MiniMax H3 Audio Refine 4/8-Step Compatibility Route / 音频精修4/8步兼容路由 (T8 Advanced EXP)

This is the node that opens the MiniMax H3 Audio Refine compatibility chain, and its whole job is to declare what actually happened before this node. MiniMax H3 generates video and audio together in one joint AV latent - that's what makes its native audio so good - but when you run a fast recipe the audio can come out flat while the video looks fine. The T8 pack's answer is a post-hoc audio refine: a short extra pass over the audio slice of the finished latent. This Route node is the honesty checkpoint that chain is built around.

Why a "route" at all

You can refine audio after a lot of different first passes: Turbo4, Turbo8, a learned latent two-pass, PDD 8-step, PDD 4+4, EAV, Prompt Relay, and long-video variants. Those are not interchangeable. The route exists so the pack can track which recipe produced your first pass, check that the model you're about to refine with is actually the right one, and not pretend the total step count means the same thing for every route.

The two inputs that matter most:

• generation_profile - the dropdown with all nine recipes. This must match your workflow. The pack is explicit: PDD/EAV are not re-applied on the refine side, and if your profile doesn't match what you ran, the audit can bypass or refuse. Misdeclare this and you're silently comparing apples to oranges.
• declared_first_pass_nfe - 4 or 8. Just what it says: how many NFE your first pass cost. Pair it with the profile and the report can state the total honestly.

You also feed it the audit from the pack's audio refine audit node, the refine_model (the explicit, non-Turbo model you want to run the tail with), and positive conditioning. The design detail worth knowing: the model connected here is explicitly the refine model. This pack never gates on filenames, sizes or disk hashes - a consistent T8 philosophy across all these nodes. That's friendlier than it sounds; it means a renamed or re-quantized checkpoint won't silently refuse to run.

Outputs are refine_model (passed through for the next node), the route object that feeds the Plan node, plus decision and report_json for your text preview. Everything downstream in the chain - the Plan node that builds the fixed four-NFE tail, the Setup node that emits sampler inputs - is downstream of this route object.

Wiring it up and installing

The chain only makes sense attached to a real first pass, so start from the pack's 18-audio-refine example workflows - there are separate ones for Turbo4, Turbo8, learned two-pass, PDD Ref2VA 8, PDD 4+4, EAV, Prompt Relay and long video with Prompt Relay. The README's own rule: run those workflows one at a time, don't queue the whole folder.

Installation is the same for every node in this pack:

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

or just search "MiniMax H3 Audio T8" in ComfyUI Manager and restart. Before anything: update ComfyUI itself - this pack leans on comfy_api.latest, comfy.weight_adapter and native H3 support, and an old ComfyUI will turn the whole pack red with an import error. The requirements.txt is deliberately empty; everything it needs comes from ComfyUI.

The trap people actually hit: picking the default turbo8 profile when they ran something else, or expecting this node to fix audio by itself. It can't and doesn't - it just signs the route. The refine quality still needs your ears on the two outputs the quality gate gives you. A route that says "compatible" is not a claim that the result is better; the pack's own blind tests came out roughly a coin flip. Use it to keep your A/Bs honest, not to declare victory.