MiniMax H3 HyperFlow 8-Step Plan (EXP/T8)

If you've spent time in the distillation end of ComfyUI, you already know the rule: a distilled model was trained on a specific noise schedule, and feeding it a different one hands it noise levels it never saw. HyperFlow is that problem with a twist - two noise times instead of one, and separate clocks for video and audio.

MiniMaxH3HyperFlowPlanT8Advanced is the node that builds the schedule. That's all it does, and it's smaller than you'd expect: one input, model.

Why one input is enough

Because the plan isn't a schedule you authorise, it's a contract you recover. The input model has to be a HyperFlow-patched MODEL - the output of the HyperFlow Loader. The node reads the loaded artifact's identity and constructs the fixed 9-point / 8-forward AV interval plan from it, then hands you a typed hyperflow_plan object.

The description says it plainly: it's "bound to that exact MODEL, not a bare SIGMAS list." That's the whole design. A raw sigma list can be wired into anything, which is exactly how people end up running a two-time adapter on a one-time schedule and blaming the model. This node makes the binding explicit and typed.

Two outputs: hyperflow_plan (feeding the sampler nodes) and report_json, which is where the plan's identity lives - owner, source hash, the raw sigma points, the video and audio sigma segments derived from them, and the absolute interval the plan covers.

What's inside the plan

The adapter's metadata describes rank/alpha 256, gate 0.25, video and audio shifts of 12 and 3, and nine raw sigma points. The plan object expands that into a video segment and an audio segment, each shifted on its own clock, plus a start_interval / stop_interval pair. NFE is simply stop - start, which is how the family stays honest about cost: the full plan is 8 evaluations, and the partial plans are 4.

Those intervals aren't arbitrary cut points. The training grid has specific boundaries - the pack's docs quote video sigma ≈ 0.923077 and audio sigma = 0.75 at the 4:4 boundary - which is why the split routes are typed nodes rather than a "split at step N" slider. If you want a partial plan, use the Head or Tail plan node; the full8 plan node in this article deliberately won't hand a half-plan to a generic restart.

Install

ComfyUI Manager → 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

Restart ComfyUI completely before judging anything, then refresh the browser. The pack's requirements.txt installs no packages on purpose; this node in particular is pure metadata work on an already-loaded model, so it costs nothing to run. Placement rule: the HyperFlow safetensors go in models/hyperflow/loras, and the H3 base stays in models/diffusion_models.

Where it bites

Wrong input, loud failure. Wire a stock H3 MODEL here and you don't get an 8-step plan for a normal model - the plan building is tied to the HyperFlow artifact's metadata, and the sampling nodes downstream validate the binding. Build order in the graph is Loader → Plan → Sampler, every time.

One plan, one model. Because the plan carries the source hash, you can't build a plan off loader A and sample loader B. That's deliberate: it's the check that catches a copy-and-paste graph where you swapped one of two model branches.

This is not a shortcut to fewer steps. HyperFlow isn't a "6 steps is fine" slider; the plan is a fixed trajectory of 8 evaluations with trained boundaries. Running it at some other length isn't a tuning option - it's a different experiment that nobody qualified.

Read the report at least once. The report is the fastest way to confirm what actually got installed, and it's the artifact the pack's parity-audit node compares against when you want proof that a split route matched the single-pass one.