MiniMax H3 Enhance-A-Video + Long Video (Advanced EXP)

Long Video in H3 is a segmentation dance: you render segment by segment with context overlap, and each segment resumes from the last. If you want Enhance-A-Video in there, you can't just wrap the whole chain in one composer - a consumed audit token from segment N would leak into segment N+1 and the math falls apart. MiniMaxH3EnhanceAVideoLongVideoComposerT8Advanced is the segment-aware version: it keeps Long Video Conditioning as the scoped layout owner and adds one fresh EAV runtime per Stock20 segment, binding segment_index/context_frames to the resume state so each segment gets a clean audit.

How it works

Inputs: model (connect the MODEL output of the matching Long Video Conditioning node - the tooltip is explicit that it must come from that node), sigmas (the exact native Stock20 sigmas for this segment's sampler), segment_index (wire from the Long Video Planner's segment_index), and context_frames (wire from the Planner too - valid values are 0, 5, 22, or 39, matching H3's context-overlap options). Then the usual EAV knobs: mode (disabled / report_only / apply_exp), tau (4), start_video_progress/end_video_progress, max_workspace_mib (32), g_hard_limit (1.5).

The segment binding is the whole design: one composer instance per segment in the chain, each with its own runtime, each fed the exact segment/context identity. That's what "prevents a consumed audit token from leaking across segments" means in practice - no shared state, no cross-segment contamination, and the per-segment report_json tells you what each stretch of video actually ran.

Outputs

model (the composed model for this segment's sampler), runtime (feed the EAV Audit node - placed after this segment's sampling), and report_json. Because each segment owns a fresh runtime, you audit per segment.

Installing it

From MiniMax H3 Audio T8 (T8mars/T8star, GPL-3.0). 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; models to models/ by hand. H3 weights' Community License excludes the US, EU, UK, South Korea. (Long Video candidate save/finalize also wants ffmpeg on PATH - most ComfyUI bundles already have it.)

The honest limits

Same story as the other EAV composers: it currently passes low-load deterministic contract and import checks, explicitly not stress-tested, and the README refuses to claim quality, audio, speed, VRAM, or 16GB-safety wins. And the classic mistake is forgetting the context identity: if context_frames doesn't match what the Planner actually used (0 vs 5 vs 22 vs 39), the resume state binding is wrong and your audit will be lying about which segment it measured. Wire it from the Planner, don't type it by hand.