Nodes/comfyui-minimax-h3-audio-T8/H3 Progressive · External LOW EAV Apply (T8 EXP)
ComfyUI Node

H3 Progressive · External LOW EAV Apply (T8 EXP)

Effects on the First Pass Only

By T8mars·Created 2 months ago·Updated about 7 hours ago· 1,158
H3 Progressive · External LOW EAV Apply (T8 EXP)
  • model
  • eav_config
  • plan
  • low_source
  • model

EAV in this pack is a bounded gain applied to H3's native temporal attention, observed at the post-norm/RoPE Q/K stage, with the original masks, projections and gates left intact. People call it "attention effects" and that's close enough: it changes how much temporal context the model weighs at each step. It is not a colour grade and it is not a quality slider.

The reason it's split into an external config node and per-stage apply nodes is that the two halves of a progressive run want different treatment. This node attaches EAV to LOW and only LOW.

What it does

Small node, short list. In: a model, an eav_config (from the pack's Stage EAV Config node), the progressive plan, and the low_source latent. Out: a model. That's it - one input removed, one model back.

No sampling. No model call. What it does is bind the effect to the stage's actual geometry and clock before the sampler runs:

  • It binds LOW's native mask and its full absolute video clock. Progress is measured as 1 - video_sigma over the whole run, not as local step progress inside this stage. So an EAV window of "15% to 90% of video progress" means the same thing on LOW as it would on a single-pass run - it isn't silently renormalised to the first four steps.
  • It doesn't touch the sampler's schedule, the noise generator, or the audio delivery.

The eav_config itself is a separate, reusable node you'll see across this family (its display name is H3 Stage EAV · External Config; that's where the settings below live, not on this node) with its own mode (disabled / report_only / apply_exp), tau gain, a start/end progress window, a workspace budget and a hard gain limit. Two things from that node's documentation are worth repeating because they trip people up constantly: disabled is an exact bypass, and tau = 0 is not an off switch - it's a gain parameter, so turning it to zero doesn't mean "no effect."

Wiring

Stage EAV Config ─────────→ External LOW EAV Apply ─→ LOW Sampler Only
plan, low_source ─────────↗        (model in, model out)

The order matters if you're also using Prompt Relay: bind Relay to the stage first, then apply EAV. The two are designed to share the network entry point, and the LOW apply node expects the Relay binding to already be there when relevant.

Set the config to report_only on your first run. It reports what would have been applied - actual calls, absolute sigmas, coverage - without changing the output. The pack's examples ship with report_only as the default for exactly this reason, and any tutorial telling you to jump straight to apply_exp is skipping the only cheap sanity check you get.

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

Restart ComfyUI completely and refresh the page. Nothing is pip-installed by the pack - requirements.txt is comments explaining that the author won't risk your Torch/CUDA install - so the usual failure is a process that wasn't actually restarted.

Where people get burned

Assuming effects fired because the node is in the graph. This is the whole reason the pack ships LOW Effects Execution Audit as a separate node: it reads the immutable completion evidence from the stage rather than a mutable counter on the apply node, and missing effects or incomplete coverage are never counted as verified. Wire the audit. Read the report.

Cranking tau for more punch. The author's own guidance is the opposite of the instinct: start restrained and increase in small increments, watching for identity drift, local warping, flicker and - because H3 samples audio jointly - indirect changes in the sound. The hard gain limit stays where it is.

And don't expect this to be a general-purpose enhancement. CFG 1, no learned-network qualification, no persistent effect identity: this is an experimental, per-stage intervention. If you can't hear or see the difference on your own clip at apply_exp, the honest conclusion is that this effect doesn't do much for your material, not that you need to push the gain until you can.

CategoryT8/MiniMax H3/Modular Sampling/Progressive Experimental

Inputs (4)

NameTypeDefaultDescription
modelMODEL—
eav_configT8_STAGE_EAV_CONFIG—
planT8_PROGRESSIVE_STAGE_PLAN—
low_sourceLATENT—

Outputs (1)

NameTypeDescription
modelMODEL—