MiniMax H3 Runtime Metrics Probe

Before you trust any claim this pack makes about a run, you need to know what actually happened: how many times H3's transformer was really evaluated, how many of those were Spectrum forecasts, what the sampler did at each sigma. MiniMax H3 Runtime Metrics Probe is the passive instrumentation node that tells you - without enabling any of the pack's active tricks.

The name undersells the guarantee. This node installs the flow-metrics wrappers and only the wrappers: no trajectory capture, no guidance, no progressive handoff, no attention changes. Your model samples exactly as it would have, and in the meantime every sampler and model call gets counted and timestamped. It's the control condition for everything else in the pack, and it's also the honest debugging tool - when a workflow "looks wrong," the probe is how you find out whether it's a scheduling problem or your eyes.

Inputs and outputs are minimal by design. Feed it your model (the README suggests placing it after Spectrum when you want exact-versus-forecast provenance recorded - put it upstream of Spectrum and it can't tell the difference). Optional metrics input lets it append to an existing H3_FLOW_METRICS object if one already exists from another node, rather than spawning a second artifact. Outputs: the patched model (identical behavior) and its own fresh metrics.

What you get out of the metrics handle: logical sampler calls vs. actual H3 transformer evaluations, Spectrum forecasts and promotions, sigma and the unshifted flow coordinate for each point, phase and outer-step metadata, and wall-time per sampler stage. That last one is what makes the probe worth wiring into a graph you're tuning: "am I paying for forecast calls I thought were real, and where did the minutes actually go?" Feed the metrics into MiniMax H3 Metrics JSON to autosave it as a file under ComfyUI's output directory.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/MiniMax-H3-Flow-Aligned-Regenerate.git

Restart. No Python deps beyond ComfyUI's, no model downloads.

One honest limitation

Telemetry tells you the intended path executed. It does not tell you the clip is good - the author is emphatic that decoded video and audio are the quality gate, and metrics are the evidence ledger, not the verdict. Use the probe to confirm the machinery, then watch the media. It's also worth noting this whole pack is a single-author research effort with essentially no community footprint yet (as of writing, these nodes show zero search impressions), so the "trusted operating points" are the author's own matched tests - the probe is how you'll generate your own.