H3 HyperFlow · Separate Full8 / Partial4 / Fresh4 Setup (T8 EXP)

Every two-stage video route in ComfyUI eventually grows the same argument: which stage am I in, which sigmas belong to it, and which output socket is the actual clean picture. This node answers all three for the HyperFlow fresh route, in one place, before any sampling happens.

It takes a HyperFlow-patched MODEL plus an AV latent, you pick the stage, and it hands you back the model, a sampler, sigmas, a typed stage_context and a report. No diffusion runs here - that's the point. Setup is separated from execution so you can see what the stage is going to do.

The stages

Four values, and they're genuinely different recipes rather than labels.

hyperflow_low_full8 runs the full absolute 0:8 at LOW resolution. Because LOW is finished, its completed output is what feeds the learned 3D latent upscaler.

hyperflow_low_partial4 runs only absolute 0:4 - a partial trajectory. Here the socket matters enormously: the predicted clean x0 is denoised_output, and the pack is explicit that the non-terminal output is not the clean picture. Get this wrong and you'll upscale a half-denoised latent and wonder why the whole clip looks like it was shot through a shower curtain.

hyperflow_high_after_full8 and hyperflow_high_after_partial4 are the HIGH halves, both running absolute 4:8 at the upscaled size with a fresh, external noise field and the original audio rebased to the HIGH start. Twelve evaluations total for the full8 route, eight for partial4.

That's the definition of "fresh" here, and it's what separates this from the continuous HyperFlow route, where the second stage inherits raw model-space x_sigma with no new noise and no resize. Both are legitimate; they are not the same graph, and swapping one for the other mid-edit is how you get a clip that looks like two different films stapled together.

Inputs and outputs

model - the output of the dedicated HyperFlow fresh loader, with that stage's content LoRAs already applied. Each stage gets its own loader; models, prompts, noise and effects are independent between stages.

av_latent - the AV latent at this stage's working size. On a HIGH stage that means the upscaled latent, not the LOW one.

stage - the combo above, default hyperflow_low_full8.

Outputs: model, sampler, sigmas, stage_context and report_json. Wire model/sampler/sigmas into your sampler, and stage_context into a Stage Sampler if you want an auditable stage result you can save and resume from. Standard Core sampling works too - the setup just prepares the contract; the actual stage still executes once through Core.

Install

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 fully. You need the H3 base, Qwen3-VL encoder, video and audio VAEs, the learned 3D upscaler in models/latent_upscale_models/, and the original HyperFlow adapter in models/hyperflow/loras/ - no auto-download for that one. The pack's requirements file is deliberately empty so nothing here touches your Torch install.

Where people get burned

Effects. If you've bound external EAV or Prompt Relay on a stage, the pack requires CFG 1 for active effects. Higher CFG with an active effect on a HyperFlow stage is not a supported combination, and it's the kind of thing that produces a subtly wrong clip rather than an error.

Object lifetime. The paired MODEL and CONDITIONING have to stay together - if a Relay conditioning node is involved, its model and positive come as a set, and the pack refuses mismatched pairs. Rebuild both per stage rather than reusing one across LOW and HIGH.

And the scope: this node wires the fresh route, not the continuous split and not P7 long-video. The docs list the qualified evidence as tiny-model CPU runs plus a few GPU probes at 0.4-0.6MP, with a full-pre-trained, full-resolution, human-reviewed pass still outstanding. It's the most interesting H3 scheduling work in the pack right now, but it is explicitly not "download and ship".