H3 Chunked v5 · External Window EAV (T8 EXP)

"EAV" in this pack is the Enhance-A-Video attention trick: a stage-local routing layer that scales attention outputs during a specified window of the sampling progress. It's the kind of thing people bolt on and then argue about for a week, because it changes the picture, is basically free at runtime, and has no quality guarantee whatsoever.

H3 Chunked v5 · External Window EAV binds a separately editable Stage EAV config to one exact joint AV PASS2 window, including the locked video/audio overlap. It's the per-window version of the pack's general Stage EAV family, and it exists because on the v5 route a "stage" is a window, not a whole clip.

Inputs and outputs

You give it model (the raw model you'd have sent to the window), sigmas, partial4_denoised_output, lifted_full_av, prepared, plan, window_index, and an eav_config from the pack's H3 Stage EAV · External Config node. Optional: previous_result to match the window's chain position, and relay_positive for the case where Relay is also in play.

Out come model - the EAV-bound model, which you wire into the matching One Joint AV PASS2 Window node - plus runtime, report_json and stage_context. Wire runtime into H3 Chunked v5 · Window EAV Actual Calls; without that audit you have no idea whether the effect executed.

The config, briefly

The external config node defaults to mode=report_only, which measures the routing opportunity and changes nothing. apply_exp is the one that actually touches the latent. tau defaults to 4.0, the progress window defaults to 0.15–0.90 of video progress, and there's a hard cap on the gain (g_hard_limit, default 1.5). That cap is not decoration: the pack's own validation runs keep reporting maximum gains well under it and treat gains as measurements, not as proof that anything got better. An applied EAV with a maximum gain of 1.0 measured, and changed essentially nothing, is a real outcome you should be prepared for.

Also in the contract: EAV is video self-attention only. Audio is not reworked by it, and EAV only fires on an explicitly wired stage - connecting a wire doesn't certify coverage.

What it is not

• Not default quality. report_only is the default for a reason. Turn on apply_exp when you're testing, one variable at a time, and keep the receipt.
• Not composable with Relay yet. The node description says Relay composition is not qualified. If you need both, expect to be on your own, and don't assume the audit numbers add up.
• Not a whole-clip effect. One window per node instance. Multiple windows means multiple apply nodes, each paired with its own audit.

Install

Manager → search MiniMax H3 Audio T8, or:

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

Fully restart ComfyUI and refresh the browser. Recent Core with native H3 support is required; if Manager's version lags, install from GitHub. The pack ships no weights and no extra Python dependencies for these nodes - the requirements.txt is empty by design so an install can't replace ComfyUI's Torch/CUDA stack. Main model, Qwen text encoder, and both VAEs come from your own downloads plus the author's t8star HuggingFace mirrors.

Where it goes wrong

The apply is invisible and you assume it failed. Check the audit report and the config mode first. A gain of ~1.0 with apply_exp is a legitimate "measurably nothing" result, not a bug.

Runtime node wired to the wrong window. The runtime object is bound to a specific source, lift, plan and window_index. Swap windows in the graph without re-projecting and the audit refuses - which is the correct behaviour and a lot nicer than silently auditing the wrong thing.

Plan/geometry edits mid-graph. The apply re-checks prepared state; if you changed the plan after building it, rebuild forward from Prepare Full AV Noise.

Duplicate pack copies. The usual suspect in this repo: an old test directory in custom_nodes serving stale schemas. Rename with a .disabled suffix (leading underscores don't disable), restart, and verify python_module in /object_info.

And the standing caveat from the README, which applies double to effects: a passing sample doesn't certify your asset, your backend or your seams. Watch the clip, don't just read the report.