H3 Motion Ref Scale (internal)

Here's a fact about H3 that determines a lot of its VRAM behavior: the model re-encodes reference videos in full and packs every frame as a token that rides through every sampling step. Feed it a full-resolution motion reference and you've roughly doubled the sequence length for the whole generation, for no quality gain at all. H3 Motion Ref Scale (internal) is the pack's fix for exactly that waste.

The insight that makes it work is a division of labor. In Mask mode, the motion reference has exactly one job: carry movement. Identity - who or what the replacement is - comes from the separate <Picture N> reference images. Motion is a low-frequency signal, so it survives being small; identity is not the motion ref's problem. So this node downsizes the tracked source crop to a small short edge (256 by default) before it's fed to H3, and the author's own changelog confirms the trade is free: tracking and replacement quality are unchanged.

Inputs:

• images (IMAGE) - the tracked source crop being used as the motion reference.
• short_edge (INT, default 256, range 32–768, step 32) - the target short side in pixels.

Output: the resized images.

The mechanism has three rules worth knowing. It preserves aspect ratio, so a 16:9 crop stays 16:9. It snaps to the 32 grid - the size lands on multiples of 32, which matters because the H3 node's own canvas logic leaves a 32-grid reference at its reduced size instead of upscaling it back. And it never upscales: if your crop's short edge is already at or below the target, the node passes it through untouched rather than blowing it up pointlessly.

What should you actually set short_edge to? Default 256 is the sane starting point and what the author ships. If you're doing fine-grained motion like a dancer's footwork or a detailed hand gesture and see the motion getting mushy, nudging to 384 or 512 costs some speed but keeps more detail - the range exists precisely because some shots want it. For anything where the camera is fairly static, 256 is honestly more than enough.

It's internal to Mask mode, so you'll meet it inside the generated workflow rather than on a blank canvas, and it needs the same Mask-mode stack as its siblings: MaskVidExperiments plus the SAM 3.1 checkpoint. The whole point of this node is that you can forget it exists and still benefit - it's H3's VRAM budget being spent on the thing that actually matters.