MiniMax H3 Dynamic Guidance Audit / 动态引导审计 (Advanced)
How many forwards did that guider actually burn?
- av_latent
- runtime
- av_latent
- report_json
The whole point of the dynamic guidance experiment is knowing what you paid for. MiniMaxH3DynamicCFGGuiderT8Advanced hands you a runtime object when it builds its guider; this node, placed after sampling, unpacks that runtime and reports what actually happened - how many guider calls were made, how many physical model forwards ran, and how the cond/uncond branches were batched. Then it hands your latent back untouched. It's the receipt for the guidance experiment, and it exists because "true CFG doubles the forwards" is a claim you should be able to verify, not just believe.
How it works
Two inputs, two outputs. In: av_latent (your sampled latent) and runtime (the H3_T8_DYNAMIC_GUIDANCE_RUNTIME from the guider node). Out: av_latent (identical - the node is explicitly non-mutating, is_output_node: true) and report_json with the counts. The mechanism matters less than the discipline: put this node in the graph when you first enable true_cfg_exp, and leave it there while you tune, so you can see whether the two-branch mode is actually doing two branches or silently collapsing to one.
Why you'd bother
ComfyUI gives you a GUIDER socket and mostly doesn't tell you what ran inside. If you're comparing passthrough vs. single-condition-gain vs. true CFG on a fixed seed - the only honest way to evaluate guidance experiments - you need a ground-truth count of forwards, not vibes. This node is also the pack's fail-closed habit in action: the audit trail is part of the feature, and the README's whole tone is "we verify mechanically before we claim anything."
Where it sits
DualClock Sampler-style setup → Dynamic CFG Guider → sampler → this audit node → AVDecode. Wire runtime from the guider to here, and av_latent from the sampler output to here and onward. If you skip the audit, nothing breaks - but you lose the one piece of evidence that tells you whether your guidance experiment did what the combo box said it would.
Installing it
From MiniMax H3 Audio T8 (T8mars/T8star, GPL-3.0). Manager search MiniMax H3 Audio T8, or:
cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8
restart. No forced pip deps; models to models/ by hand. H3 weights' Community License excludes the US, EU, UK, South Korea.
The trap
If you get a hard error telling you the runtime didn't come from the guider, you've fed it the wrong object - the runtime socket is a typed H3_T8_DYNAMIC_GUIDANCE_RUNTIME, and the node will not accept a report string or a foreign wrapper. And remember it doesn't judge: it counts. Seeing "one branch, N forwards" after you enabled true_cfg_exp is a bug report, not a pass - go check your negative wiring.
Inputs (2)
| Name | Type | Default | Description |
|---|---|---|---|
| av_latent | LATENT | — | |
| runtime | H3_T8_DYNAMIC_GUIDANCE_RUNTIME | — |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| av_latent | LATENT | — |
| report_json | STRING | — |