MiniMax H3 HyperFlow Tail 4 Plan (EXP/T8)

MiniMaxH3HyperFlowTailPlanT8Advanced is the mirror of the Head plan: it recovers the absolute 4:8 intervals of the HyperFlow trajectory. The description is short and does not oversell - "for explicit new-noise 8+4 upscale experiment only."

That phrase, new-noise, is the entire character of this half. The tail plan doesn't continue a trajectory where the head left off; it re-noises the AV latents and runs the final four evaluations at the higher resolution. Which is why the node's own docs warn against calling the result a seamless continuation.

What it does mechanically

One input, model - a HyperFlow-patched MODEL, same as the Head plan - and two outputs, hyperflow_plan and report_json. It reads the artifact's trained grid, takes the 4:8 window, and produces a typed plan with start_interval = 4, stop_interval = 8, so NFE is 4. Both the video and audio sigma segments come from the artifact's own metadata, including the separate shifts, so the tail plan knows the dual clock the same way the full plan does.

The node that consumes it enforces the pairing: the tail-side samplers raise if they aren't handed an absolute 4:8 plan. Mixing Head and Tail outputs is a canvas-level mistake that fails loudly, which is the right design - the two outputs have identical types and identical-looking wires.

Where it fits

Two routes use it, and they differ in what happened before:

• Partial 4+4 upscale - Head plan drives a low-resolution 0:4 coarse stage, socket 1 of that sampler goes to the learned 3D latent upscaler, and this tail plan finishes the job at high resolution. Eight evaluations total, four per stage.
• 8+4 refine - a full 8-step pass at low resolution, learned 3D upscale, then this tail plan's four evaluations at the final size. Twelve evaluations, and again not a continuation.

In both cases the high stage gets fresh noise. The refine-side sampler requires a distinct noise seed supplied by the generic SamplerCustomAdvanced, so don't expect re-running give you the same tail twice unless you hold the seed.

The receipts are explicit about this being a quality-uncertified experiment: the report carries quality_status: "unverified_experimental" and names the recipe. The pack's evidence for the 8+4 route is execution-level - a 256²→512² probe completed the whole chain and delivered H.264 with 32 kHz AAC - which proves it runs and produces a valid file, not that it looks better than the single-pass route.

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

Quit ComfyUI entirely and start it again, then refresh the browser - loading new node classes doesn't happen on a page refresh. The pack ships an empty-by-design requirements.txt, so nothing here will replace your Torch/CUDA stack. Assets for this route: HyperFlow safetensors in models/hyperflow/loras, unpruned H3 base in models/diffusion_models, learned 3D latent upscaler in models/latent_upscale_models, VAEs in models/vae.

Where it bites

Plan/model mismatch. The plan is bound to the model it was built from. Cloning the base into two stages is expected; swapping in a different base between stages is not, and the binding check is there to catch it.

Seed discipline. The high stage is a new-noise restart, so the seed you set on the generic sampler is doing real work. A/B tests that change seed and resolution produce differences you can't attribute.

Do not read "8+4" as "12 steps is better". Both pack routes are recorded as unverified for quality. The measured, checkable claim in this family is the parity one - that the continuous split matched single-pass latents exactly - and it belongs to a different node.

Keep sizes legal. Learned 3D upscale wants a clean 2x on the H3 32-pixel grid. "Half, roughly" is how you get soft output and blame the adapter.