H3 Continuation · External LOW EAV (T8 EXP)

"EAV" in this pack means an enhancement applied at attention level within a progress window - a measured, bounded gain on the video stream, not a slider that makes your video better. The Config node decides whether anything happens at all; this node is just the place it gets attached, and it is scoped so tightly that it only ever touches the low-resolution phase of one continuation window.

That scoping is the interesting part. Applying a detail gain on the cheap pass and not on the final pass is a legitimate move - you are sharpening the structure the upscaler will see, before a full-resolution pass can turn an over-applied gain into flicker and identity drift.

What it does

It attaches a stage EAV config to the LOW model, preserving the accepted parent's native motion guides and masks and the global 1−video-sigma clock. Then the LOW sampler runs with a model that has the effect available inside the configured window.

Three details from the pack's own documentation are worth internalising:

• The window is measured as 1 − video_sigma on the global progress clock. It is not per-stage progress, and HIGH does not reset the clock. So "15%–90%" means 15%–90% of the whole window's schedule, not of the low pass.
• The config's mode decides everything: disabled is a bypass, report_only measures without applying, apply_exp applies. If nothing changed in your output, the first thing to check is the config node, not this one.
• disabled returns the input MODEL object unchanged. That matters if you are debugging by elimination: no silent replacement, no identity mismatch to chase.

Inputs and output

• model - the phase's patched model, after Relay if you use Relay. Relay first, EAV second; the pack refuses a stacked owner.
• prepared_phase - and it must be the LOW phase. Hand it a HIGH phase and you get a flat ValueError telling you exactly that. There is a separate HIGH EAV node for the other half, which this one deliberately does not replace.
• plan - the shared plan object; the phase and plan are cross-checked so the effect is bound to the schedule you think it is.
• eav_config - the stage EAV config from the pack's EAV Config node. Wire an actual config object; this node does not invent defaults.

It outputs a single model. That one output goes into the LOW sampler's model input. There is no conditioning here, no latent, no sampling - deliberately.

Wiring

Conditions (low) → Relay Apply (if used) → LOW EAV Apply → LOW sampler.

The EAV gain is constrained: there is a hard limit on the configured gain and the node errors rather than quietly clamping past it, and report_only does not disable the resource and numeric protections either - it just declines to apply the gain. The author's own tested candidate used a tau of 8 over roughly 15%–90% of progress and described it as restrained; the node default is more conservative still. Raising it is a one-step-at-a-time thing, because higher gains are documented as more likely to amplify identity drift, local warping, flicker, and change through into the joint audio.

Install

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

Then fully restart ComfyUI - an EXP class appearing on the next page refresh without a restart is not a thing - or install through ComfyUI Manager by searching "MiniMax H3 Audio T8". No pip packages are installed by design; optional features pull their own dependencies only when used. Recent Core with native H3 support plus the H3 checkpoint, Qwen encoder, video VAE and audio VAE remain prerequisites.

Things that will bite you

The node being present in your workflow proves nothing about whether the effect ran. The pack's own evidence notes that a run can show one active forward with a gain of 1 - real execution, no measurable benefit. Read the audit: the common Progressive Effects Audit nodes record the actual calls, and that record, not the Apply node, is what tells you the effect was live during sampling.

Also expect active stage effects to want CFG1, and be prepared for a refusal if your model already carries an owner. This is the least glamorous node in the pack and the most likely to be wired in the wrong order.