H3 Probe Schedule (early-oracle head)

The de-rope pipeline works by reading a baseline render, deciding where the motion is too fast, then regenerating a slowed version. The baseline costs real money - it's a full pass. H3 Probe Schedule is the cheat: it runs only the head of that baseline schedule, because the burst timing is already readable by step 4–5 of 25. Run the probe, read the oracle and the init from the early x0 estimate, and skip roughly 75% of the first pass.

How it works

It's a scheduler node - you wire it where you'd wire a stock scheduler, and it hands the sampler a SIGMAS list that stops early. The node itself is deliberately dumb; the cleverness is in what you do with its output. Wire the sampler's denoised_output (the x0 estimate) into H3 Jerk Oracle, and into the decode that feeds H3 Time Smear. The oracle still sees the motion, and the early x0 is a workable init for the regeneration pass - partly because injection destroys fine detail anyway, so a fully finished baseline was never going to survive the regen intact.

Inputs are minimal and you mostly leave them alone:

• model - the diffusion model (required, it's a sampler-facing node).
• scheduler - simple (default), normal, beta, sgm_uniform, karras, exponential. simple is the shipped choice.
• total_steps (25) - what the full schedule would be.
• probe_steps (6) - how much of it actually runs. This is the one dial. Raise it if the init loses too much choreography on your content; drop it if you're just iterating.

The honest trade-off

There is one, and it's audio. No finished baseline means no finished baseline audio: H3 Audio Recover's reference input has nothing full-speed to blend with, and the probe's own audio estimate is rough. For a dialogue clip this matters. For a motion-iteration pass - "does the choreography land before I pay for a final?" - it's the right price. The pack's own example (motion_pipeline_probe_expert.json) pairs this with an expert schedule that regenerates with a base-model head and a turbo tail, and it's the cheapest full de-rope in the set at roughly 8.5 minutes for a 5-second 1024 clip on the hardware they measured.

Install

This node ships with ComfyUI-MAINodes, so it installs with the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/matlowai/ComfyUI-MAINodes

or via ComfyUI Manager (search ComfyUI-MAINodes), then restart. It shows up under sampling/custom_sampling/schedulers. No extra Python deps; the example graphs that use it also want ComfyUI-KJNodes.

Common issues

If the regenerated clip loses choreography - motions that were in the baseline come back wrong or missing - the early x0 was too early for your content; raise probe_steps. If you wire the probe into a graph that expects a finished baseline downstream (like one feeding H3 Audio Recover's reference_mix), you'll get a thin or wrong audio track, and that's not a bug - it's the documented trade of this node. Keep the probe for iteration and the full pipeline for the keeper.