MiniMax H3 Audio Refine Dual-Model Setup / 音频精修双模型装配 (T8 Advanced EXP)
The strict sibling of the dual-clock setup
- plan
- refine_model
- positive
- av_latent
- model
- noise
- guider
- sampler
- sigmas
- latent
- report_json
In the dual-model Audio Refine branch, this node is where everything comes together - and where the chain gets strictest. The single-model Dual-Clock Setup assembles sampler parts from a plan; this node does the same job for the dual-model path, but before it assembles a thing, it revalidates the entire signed chain: the audit, the route, the Phase 2 plan, the refine model, the conditioning, the latent, and the resource gates. Only with all of that intact does it emit the exact, uncached dual-clock partial tail for SamplerCustomAdvanced.
"Uncached" is worth pausing on. A dual-model tail is the kind of thing a cache could quietly poison - you re-run with a different model pairing and the sampler happily reuses a cached tail from the wrong one. This node deliberately doesn't, so every run starts from the state it actually validated. That's also why the revalidation exists: in a chain this opinionated, the failure mode people actually hit is feeding it a plan from a different chain (or a stale route) and getting a tail that doesn't match the artifacts they connected.
What goes in, what comes out
- plan (
H3_T8_AUDIO_REFINE_PHASE2_PLAN), refine_model, positive (CONDITIONING), av_latent - Outputs: model, noise, guider, sampler, sigmas, latent, and report_json - the five sampler parts in the same positions as the single-model setup, wired into
SamplerCustomAdvanced
If you've run the single-model chain, the shape is familiar: deterministic noise, a BasicGuider-equivalent guider, the T8 dual-clock sampler, the exact tail sigmas, and a latent carrying the audio-only nested mask (video 0, audio 1).
Installing it
Part of the T8mars/comfyui-minimax-h3-audio-T8 pack - search "MiniMax H3 Audio T8" in ComfyUI Manager, install, restart, or:
cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8
Update ComfyUI itself first (recent core: comfy_api.latest, comfy.weight_adapter, comfy.patcher_extension, comfy.ldm.minimax); requirements.txt is intentionally empty. As with every EXP node, start from the bundled workflow.
Where people get burned
Two failure shapes, both about mixing chains. First: wiring a single-model H3_T8_AUDIO_REFINE_PLAN into this node instead of the Phase 2 plan - the schema types are distinct for exactly this reason, and the revalidation should refuse it. Second: changing the refine_model between the Route and this node and expecting it to just work; the signed route is bound to the model it validated, and swapping models means re-running the route. If the assembly refuses, read report_json rather than fighting it - in this family, a refusal is almost always a correct refusal. And remember the standing pack rule: this node's sampler is the owner of the dual-clock path; don't layer another sampler or attention-owner node on top of it.
Inputs (4)
| Name | Type | Default | Description |
|---|---|---|---|
| plan | H3_T8_AUDIO_REFINE_PHASE2_PLAN | — | |
| refine_model | MODEL | — | |
| positive | CONDITIONING | — | |
| av_latent | LATENT | — |
Outputs (7)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |
| noise | NOISE | — |
| guider | GUIDER | — |
| sampler | SAMPLER | — |
| sigmas | SIGMAS | — |
| latent | LATENT | — |
| report_json | STRING | — |