H3 VDN · External Prompt Relay Apply (T8 EXP)

OpenVDN's distilled H3 branch doesn't run attention the way the dense base model does - it's a window-plus-linear-scan path, the cheap-attention shape that makes a DMD8 route fit on modest hardware. That's great for memory and terrible for a prompt that wants to talk about time ("then the camera pulls back"), because the original path has no explicit sense of when.

MiniMaxH3VDNRelayApplyEXPT8 adds one. It's the VDN-specific complement to the pack's Prompt Relay system, and it's separate from Stage EAV on purpose: Relay is a prompt-and-timing effect, EAV is an attention-energy effect, and the pack tracks them with separate counters so you can tell which one you actually applied.

What it changes

Two things, per the author's description, both inside the existing path:

• Real temporal bias on the original VDN window keys. Not a re-implementation, not a softmax replacement - a bias term added to the window key sets the branch already uses.
• Beta-weighted nonlinear text seeds into the original linear scan. The relay's prompt events seed the linear attention scan with nonlinear weighting, which is how a time-tagged prompt reaches a path that otherwise treats text as one blob.

Neither adds diffusion NFE. You're not paying steps for this, you're paying a little workspace, which is what max_workspace_mib bounds.

Inputs and outputs

• model - the VDN stage model, paired with its authenticated Relay conditioning. The documented wiring is Relay Conditioning → VDN stage → this node → optionally Stage EAV.
• sigmas, av_latent - this stage's, same as you'd give the EAV apply node. Mismatched stage objects are the usual cause of a silent no-op.
• stage_context - from MiniMaxH3VDNStageSetupEXPT8 for the same stage.
• mode - apply_exp (default) or disabled. Note that unlike EAV, the default here is the active one, and there's no report_only middle position. If you want to compare, you set disabled and re-run.
• max_workspace_mib (default 64, 4–1024).

Outputs: model (into the sampler), runtime (a T8_VDN_RELAY_RUNTIME - hand it to MiniMaxH3VDNRelayAuditEXPT8 after sampling), and report_json. The report records the binding: the stage context, the block count and a binding hash, so a later audit can confirm the runtime belongs to the model it ran on.

The honest disclaimer, which is in the node itself

The author's description says this is an "explicit experimental VDN extension, not paper-softmax or trained quality equivalence." The boundary text in the runtime is even more specific: the window key bias and temporal logit bias are the real thing on the original path, while the linear beta-weighted text seeding is an experimental extension - and the workspace number is an adapter working-set estimate, not your total VRAM.

Two practical consequences. Unsupported kernels fail normally rather than falling back to a dense path - good, because a silent fallback would change your memory profile and your speed entirely. And if you're on the report_only-happy part of this family and wondering where that went: it doesn't exist here. disabled or apply_exp, that's it.

Install

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

Exit ComfyUI completely, restart, then refresh the browser. Manager search: MiniMax H3 Audio T8; install from GitHub if the registry listing lags. There's no pip install step - the pack's requirements file is empty by design so it can't replace ComfyUI's Torch/CUDA build - but recent core is required, because the pack builds on native H3 support and the newer node API.

You'll also need the OpenVDN bundle (t8star/Vdn-Minimax-H3-Comfy) in place; a plain H3 model won't satisfy the VDN stage's own checks.

Gotchas

This node is not effect-execution proof - the docs are explicit that you should inspect the downstream audit. A runtime that was never consumed by a sampler will happily report its configuration.

The pairing rule is the one that trips people: Relay conditioning and the VDN stage model go together, and the sigmas and AV latent have to be the same stage's. Wire a well-formed runtime to the wrong stage and you'll get bias applied at the wrong sigma range, which shows up as a clip that gets sharper somewhere it shouldn't and not where you wanted.