H3 Chunked v1 · Segment Relay Actual Calls (T8 EXP)

Here's the thing nobody tells you about attention-level effects in ComfyUI: a connected wire proves nothing. You splice in a patch node, the sampler runs, the picture changes a bit, and you assume the effect fired. Half the time it silently delegated, or the patch got overridden by the next node in the chain.

H3 Chunked v1 · Segment Relay Actual Calls exists to kill that doubt for exactly one v1 PASS2 segment. It's an audit node, and it's an output node - nothing downstream needs it, you just run it and read the numbers.

What "Relay" means here

Prompt Relay is this pack's timeline-script binding: a plan lays out what happens when - persistent scene text as a global layer, one-off actions and dialogue as timed local events - and then gets bound to a specific MODEL/CONDITIONING pair so the model's attention is routed per event, not per prompt. It's the same impulse as LTX's Director-style timeline editing: give the model a schedule instead of a paragraph. In the chunked v1 route, that binding has to be re-projected onto each temporal segment, which is what MiniMaxH3ChunkedV1RelayProjectEXPT8 does.

This node tells you whether the binding actually executed.

The mechanism

You give it three things:

• segment_result - the typed T8_CHUNKED_PASS2_RESULT from the PASS2 sampler for this segment.
• segment_spec - the same typed slice identity the sampler used.
• runtime - the T8_CHUNKED_V1_LOCAL_RELAY_RUNTIME returned by the v1 Relay Project node you wired in front of that PASS2 segment.

The runtime object carries the binding hash and the execution counters captured when the segment started, plus the live counts at the end. The audit diffs them and reports actual forwards and routed attention calls for the segment. If the runtime was never prepared, was already closed, or belongs to a different segment/spec, you get a refusal rather than a reassuring zero. Mismatched pairings failing loudly is a recurring theme in this pack, and it's a good instinct.

It outputs cumulative_av_latent (the whole clip as completed so far - same value as the sampler's cumulative output) and report_json. Read the report; that's the product.

What it will not tell you

• It won't cover Relay+EAV combined. If you wired both an EAV config and a Relay pair onto the same MODEL, those combined calls are attributed to the EAV audit, not to this standalone report. Two audits, two claims, no double counting.
• Calls are not quality. A non-zero routed-attention count means the routing happened. It does not mean the picture got better, that the effect reached every layer you care about, or that your script's timing reads correctly on screen.
• It's one segment. In a three-segment graph you need three audits, each paired with its own segment's runtime. This is the "one node per segment" pattern again - the family refuses to do a loop behind your back.

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

Fully restart ComfyUI, then refresh the page. No extra Python packages for the base and EXP nodes; the pack keeps its requirements.txt empty on purpose so installing it can never replace your Torch/CUDA stack. Weights - the H3 main model, Qwen text encoder and the video/audio VAEs - come from the author's HuggingFace account (t8star) and go in models/diffusion_models, models/text_encoders and models/vae respectively, per the README's model table.

Troubleshooting

Missing node or missing inputs after updating. Check for a duplicate copy of the pack in custom_nodes - an old test-copy directory shadowing the real one is the number one cause of "the panel doesn't match the workflow". Rename it with a .disabled suffix (leading underscore doesn't count), restart, verify python_module in /object_info.

"Runtime not prepared" or a binding-changed error. You moved the Relay Project node, changed the plan, or swapped the MODEL between the projection and the segment run. The audit wants the exact runtime that the PASS2 segment consumed - rewire it, don't work around it.

EAV applies are where people get burned. The default EAV mode is report_only; it measures and changes nothing. If you expected a visible effect, that's why. The pickles are yours to opt into.