H3 Mid Insert (change the token grid MID-denoise) [experimental]

Experimental proof of concept, flagged off-distribution by the author themselves. Where the rest of the Motion Lab pack inserts extra token-times between two passes (coarse pass, then a fresh dense pass), H3 MidInsert does the insert in the middle of a single denoise: run the coarse grid down to a handoff sigma, insert the in-between tokens, and finish the schedule on the dilated grid. The model then decides the motion with the in-betweens already in the sequence, and the coarse pass is partial rather than finished.

How it works

You split the schedule with SplitSigmas - pass A runs the high half down to sigma_s - and hand the still-noisy latent to this node. It inserts the held token-times using the same grid arithmetic as H3 Temporal Insert (shared code, so the two routes can never disagree about which token-times exist). The clever bit is the init: lerping two noisy neighbours is variance-deficient, so noise_topup (default 1) adds fresh gaussian noise sized from the measured per-token, per-channel residual of the neighbours. 0 is raw lerp, 1 is full top-up, and the report prints the measured correlation and deficit so you can see how much of it was real motion vs. noise. Wire sigma_s in and it also prints the noise-only bound the flow parameterisation predicts, which is how you judge that.

The wiring rules, and they matter

This node is not plug-and-play, and the description is explicit about the two traps:

• Pass B must use DisableNoise. Pass A's output has already been divided by 1 - sigma_s (inverse noise scaling). RandomNoise on pass B would add a second full noise draw on top and destroy the handoff. Zero noise makes noise_scaling multiply by 1 - sigma_s again and hand the sampler back exactly the state pass A left.
• No noise mask, deliberately. A repaint mask re-noises from a clean latent every step, and this latent is not clean - it's already at sigma_s. Any noise_mask on the incoming latent is dropped.
• Audio is passed through unchanged. It stays on the base clip's clock while the video moves onto the dilated one, so pass B sees an audio stream covering only part of the video span. Take the delivered audio from elsewhere. Unlike H3 Temporal Insert, this node will not fabricate a zero audio track, because a clean zero mid-schedule is not a valid state.

Inputs: samples (the mid-schedule nested AV latent), hold_map, noise_topup, seed (for the top-up draw). Optional expand_to_end (default false - a mid-denoise arm wants the same grid its between-passes control used) and length / sigma_s (the latter is report-only).

Should you use it?

If you're running the settled pipeline, no - this is research surface, the author calls it a PoC. It exists to answer whether mid-schedule insertion holds choreography better than the between-passes route, and it needs a user comfortable debugging a split schedule. If you want to poke at it, start from a graph that already does the temporal-insert dance and swap the insert stage in; the shared grid code means the token-times stay consistent, so you're comparing inits, not grids.

Install

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

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