H3 HyperFlow · Bind External TAIL Effects (T8 EXP)

The continuous HyperFlow split is the "other" two-pass idea in this pack, and it's worth understanding how it differs from P7 before you plug anything in. P7 stops at step four, upscales, re-noises and finishes at a higher resolution. The continuous split doesn't: HEAD samples the first part of the absolute interval and hands over the raw model-space AV state, TAIL continues from exactly that state with no new noise, no resize, no clock reset. One trajectory, two nodes, two independently patched models.

Bind External TAIL Effects is the TAIL half's binding step. It attaches an independent TAIL model and its LoRAs to the actual HEAD boundary, keeps the paired Relay conditions together, and prepares the optional external Stage EAV wiring.

Inputs

continuous_boundary - the typed boundary from the HEAD-only sampler (or from Load Exact Frozen HEAD if you're resuming). This is a captured model-space x_sigma, not a clean x0 and not a regular latent; the pack is emphatic about that, and about not treating it as one.

model - the TAIL model branch. It can carry its own content LoRAs and its own patches; both halves must come from the same H3 base and the same original HyperFlow adapter, but the content patches are deliberately independent.

positive / negative - the TAIL's conditions. If you're using Relay, these come from the paired Relay conditioning for this half, and the binding keeps them with the MODEL.

Outputs

model, positive, negative go down to the EAV apply (optional) and then the TAIL-only sampler. sigmas and stage_context are what the external Stage EAV node needs. report_json records the binding.

And then the one that catches people: stage_template is a geometry/scaffold latent, not the raw x_sigma. It exists so the EAV path has something to work against. The sampler still takes the real continuous_boundary - the docs are explicit that the template must not be substituted for it, and if you wire the template into the sampler's latent input you get a run that executes happily and produces something that isn't a continuation.

CFG and effects

Active stage effects in this pack want CFG1, both halves. Mixed CFG with an effect applied is not a supported combination, and the reason is fairly obvious once you think about it - the effect path assumes the phase is running without classifier-free guidance pulling in a separate negative branch.

The HEAD/TAIL wiring, in full

HEAD HyperFlow MODEL → optional Relay → HEAD Effects Bind → optional Stage EAV → HEAD Only
HEAD boundary → [HEAD Audit] → [HEAD Save/Load, optional] → TAIL Bind (this node)
TAIL HyperFlow MODEL → optional Relay ─┘
   → optional Stage EAV → TAIL Only → TAIL Audit → Save / decode

You can run only one half if that's what you want - HEAD alone gives you the first part and a boundary you can freeze; the pack supports saving HEAD and having a TAIL-only graph open it later, with the base and adapter contents re-checked at load.

Install

Manager → search 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 and browser refresh. The pack's requirements.txt installs nothing on purpose, so this can't change your Torch/CUDA versions. You need H3 base weights, the Qwen text encoder, both VAEs, and the original non-pruned HyperFlow adapter under models/hyperflow/loras/ - loaded per branch with the pack's dedicated HyperFlow loader. A pruned adapter file is rejected outright, and so is a generic LoRA whose metadata declares itself a HyperFlow artifact.

Reality check

This route's node descriptions carry a blunt limitation: the first version's boundary identity was process-local. That's been extended with explicit save/load and certified HEAD reuse, but the pack still lists broader backend support, full-asset validation and human quality review as pending. Which is a long way of saying: excellent for controlled experiments, not yet the thing you build a production pipeline on top of.