H3 Plan Estimate (alpha)

Alpha node from the timeline surface, and one of the more honestly-designed utilities in the pack. The question it answers is "how much is this retiming going to cost me" - before you commit the GPU. The interesting part is how it answers: it refuses to guess from your GPU's spec sheet. Its cost estimates come from your own flight recorder, the JSONL that H3 Record Start / H3 Record Stop write. An uncalibrated guess is worse than no answer - that's not my spin, it's in the node's description.

The three verdicts

• equivalent_clip_time (FLOAT) - how many times a plain render of this clip the plan costs. This one is exact from the plan and needs no calibration; the dilated frame count is arithmetic.
• seconds (FLOAT) - real minutes from YOUR flight recorder. It comes back -1 when your box has no recorded runs, and that's deliberate: no runs on record, no estimate. Run a render or two through the recorder pair first, and it calibrates from those.
• work_units (INT) - how many latent tokens / dilated frames the plan generates, the raw cost driver.
• vram_band (STRING) - a bracketing verdict from your recorded runs (does this fit your card?), not a peak-bytes prediction.
• report (STRING) - the human-readable version of all of the above.

Inputs: plan_path (required - a plan.json from the editor, the oracle node, or the CLI), plus optional plan_json (pasted, wins over plan_path), ignore_uncompiled_lanes, port (which instance's /system_stats the VRAM reading comes from, default 8188), and recorder_path (the flight-recorder JSONL, default /root/.cache/mainodes/flight_recorder.jsonl).

Why this design is right

Two nodes in this pack make the same principled choice: H3PlanEstimate and the oracle's cost report both take s_per_step from your measured runs rather than multiplying a generic number. It's the difference between a pricing model and a metronome - a GPU spec doesn't tell you how long your clip's token count takes on your card with your attention patches. If this pack has a house philosophy, that's a piece of it: measure, don't assume.

Install

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

or ComfyUI Manager → ComfyUI-MAINodes, restart. Category latent/minimax/timeline.

Common gotchas

seconds at -1 isn't a bug - it's the node being honest that your recorder is empty. And if you point port at the wrong ComfyUI instance, the VRAM band is read from a box that isn't the one running your render. The ignore_uncompiled_lanes default (off) means a plan carrying lanes this backend can't compile gets refused rather than silently mis-estimated - turn it on only if you understand what gets dropped.