[Legacy/Research] MiniMax H3 RefDelta Scheduler

The honest headline

This node is deprecated, and the deprecation is the whole story. The display name says [Legacy/Research], and the README is even blunter: for production, use ComfyUI's stock BasicScheduler with scheduler = beta. This node exists so old saved workflows still load and so people doing scheduler research have something to poke at. It is not where you should be pointing your sigma generation today.

What it is

A custom scheduler that shapes the noise schedule (sigma curve) from a "profile" - a list of difficulty values across denoising progress, weighted against a continuous beta prior. The idea was that a profile could redistribute sampling effort to match how hard each phase of the RefDelta trajectory actually is.

That idea died for a mundane reason: there was never a calibrated profile. The bundled r1024_provisional profile is, by its own metadata, a neutral beta(0.6, 0.6)-prior compatibility profile with flat difficulty of 1.0 everywhere. It is the scheduling equivalent of "we have no data, so here's a safe middle." The author's words: "There is no production calibration file to select from the FL2VA/Ref2VA replay workflow." The diagnostic comparison data is explanatory, not a scheduler oracle - explicitly so you don't get clever and start densifying schedules based on it.

The inputs

• model - any H3 MODEL; the node pulls the model's model_sampling to build sigmas from.
• steps (20) - step count.
• denoise (1.0) - denoise fraction, same as any scheduler.
• profile - the only choice is r1024_provisional, a neutral compatibility profile. If a requested profile is unavailable it falls back to a uniform profile rather than erroring.

Output is SIGMAS, which feeds the sigmas socket of a KSamplerAdvanced or SamplerCustomAdvanced.

If you're doing research

The research tooling lives in tools/build_profile.py in the repo. It's deliberately hobbled: it strips comparison_* and ref_delta_* fields before binning, deduplicates replay copies of the same telemetry, emits neutral difficulty by default, and requires an explicit --experimental-stability-density flag before it will emit any non-neutral density. Generated profiles are marked non-production. That's the author being careful about the exact trap mentioned above - building a schedule from diagnostics that weren't validated as schedule evidence.

python tools/build_profile.py \
  ComfyUI/output/refdelta_telemetry/*.csv \
  --output /tmp/r1024_scheduler_research.json \
  --id r1024_scheduler_research \
  --experimental-stability-density

Install and practical advice

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/ComfyUI-MiniMax-H3-RefDelta-Solver.git

Restart ComfyUI; no pip dependencies. If a workflow you downloaded still contains this node, the low-effort fix is to delete it and wire BasicScheduler (scheduler beta) into the sampler instead - that's literally the production recommendation the pack ships with. Keep the node only if a saved workflow needs it to load, or you're actively researching schedule density.