MiniMax H3 HyperFlow Continuous Split (EXP/T8)

This is the cleverest node in the HyperFlow family and the one most likely to be misunderstood. MiniMaxH3HyperFlowSplitT8Advanced runs one trained 8-step HyperFlow trajectory across two independently-patched models - 4 and 4, or 1 and 7, or any split from 1 to 7 - with an exact model-space x_sigma handoff and no new noise.

Say it in plain terms: it slices the sampler, not the video. Same resolution on both halves, no upscale, no re-noising, no seam-by-design. The node's own description draws the boundaries: this is not the legacy Director 4+4, and it is not a low-to-high upscale.

Why anyone wants that

Because patches are stage-scoped. If you want different content LoRAs or a different attention configuration driving the first half versus the second, the naive approach is to run two passes and glue them - which is where seams come from. Here there's exactly one latent, one schedule, one seed, and the only thing that changes at the boundary is which patched model is being evaluated.

The handoff is the whole trick: the exact captured model-space x_sigma at the split point is passed through untouched. No add-noise step, no rescale. That's what makes it a continuous trajectory rather than a two-pass generation.

Inputs and outputs

• model_low, model_high - two HyperFlow-patched MODELs. The docs are firm here: clone one common full H3 base for both stages. Two unrelated bases will produce something, but not the thing this node is for.
• positive, and optionally negative
• av_latent - native packed H3 AV latent
• seed (control-after-generate is on)
• split_interval - 1 to 7, default 4. This is the absolute evaluation index where the high model takes over, not a frame count.
• cfg - default 1.0

Outputs are av_latent and report_json. Straight into your normal decode.

The claim, and how to check it

The pack measured the interesting thing: final video and audio latents from the continuous split versus the single-pass 8-step run differed by max/mean/RMS absolute difference of 0 on the tested T2VA case, and that held for both a 1+7 and a 7+1 split across two independent loader owners.

That's a strong claim and the pack ships a way to audit it locally rather than asking you to take its word: MiniMaxH3HyperFlowLatentParityT8Advanced takes both latents and produces a read-only difference report. Use it. If your split disagrees with a single-pass run, that's worth knowing before you build a pipeline around it.

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, then browser refresh. requirements.txt installs nothing - by design, so the pack can't swap out ComfyUI's Torch. Beyond the usual H3 models and the HyperFlow safetensors in models/hyperflow/loras, the practical requirement is enough VRAM for two patched models in one graph, or a Core configured to page them.

Where it bites

The two-owner crash is real. The pack's notes record a failure on the first attempted two-loader split inside the native graph compiler on Windows (malloc_graph_pop / aimdo memory compile error), and that the identical graph then succeeded on an isolated Core started with --disable-comfy-compiler. If you hit a native compiler crash on this node specifically, try that flag before you go bug-hunting.

Same base, or don't bother. Cloning one full H3 base twice is the instruction. Mixing a pruned and a non-pruned base across the stages is not the experiment.

Don't confuse it with the Director's 4+4. The old standard_4plus4_v1 is a different graph with a different scheduler and different boundaries. The split boundary here is the trained grid (video sigma ≈ 0.923077, audio sigma 0.75 at the 4:4 point), and the pack records receipts for exactly those intervals - not for an arbitrary "step 3.5".

Zero parity is file-level proof, not taste. Identical latents to a single-pass run means the split is faithful. It does not mean either is good-looking. Judge the picture separately.