MiniMax H3 Enhance-A-Video Runtime Audit (Advanced EXP)

Enhance-A-Video (FETA) is the trick where cross-frame attention is computed from the target video's Q/K and used to scale the target's attention output - the paper's way of sharpening motion without touching the whole architecture. In this pack it's wrapped in auditing: every EAV composer returns a runtime object, and MiniMaxH3EnhanceAVideoAuditT8Advanced is the node you place after sampling to verify the whole thing actually ran as claimed - the selected schedule's forward count, all 50 H3 DiT attention measurements, the observed g range, and the chunk workspace. Then it passes your latent through unchanged.

Why 50 measurements matters

This is where the pack's honesty shows. An H3 DiT block's attention isn't one measurement - the FETA route touches a defined set of attention positions per forward, and the pack's contract is "50 measurements on a full forward." The Audit node checks that the count you were promised actually happened. If a wrapper silently skipped the FETA computation, the runtime audit is what catches it. The same discipline applies to the block-cache variant, which audits one measurement on a cache hit versus 50 on a full forward - because a cached forward legitimately does less work, and the audit tells you which world you were in.

How it works

Two inputs: av_latent and runtime (the H3_T8_EAV_RUNTIME from whatever EAV composer you used). Two outputs: av_latent (bit-for-bit your latent, is_output_node: true) and report_json with the forward count, measurement count, observed g range, and chunk workspace. It's a pure verification node - no gain, no modification, no sampling. Think of it as the receipt at the end of an expensive meal: it doesn't make the meal better, it tells you whether you ate what you ordered.

Where it fits

... → EAV composer (model + runtime) → sampler → AVDecode, with the Audit node sitting between sampler output and decode, fed by the composer's runtime. The workflow template in examples wires it exactly that way. Leave it in while you tune tau and the progress windows - it's the only way to know whether your apply_exp run actually applied FETA on the frames you intended.

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.

The trap

Don't treat a clean audit as a quality pass. The pack says it repeatedly: automated metrics prove the trajectory and joint audio changed, not that the picture is better or the sound is non-worse. A green report_json means the mechanism ran; it does not mean the clip improved. And if the audit errors on a consumed token across segments, that's the Long Video case - each segment needs its own fresh runtime, which is exactly what the segment composer exists to provide.