MiniMax H3 HyperFlow Tail New-Noise Sampler (EXP/T8)

The 8+4 route, in one node: run the complete 8-step HyperFlow trajectory at the low resolution, upscale the clean AV latent with the learned 3D upscaler, then do four more evaluations at the final size on a freshly noised tail. MiniMaxH3HyperFlowRefineSamplerT8Advanced is the high-stage half of that.

Twelve network evaluations total, and the author's own description draws the line you need to hear: this is not a seamless 4+4 continuation. The trajectory ended. What the tail stage does is start a new one at the higher resolution.

What it needs

Three inputs:

• model - the HIGH stage MODEL. HyperFlow-patched, and typically its own loader instance so it can carry its own content LoRA or attention patches.
• av_latent - the upscaled clean AV latent, i.e. after the learned 3D latent upscaler.
• hyperflow_plan - and it must be the absolute 4:8 Tail plan. Anything else raises before sampling.

Outputs are model, sampler, sigmas, report_json - the standard trio for a stock SamplerCustomAdvanced, plus a receipt. Internally the sampler is set up as an internal continuation with a new-noise restart, which is why the report names the handoff as explicit_clean_latent_new_noise_dual_av_clocks and the recipe as hyperflow8plus4_new_noise_upscale_exp_v1.

Two practical details from the description: the generic SamplerCustomAdvanced supplies a distinct noise seed for the tail, so the seed control downstream is meaningful and not just cosmetic; and the node declines to inject a continuous x_sigma override, because that would blur the line between this route and a real continuation.

Note the plan node isn't a "High" plan - this is where people trip. The Head plan is 0:4, the Tail plan is 4:8. The 8+4 route uses the Tail plan, because the high stage owns intervals 4:8 either way. The difference between the two tail-side samplers is what fed the upscaler, not which half of the schedule you're sampling.

Route vs route

Concretely, the two high-stage options in this family:

| High-stage node | What it expects upstream | Total evaluations | |---|---|---| | This node | LOW ran the full 8 intervals, then learned 3D upscale | 8 + 4 = 12 | | Partial 4+4 Refine | LOW ran only intervals 0:4 and you took the predicted clean output | 4 + 4 = 8 |

Same handoff shape, different amount of work at the low resolution. Eight evaluations is cheaper; the T8 docs are explicit that neither route's picture quality is certified.

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

Full restart of ComfyUI, then a browser refresh. The pack installs no Python packages at all - requirements.txt is deliberately empty so it can't replace ComfyUI's Torch/CUDA stack - which means there is nothing to pip-install when a node misbehaves. For this route you need the HyperFlow safetensors in models/hyperflow/loras, an unpruned H3 base, the learned 3D upscaler under models/latent_upscale_models, and the usual H3 video/audio VAEs in models/vae.

Where it bites

Predecessor identity matters. Two stages of the same route should come from one full H3 base, cloned for each stage. The pack's long-video and director notes inspect exactly this and reject mismatches, and the reason is that a mixed pair gives you a number nobody can compare to anything.

Upscaled latent, not any latent. If you're unsure whether you're holding the clean upscaled AV latent or a partially-denoised intermediate, the parity audit node is the way to tell - it'll show you the difference between what you have and a reference rather than making you eyeball it.

Twelve evaluations is not a free lunch. Call it what it is: a low-res pass plus a short high-res refine. The pack declines to claim speed or quality superiority for it over the single-pass route, and the runtime table labels the whole 8+4 column "quality not certified".

Seam behaviour doesn't travel. If you've tuned a recipe on the pack's older long-video or Director 4+4 paths, those seam conclusions don't carry over to this route. Re-noise means a new start, and the pack says so in both the node description and the docs.