-encoding nearly the whole scene. And remember the chain's iron rule - this node plans, it never executes. The plan is a wish list until a human accepts each item.