H3 Chunked · PASS2 EAV Actual Calls (T8 EXP)

There's a specific kind of disappointment in ComfyUI: your effect node is green, the render completes, and the effect did nothing. Usually because a wrapper, a backend or an attention patch sat between your node and the model and quietly ate it.

This audit node exists for that exact failure. It reports the actual EAV forward and attention calls that happened during one chunked PASS2 segment, on the pack's stated principle that a graph connection alone is not proof of effect coverage.

Inputs and outputs

Three inputs, all required. segment_result is the T8_CHUNKED_PASS2_RESULT from the segment's PASS2 sample. segment_spec identifies which segment you're auditing. runtime is the T8_STAGE_EAV_RUNTIME that MiniMaxH3ChunkedPass2EAVApplyEXPT8 produced for that same segment - and it needs to be the same segment, not a sibling.

Outputs are cumulative_av_latent and report_json. It's an output node, so the report also renders on the node. cumulative_av_latent is the running full-video result: in a chunked run you carry it forward, segment by segment, and it becomes the AV you eventually decode. That's the structural difference between the chunked route and a single 4+4 pass - each window is sampled against the global context and stitched through this cumulative output, not concatenated afterwards with a blend.

Read the report. The call counts in it are the answer to "did my effect actually reach the model", and in this pack a count of zero is a real, reportable, non-crashing outcome.

Unit gain is not a win

Here's the thing the pack shouts in every document and that people still miss: a measurement existing is not the same as an effect working. EAV reports a gain, and the pack has published qualification runs where the effect executed correctly and every single gain measured 1.0 - proving the routing worked and nothing else.

So when your audit reports hundreds of calls and the picture looks identical, that's consistent. Check the gains, not just the counts. And if you're using report_only in your EAV config, expect exactly this: the pack's design is that report_only doesn't apply the bias at all, so the value of the run is the telemetry, not the output.

Scope limits worth knowing

Only full_frame_safe single-tile plans are supported in this adapter path; the pack is explicit that multi-tile producer coverage isn't certified, and it prefers refusing to half-serving you.

If you're running Relay and EAV together on the same segment, the combined Relay+EAV calls are counted by this audit, not by the Relay audit. That division is stated in the node descriptions of both, and getting it wrong makes you think one of the two effects vanished.

Install

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

Manager equivalent: search MiniMax H3 Audio T8. Then fully quit ComfyUI and relaunch - a browser refresh is not a restart and won't register new nodes. No dependencies to install, intentionally: the pack's requirements file declares no packages at all so it can't displace ComfyUI's Torch or CUDA stack. You do need a recent ComfyUI with native H3 support, and you supply the weights: H3 diffusion model to models/diffusion_models, Qwen text encoder to models/text_encoders, video and audio VAEs to models/vae. Third-party helper packs some of the optional example graphs use (KJNodes, Comfyroll and friends) are not installed by this pack - the README is clear about that, and if you're following a chunked example that leans on them, you install them yourself.

Failure modes

• Runtime from a different segment. The audit is bound to a specific segment spec; mismatch is an error, not a merge.
• Expecting it to apply anything. It audits and returns the cumulative latent. The effect, if any, happened in the PASS2 sample.
• Reading a clean report as an accepted render. It reports mechanical coverage. Quality is a human question, and the pack says so.
• A stale copy of this pack in custom_nodes shadowing the live one. It has produced confusing keyword-argument errors for users before. Keep one install.