H3 HyperFlow · Bind External HEAD Effects (T8 EXP)

The continuous HyperFlow route runs one trajectory cut at split_interval. That's great for coherence and slightly awkward for effects: the stage-local EAV and Prompt Relay machinery in this pack wants to know which stage it's attaching to, and "the first half of a continuous run" isn't a stage in the usual sense.

This node is the adapter. It binds external HEAD effects to the continuous contract, handing your model and conditioning back with the right patch attached and the right scaffold prepared. It does not sample - the HEAD stage still does that.

What it hands back

Seven outputs, and they're not decoration: model, positive, negative, stage_template, sigmas, stage_context, report_json.

model and positive (and negative if you're using a real one) go into the HEAD sampler. The pack requires the MODEL and its CONDITIONING to stay paired - if a Relay conditioning node produced them, they come as a set, and a mismatched pair is a refusal rather than a warning. That's the single most common wiring error with Relay in this pack.

stage_template is geometry/scaffold only. It is not a raw x_sigma boundary and cannot be used in place of the real continuous_boundary for a TAIL stage. The docs call that out specifically because the sockets look adjacent and someone will try it.

stage_context is what the stage EAV Apply node wants. sigmas here are for effect validation, not for feeding another sampler or building a Progressive plan - a small but sharp distinction, since handing these to a sampler produces something that runs and is wrong.

Wiring

Dedicated HyperFlow loader → optional paired Prompt Relay conditioning → this node → optional external Stage EAV Apply → the HEAD Only stage node. Same split_interval on this binding and on the HEAD stage - if they disagree, the effect is bound to a different cut than the one you sample.

Then, if you care whether the effect actually ran (you should), pass the HEAD boundary through the HEAD effects audit on its way to the TAIL stage or a save node.

CFG, and where effects go to die

Active effects on this route require CFG 1. That's not a style preference - the underlying grid is trained for guidance-free sampling, and the pack won't pretend a higher CFG with an active effect is a supported configuration. You can change CFG if you supply a genuine negative conditioning, but then you're outside the qualified envelope, which the pack will tell you in its report rather than hiding.

Effects also default to report_only in the pack's EAV config: the audit observes, the numbers get recorded, and no gain is applied. Applying a gain is a separate explicit choice. If your clips look identical with and without effects bound, check that first - it's a feature, not a bug.

Install

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 afterwards. You need the H3 base model, Qwen3-VL text encoder, both VAEs, and the original non-pruned HyperFlow adapter in models/hyperflow/loras/ - no automatic download, and a pruned base is rejected rather than approximated. The pack ships no pip dependencies on purpose, so updates can't disturb your Torch stack.

Where people get burned

The order. Relay before EAV, always. The docs are explicit that when both effects are in play you apply Relay first and the EAV wrapper second, and that the new EAV entry composes the stage's Relay wrapper into a single runtime. Reverse them and you're asking two wrappers to own the same hook.

Unpaired sockets. Keep the MODEL and CONDITIONING together, and remember that the TAIL stage can be configured with entirely different effects from HEAD - the two halves can use different content LoRAs, prompts, Relay plans and EAV configs, or only one of them can have effects at all. That flexibility is the point, and it's also more places to wire something wrong. Audit the HEAD half before you trust the TAIL.