H3 HyperFlow · Continuous TAIL Only (T8 EXP)

What it is

HyperFlow continuous splits one native H3 run into two editable halves at an absolute point in the schedule. The HEAD node runs 0:split and captures the raw model-space joint AV state (x_sigma) as it stood mid-run. This node runs split:8 from that captured state and finishes.

That sounds like a small detail. It's the entire reason the node exists:

• No fresh noise. You're continuing an in-flight trajectory, not starting a second denoising pass. That's what makes it continuous rather than a two-pass recipe.
• No resize, no rerun of HEAD. The boundary is handed over directly.
• seed is a model execution seed, not a noise seed. If you're hunting for a random seed to reroll the look, you're looking at the wrong field - the noise was already consumed upstream.

The point of splitting at all is editability: the tail gets its own MODEL branch (so its own content LoRAs), its own positive and negative conditioning, and the HEAD half stays cached. Change only the tail prompt and only the tail re-runs.

Inputs

• continuous_boundary - the typed state from the HEAD stage. It carries the absolute interval, so this node can read where it had to start.
• model - a clone of the same full H3 base, with the same original HyperFlow adapter loaded through the pack's dedicated HyperFlow loader. The adapter is not a normal content LoRA; wiring it as one is the mistake to avoid. Content LoRAs are separate and can differ per half.
• positive / negative - independent per half.
• seed - default 26092301, 64-bit range.
• cfg - default 1. At CFG 1 the candidate graphs reuse this stage's positive conditioning as the negative input; the moment you raise CFG you owe the graph a real negative.
• reserve_vram_mib - default 1024, floor 512. Checked at start and on every callback; failures propagate as real errors, there's no silent retry or degradation.

Outputs: av_latent (the finished joint AV), report_json (this run's receipt - the actual calls made, not a UI counter) and completed_result, the typed value that the disk-save and audit nodes consume. Wire completed_result to MiniMaxH3HyperFlowTailSaveEXPT8 if you want it on disk, or straight to decode if you just want pixels.

What it is not

Read this twice, because the names in this pack are crowded:

• It is not fresh-noise 8+4, not partial4+4, not P7. Those are separate routes with separate nodes and separate noise semantics.
• There are no denoise masks in the continuous contract. Same-resolution split, no mask.
• There is no persisted-effect, GPU-throughput or media-quality claim. The docs say the quiet part: a passing sample proves the wiring, not your footage.

If you want the second half to be a genuinely fresh 4-step pass over an upscaled preview, you want the HyperFlow fresh family instead. Continuous is for when you want the original trajectory continued.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

ComfyUI Manager: search MiniMax H3 Audio T8. Then fully quit and restart ComfyUI and refresh the page - this pack won't hot-register. Manager can lag GitHub, so clone manually if the version looks old.

Model layout is the standard H3 set: diffusion model in models/diffusion_models, Qwen3-VL text encoder in models/text_encoders, video and audio VAEs in models/vae. The repo contains no weights. requirements.txt deliberately installs nothing so a fresh install can't replace your torch/CUDA build. Start from a bundled 47-hyperflow-continuous-split graph rather than building the wiring from scratch, and swap the placeholder models out before you run anything.

Common issues

Red nodes or missing sockets. Update ComfyUI core, frontend and Manager together, then restart fully. These EXP nodes track native H3 APIs; a half-updated install is the usual cause.

"The two halves don't match." The base model and the HyperFlow adapter loader must be the same for both stages. Different HEAD base, or the adapter loaded as a content LoRA, is exactly the error the boundary check is there to catch.

OOM, or it dies at the tail callback. Raise reserve_vram_mib above the default 1024, and lower resolution/frame count. On a 16 GB card, one H3 job at a time - the README is blunt that concurrent H3 jobs on 16 GB is how you lose an afternoon.

Audio goes wrong when you changed settings. Audio drift in this pack nearly always traces back to sampler, scheduler, step count or the video/audio shifts, not the node you just touched. Don't mix in a generic EMA/Ref2VA/OpenVDN turbo LoRA to "fix" it.