Nodes/comfyui-minimax-h3-audio-T8/H3 Chunked · Bind External PASS2 Relay (T8 EXP)
ComfyUI Node

H3 Chunked · Bind External PASS2 Relay (T8 EXP)

Pairing An External Prompt Relay With One Full-Clip H3 PASS2

By T8mars·Created 2 months ago·Updated about 7 hours ago· 1,158
H3 Chunked · Bind External PASS2 Relay (T8 EXP)
  • relay_model
  • relay_positive
  • relay_av_latent
  • prompt_relay_plan
  • source_segment
  • lifted_segment
  • segment_spec
  • pass2_context
  • plan
  • sigmas
  • model
  • positive
  • runtime
  • report_json

Prompt Relay is the pack's timeline-routing tool: instead of one global prompt for a whole clip, you declare local events at absolute time positions and the relay conditions steer content attention to the right thing at the right second. Powerful, fiddly, and - in a chunked two-pass graph - easy to bind to the wrong geometry. This node is the adapter that makes the pairing checkable.

What it pairs

Ten inputs, and they come in groups. relay_model, relay_positive, relay_av_latent and prompt_relay_plan are the external Relay stack you already built - the plan, its conditioning, the model carrying it, and the AV template. source_segment, lifted_segment, segment_spec, pass2_context and plan describe the segment you're attaching it to. sigmas comes along too, because the binding is geometry-aware and the schedule is part of that.

Outputs: model, positive, runtime and report_json. Wire model and positive into this segment's PASS2 sampling, and send runtime to MiniMaxH3ChunkedPass2RelayAuditEXPT8. The report is where the validation outcome lives.

Note that it validates rather than trusts: it checks the native packed layout, the model-condition hash, and the lifted target geometry. This is the same posture as the rest of the pack - a Relay plan written for one target size can't be silently applied to a differently-lifted segment, because the geometry is re-derived and compared.

The scope line you need to read twice

From the author's description: this is for one v2–v4 full-clip/full-frame PASS2. V1 temporal chunks and spatial multi-tile need separate local-layout adapters, which don't exist here.

So if your chunked plan is doing spatial tiling, or you're on the v1 chunk contract, stop - this node isn't the one you want yet. It'll refuse, which is better than the alternative. The pack's habit of refusing rather than approximating is the reason it's usable at all on a family this big, but it does mean the limitations are load-bearing and you should read them before wiring rather than after.

Relay discipline in this pack

Three things the pack keeps saying and that are worth internalising here:

Real routing needs at least two local events. One event, or none, is a legal configuration that does approximately nothing. The default is deliberately empty - the examples ship with no events so nothing surprising happens when you open them.

report_only doesn't install routing bias, so a report-only Relay run returns a model patch and leaves the numbers alone. That's the mode you use to prove the plumbing before committing to an effect.

Don't stack two Relay conditionings onto the same model. The pack documents this directly: if your recipe already has a Relay, modify that plan rather than bolting another conditioning onto the same MODEL. And a Relay's freshly created empty AV is not a replacement for the frozen AV you're refining - that's a specific warning in the Audio Refine docs and it applies to the chunked route just as much.

Install

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

Manager users can search MiniMax H3 Audio T8; note the author ships GitHub and Registry independently, so clone when the version matters. Fully quit ComfyUI and start it again - refreshing the browser will not register new nodes. No dependency install: the pack's requirements file is deliberately empty so it can't replace ComfyUI's Torch or CUDA stack. Requires a recent ComfyUI with native H3 support. Weights are your own: H3 in models/diffusion_models, Qwen in models/text_encoders, video and audio VAEs in models/vae.

Where it breaks

  • v1 chunks or spatial tiles. Not supported by this adapter. Expect an explicit refusal.
  • A Relay plan whose timeline doesn't line up with your segment. Geometry validation is there to catch it; read the report rather than assuming it passed.
  • Wiring the runtime to the wrong audit or wrong segment. These are paired objects, not interchangeable tokens.
  • Treating a validated binding as an effect that worked. The audit tells you what actually executed, and a green wire has never been proof of coverage in this pack.
CategoryT8/MiniMax H3/Modular Sampling/Chunked Experimental

Inputs (10)

NameTypeDefaultDescription
relay_modelMODEL—
relay_positiveCONDITIONING—
relay_av_latentLATENT—
prompt_relay_planH3_T8_PROMPT_RELAY_PLAN—
source_segmentLATENT—
lifted_segmentLATENT—
segment_specT8_CHUNKED_SOURCE_SEGMENT—
pass2_contextT8_CHUNKED_PASS2_CONTEXT—
planT8_H3_CHUNKED_TWO_PASS_PLAN—
sigmasSIGMAS—

Outputs (4)

NameTypeDescription
modelMODEL—
positiveCONDITIONING—
runtimeT8_CHUNKED_PASS2_RELAY_RUNTIME—
report_jsonSTRING—