MiniMax H3 Metrics JSON

Every node in this pack that does something interesting also spits out a H3_FLOW_METRICS object - but a dataclass in a wire isn't a record. MiniMax H3 Metrics JSON is the terminal node that turns that object into something you can save, diff, and point at later: it renders the whole thing as structured JSON and autosaves it under ComfyUI's output directory.

It's the natural sink for any of the pack's metrics outputs. Chain Trajectory Capture's metrics into Flow-Aligned Regenerate's optional metrics input and you get one combined artifact spanning both passes, rather than a pile of fragments. Wire a probe's metrics in for a pure diagnostic run and you get a clean before/after record for a settings experiment. The node is an output node, so it sits at the end of a chain and shows its message in the UI.

What you set. One required input, metrics (the H3_FLOW_METRICS object from any of the other nodes), and one optional: filename_prefix, default h3_flow_regenerate/metrics, which becomes the output path. If the metrics object already knows an autosave path from a prior stage, the node respects that instead. The single STRING output is the JSON payload itself - handy if you want to pipe it somewhere else, though most people just let it land in output/h3_flow_regenerate/metrics_XXXXX_.json.

What's actually in the file is worth knowing before you open it and feel lost: counters and timestamped events covering logical sampler calls versus actual H3 transformer evaluations, Spectrum forecast/promotion events, progressive stage (low/probe/high), sigma and the unshifted flow coordinate at each point, resolved handoff and source/target geometry, trajectory commits with provenance, guidance-component RMS ratios, sampler/stage wall time, and the resolution-sigma-map diagnostics from the experimental sigma node. For the autosave to stay live, later sampler events refresh the same file, so the copy on disk tracks the run as it completes.

Install

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

Restart. Zero extra dependencies, no models.

The one thing to remember

Metrics JSON establishes whether the intended path executed - that the handoff landed where you asked, that the guidance mode actually applied, that you paid for the H3 calls you think you did. It is not a quality gate. The author's own words: decoded video and audio remain the pass/fail criterion. Save the JSON for reproducibility and debugging, then go look at the clip. And when you're sharing results from this pack - still a niche single-author effort with essentially no community consensus to lean on - attaching the metrics file is the difference between "trust me" and "here's the ledger."