H3 HyperFlow P7 · HIGH 4:8 Setup Only (T8 EXP)

Setup nodes feel like nothing when you first drop them in - they don't generate anything, they just open sockets. Then you realise what they buy you: because the sampling contract is built here rather than inside a monolith, you can slip an effect node between the setup and the sampler without invalidating anything, and the HIGH phase gets its own model and its own patches independently of LOW.

H3 HyperFlow P7 · HIGH 4:8 Setup Only patches the HIGH model you hand it and builds the original P7 HIGH stage: the absolute 4:8 interval, the sampler, the sigma schedule, and the source latent for the phase. No diffusion happens here.

Inputs

high_handoff - the typed reconcile object from Original HIGH Audio Reconcile. It carries both the enlarged picture (from the learned lift) and the audio policy, plus the identity of everything upstream. This is not a LATENT socket; wire it from the handoff node.

model - the HIGH model. In this route you typically run a second branch off the same H3 base with its own content LoRA, which is exactly what P7 wants: LOW and HIGH are allowed to differ. The patch belongs to this MODEL object only.

Outputs

model (patched, goes on to the guider and sampler, or through an EAV apply first), sampler and sigmas (the absolute 4:8 table - the stage is not restarting the clock at zero, which is the thing that makes HIGH continuous with LOW), source_av (the latent the sampler starts from), stage_context (typed, needed by the optional Stage EAV apply node), positive / negative (carried through from the handoff so the Relay pairing isn't broken), and report_json.

Typical wiring

Original HIGH Audio Reconcile → HIGH 4:8 Setup
    ├→ model → [External Stage EAV for This Phase, optional] → BasicGuider
    ├→ source_av + sigmas + sampler → ONE Stage Sampler
    └→ stage_context → EAV apply (optional)
ONE Stage Sampler → Bind Completed HIGH Result

Two habits worth forming. First, put the effect nodes after setup if you want them: the pack's own guidance is Relay before Setup and EAV after, and the stage result then records actual calls - an Apply node sitting in the graph proves nothing on its own, which the audits exist to catch. Second, keep the paired MODEL and CONDITIONING objects together; the Relay pairing is carried by both and separating them is how you get a stage that silently ignores your routing.

Install

ComfyUI Manager → search MiniMax H3 Audio T8, or clone into your custom_nodes folder:

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

Full restart and a browser refresh, as with any node pack of this size. The requirements.txt is empty on purpose - the base nodes use only what ComfyUI already ships, so installing can't replace your Torch/CUDA. You'll still need H3 base weights, the Qwen text encoder, video and audio VAEs, the original non-pruned HyperFlow adapter in models/hyperflow/loras/ (you cannot load it through a generic LoRA loader; the dedicated loader rejects a pruned file and refuses a preceding generic LoRA that declares itself a HyperFlow artifact), and the learned 3D latent upscaler.

When it misbehaves

The most common setup failure is scale mismatch: if the handoff was reconciled against HIGH geometry that doesn't match what you configured in the segment selector, the contract check trips before anything samples. The second is trying to run HIGH with active Relay or EAV effects at CFG above 1 - active stage effects in this pack assume CFG1, so if you're chasing a bigger CFG you need real negative conditions and a workflow that isn't the stock one.

GPU-wise, this is the node after which your VRAM usage becomes real. LOW ran at a fraction of the resolution; HIGH runs the full target, over a latent that already contains audio, with a second model resident. If you're borderline, that's where you'll find out.