H3 Progressive · External HIGH EAV Apply (T8 EXP)

The HIGH half of a progressive run is where the detail gets built. It's also where a badly anchored attention effect does the most visible damage, because errors that were invisible at half resolution are suddenly four times as many pixels.

This node attaches the pack's EAV effect to the HIGH stage. The naming is confusing until you see the wiring: it doesn't take the LOW boundary, it takes the high_restart - the typed object produced by HIGH Handoff Only, which already carries the frozen boundary, the restored native AV coordinates and the clean-source masks.

What it does

Inputs: model, eav_config, high_restart. Output: model. Nothing samples, nothing is decoded, and the node's description is explicit about the important detail - it uses the exact HIGH restart plan and the clean-source mask, not the noisy state as its anchor.

That distinction is the whole point. When you're partway into a second pass, the latent you're looking at is noisy; anchoring an effect window to "where we are right now" gives you a window that means something different every step. Anchoring to the clean source means the same window means the same thing on step one and step three.

LOW effects may be absent entirely, or configured differently - separate Stage EAV Config nodes are supported, so you can run a restrained effect window on the small pass and a different one at full size. You can also run HIGH only. Neither is an error.

Same config node as the LOW side (H3 Stage EAV · External Config - the mode/tau/window fields below are its inputs, not this node's), same semantics, same two classic misreadings: disabled is the bypass, and tau = 0 is a gain value, not an off switch. The progress window is still absolute 1 - video_sigma, so it spans the whole trajectory, not just the four or so evaluations this stage owns.

Wiring

high_restart ─→ External HIGH EAV Apply ─→ HIGH Sampler Only
model ────────↗
eav_config ───↗

And, as with LOW: if you're using Prompt Relay, the Relay binding for this phase has to be established before EAV is applied. Optional Relay must be stage-bound first; the two share a network entry point on purpose.

Start on report_only. Always. It executes the same graph and reports what would be applied - absolute sigmas, real effect calls, coverage - while leaving your output untouched. Only then flip the config to apply_exp, and only on a clip you're willing to watch twice.

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

Then fully exit and restart ComfyUI, and refresh the browser tab. The pack's requirements.txt installs nothing by design; optional features check their own dependencies only when reached, which is why a working base install and a working EXP install can have different requirements.

Where people get burned

Wiring this on the LOW side of the graph by mistake. Both apply nodes take a model and return a model, so a mis-wire is electrically valid and produces no error - just an effect applied at the wrong stage. The difference in the input list is the tell: HIGH takes high_restart, LOW takes plan and low_source.

Skipping the audit. HIGH Effects Execution Audit reads the actual immutable completed-HIGH evidence rather than a counter that could be stale, and it does not run LOW or trust an old cached counter. If you can't tell whether your effect fired, this node answers that question in a report you can read - while a visual eyeball at full resolution mostly answers "hmm."

And the standing caveat: EAV here is CFG-1, per-stage, explicitly not qualified for GPU/media quality or persistent effect identity. Raise tau in small steps if you must raise it at all, and check the audio too. On H3, audio is generated jointly, so a change in the video clock can show up in the sound - which is not a bug, just a consequence of not bolting audio on afterwards.