H3 HyperFlow P7 · External Stage EAV for This Phase (T8 EXP)
After setup, and only if you opt in
- model
- sigmas
- av_latent
- stage_context
- eav_config
- prepared_phase
- model
- runtime
- report_json
EAV is the pack's external stage-effect hook - an attention/value-side perturbation applied during sampling rather than a LoRA or a post-process. The generic Stage EAV nodes exist for the plain pipeline; this one is the P7 adapter, and the reason it exists rather than letting you reuse the generic node is that P7 phases carry their own motion-keyframe layout and their own absolute time base, and an effect applied against the wrong layout is worse than no effect.
Position in the graph matters
The pack's wiring rule is simple and you should follow it: Relay before the phase setup, EAV after it. So the apply node sits between LOW 0:4 Setup (or HIGH 4:8 Setup) and your guider, consuming the setup's outputs.
Inputs
model - the patched model from setup. sigmas - the phase's absolute sigma table, passed through from setup; the apply node needs it to know where in the schedule the effect window lands. av_latent - the phase's source latent. stage_context - the typed context the setup produced, which is how the adapter knows it's talking to a P7 phase rather than some other route's stage.
eav_config - the T8_STAGE_EAV_CONFIG, produced by the pack's Stage EAV Config node. This is where report_only lives. That default is a no-op: the config reports what it would do without changing anything. To actually apply a gain you have to select apply_exp, and choosing disabled bypasses entirely with the original model object preserved.
prepared_phase - the LOW or HIGH phase this effect belongs to. The phase binds the original Long Video motion-keyframe layout, so the effect lands on the right rows rather than a generic grid.
Outputs
model continues to the guider and sampler. runtime is a typed handle for the applied effect runtime. report_json is the immediate report.
Where you should actually look to know whether the effect executed is the completed stage result, not this node. The pack is consistent on that point across every route: the stage result records the actual Relay and EAV calls, and an apply node's own counters aren't proof of execution - they're mutable and can be left over from another run. If an effect was bypassed, the honest report says unverified application rather than claiming success.
Practical notes
Active stage effects in this pack assume CFG1. The stock P7 graphs get CFG1 by feeding positive conditioning into the negative socket, so if you're pushing CFG up for a better-looking result, expect the effects path to be unhappy - the docs say so directly rather than letting you discover it.
Effects are per-phase and optional. You can apply EAV to LOW only, HIGH only, or both with different configs, and the same goes for Relay. What you shouldn't do is assume an effect applied on LOW carries into HIGH: they're separate sampling passes with separate patches, prompts and noises, and the pack treats them that way.
Also, EAV does not add diffusion steps. It changes what happens inside the steps you're already running, which means it's cheap in time and not necessarily cheap in subtlety - with the gain applied there's a window and a magnitude, and beyond the configured limit the node errors out rather than clipping silently.
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
Full restart, browser refresh. requirements.txt is empty by design - no extra pip packages, so installing this can't disturb ComfyUI's Torch/CUDA build. The P7 route needs H3 base weights, the Qwen text encoder, video and audio VAEs, the original non-pruned HyperFlow adapter in models/hyperflow/loras/ and a learned 3D latent upscaler in models/latent_upscale_models.
When it goes wrong
If your run errors out at the apply node, check the order first: EAV after setup, not before. If the effect seems to do nothing, check that the config isn't sitting at report_only - that's the default and it's a numerical bypass, so a "no visible change" result is the expected behaviour rather than a failure. And if you're mixing Relay and EAV on the same phase, do Relay first; the pack's ordering isn't arbitrary.
Inputs (6)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| sigmas | SIGMAS | — | |
| av_latent | LATENT | — | |
| stage_context | T8_STAGE_CONTEXT | — | |
| eav_config | T8_STAGE_EAV_CONFIG | — | |
| prepared_phase | T8_HYPERFLOW_P7_PREPARED_PHASE | — |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |
| runtime | T8_STAGE_EAV_RUNTIME | — |
| report_json | STRING | — |