MiniMax H3 HyperFlow Partial 4+4 Refine (EXP/T8)

MiniMaxH3HyperFlowPartialRefineSamplerT8Advanced is the high stage of the cheapest complete HyperFlow upscale route: low-resolution coarse sampling over intervals 0:4 only, take the predicted clean output, learned 3D upscale, then this node re-noises and finishes intervals 4:8. Eight evaluations for the whole two-stage job - the same count as the full single-pass 8-step run, but spent at two different resolutions.

If you're comparing this against the pack's other tail-side sampler, the distinction is upstream, not here: this one is for when LOW ran four intervals and you took the partial-x0 output. That other one is for when LOW ran all eight. Both hand the high stage the same 4:8 tail, and both need the Tail plan, never the Head plan.

Under the hood

It's literally a subclass of the 8+4 refine sampler - same inputs, same output sockets, changed node id, display name and description. So the behaviour is identical: internal continuation with a new-noise restart, a distinct noise seed from the generic SamplerCustomAdvanced, and an absolute 4:8 plan requirement that raises if it isn't satisfied.

What differs is the reported recipe line. This node stamps hyperflow4plus4_partial_x0_upscale_exp_v1 and a handoff of low_partial_x0_learned_3d_high_new_noise_dual_av_clocks, versus the other node's low_denoised_output_learned_3d_high_new_noise. If you're debugging a graph you didn't build, the recipe field in the report JSON is how you tell these two apart - they're visually identical on the canvas.

Inputs: model (the HIGH stage MODEL), av_latent (the upscaled clean latent), hyperflow_plan (4:8 Tail). Outputs: model, sampler, sigmas, report_json, feeding a stock SamplerCustomAdvanced.

Why you'd pick it

Eight evaluations instead of twelve is a real difference if you're iterating on a prompt. The trade is that the low-resolution half of the trajectory gets only four evaluations with a partial x0 prediction feeding the upscaler, and the pack is blunt that neither quality figure is certified - the formal evidence here is execution-level: 832×480 and 544×960 outputs that decoded completely, with two 4-NFE stages and full H.264/AAC delivery.

The critical wiring rule for the low side, which isn't this node's fault but always gets blamed on it: the coarse sampler's denoised_output (socket 1), not socket 0, is what goes into the learned 3D upscaler. Socket 0 is the nonterminal x_sigma result. The pack's docs note that the HIGH plan check can't detect a wrong LOW socket, so a wrong connection produces no error at all.

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

Fully quit and restart ComfyUI, refresh the page. No package installs: the pack's requirements.txt is intentionally empty so nothing can overwrite ComfyUI's Torch stack, and every optional feature imports its own dependency only when used. Required files for this route: HyperFlow safetensors in models/hyperflow/loras, unpruned H3 base in models/diffusion_models, learned 3D upscaler in models/latent_upscale_models, H3 VAEs in models/vae.

Where it bites

Tail plan, not Head. Same failure as its sibling: hand it a 0:4 plan and it raises. That's the good outcome; the bad outcome is the socket mistake one step earlier.

One base, two stages. Clone one full H3 base for both model slots. Mixing bases across stages is not a qualified configuration, and receipts across the pack's routes hash the model identity precisely so a mismatch is visible.

Don't expect the low-res half to be free of structure decisions. Four evaluations at a fraction of the resolution is doing a lot of work; if composition or motion is wrong, going to the 8+4 route gives the low half more room before the resize.

Quality is unverified, and labelled so. The report literally says unverified_experimental. That label isn't boilerplate - it's the pack telling you the picture passed mechanical decode, not a human review of whether it looks as crisp as the single-pass route.