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

BlockCache is the trick that reuses cached blocks across a sequence so a forward doesn't re-run attention everywhere. Slap Enhance-A-Video on top of that naively and you've got a problem: FETA is supposed to measure cross-frame attention, but if the block didn't actually execute, there's nothing to measure. MiniMaxH3EnhanceAVideoBlockCacheComposerT8Advanced is the explicit Stock20 composer that reconciles the two - it runs FETA only in blocks that actually execute, and audits one measurement on a cache hit versus fifty on a full forward. It knows which world it's in, and it tells you.

The contract is strict on purpose

The model input tooltip is a contract, not a suggestion: connect the DualClock MODEL after the separately-installed MiniMax H3 Block Cache (T8) node - the first contract requires a CPU cache and refuses every other wrapper/patch. This composer is a bridge to a plugin you have to install separately, and it only works if the thing it's bridging is exactly the thing it expects. sigmas must be the exact native Stock20 SIGMAS sent to the sampler.

Inputs you'll actually touch

• mode - disabled preserves the BlockCache model exactly (FETA off); report_only audits FETA without gain; apply_exp turns the target-video gain on.
• tau (4) - experimental FETA weight, a candidate not an optimum.
• start_video_progress / end_video_progress - the schedule window FETA acts in (defaults 0→1).
• max_workspace_mib (32) - FETA's score-buffer budget only, not total workflow VRAM.
• g_hard_limit (1.5) - caps observed gain before the audit flags it.

Outputs: model, runtime (feed the EAV Audit node), and report_json.

The one measurement vs fifty thing

This is the honest bit, and the thing to understand before you read any report: with BlockCache, not every forward re-runs all attention. A full forward gives FETA fifty attention measurements; a cache hit gives it one (the one block that actually executed). The composer audits accordingly, and the difference is information, not failure. If your report_json shows single measurements everywhere, you're on cache hits - which is the point of BlockCache, but it also means FETA had very little to work with. Read the audit in that light.

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. And remember the BlockCache plugin itself is a separate install.

The trap

The README is explicit that this composer (and the STG/Long Video composers alongside it) currently pass only low-load deterministic contract, registration, and workflow-import checks - no stress testing, no quality claims. So: don't stack it with other global attention patches, don't expect a speedup guarantee, and don't treat a clean audit as "better video." It's a mechanism, verified to run; the artistic verdict is still out.