H3 Stage EAV · External Config (T8 EXP)

EAV here means Enhance-A-Video - the paper (arXiv 2502.07508) with the FETA trick, where you re-weight cross-frame temporal attention energy during part of the sampling trajectory to sharpen motion and detail. This pack ships a clean-room adapter of it. MiniMaxH3StageEAVConfigEXPT8 is the settings object: one small node you configure once and wire into the EAV apply node on any stage, or into two different apply nodes with two different configs if you want LOW and HIGH behaving differently.

It holds six numbers, and three of them are the ones that matter.

The six

• mode - report_only (default), apply_exp, or disabled. Read this as a three-position switch with very different meanings. report_only computes and reports what it would do but applies no gain. apply_exp is the real thing. disabled is an exact bypass - it doesn't call the adapter at all.
• tau (default 4, −32 to 32, step 0.25) - the FETA temperature that scales how strongly the enhancement is applied. The author's warning is worth taking literally: tau = 0 is not an off switch. If you want the effect gone, use disabled.
• start_video_progress (default 0.15) and end_video_progress (default 0.90) - the window of the diffusion trajectory where EAV is active, expressed as video sigma progress.
• max_workspace_mib (default 32, 4–512) - a bound on the adapter's working set.
• g_hard_limit (default 1.5, 1–3) - the hard ceiling on gain. If the computed gain would exceed it, the node errors rather than quietly clamping.

Output is a single eav_config socket, typed T8_STAGE_EAV_CONFIG, which goes into MiniMaxH3StageEAVApplyEXPT8.

The one genuine footgun: absolute progress

The description states it as a flat fact: "Progress is absolute 1-video_sigma; it is not reset at the stage boundary."

That sounds like trivia. It isn't. In a split workflow you're running two sampling calls, but the schedule is one timeline - if your LOW stage covers sigmas 1.0→0.4 and your HIGH covers 0.4→0, then an EAV window of 0.15–0.90 means LOW sits inside the window and HIGH sits entirely outside it. If you were imagining a nice symmetric "last 30% of each stage", you'd be wrong, and the last stage would get no enhancement at all. On a cold-resume graph where you load a saved LOW and only run HIGH, that's the difference between an effect and no effect - and the downstream audit is what tells you which you got.

Also note the ordering constraint from the pack's own docs: active stage effects currently want CFG 1. H3 is a guidance-distilled family, so this is usually where you are anyway, but if you've cranked CFG you should expect the effect to behave oddly or refuse.

Install

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

Restart ComfyUI completely and refresh the frontend. Manager users: search MiniMax H3 Audio T8. There's nothing to pip install - the pack's requirements file is deliberately empty so it can't touch ComfyUI's Torch/CUDA stack - but you do need a recent ComfyUI core, because the pack targets native H3 support and the newer node API.

Gotchas

A config node is cheap, so keep a couple: one defaulting to report_only so you can A/B what the effect wants to do, one with your tuned values. If you wire the same config to both stage apply nodes, understand you're giving both stages the same absolute window - which, per above, is often not what you want. And if the audit comes back observed_no_steps_in_effect_window, the config isn't broken, your window just doesn't intersect the stage's real sigmas.