Deterministic 16-bit console pixel art from video, as ComfyUI custom nodes. No sampler, no seed, no flicker.
Diffusion models restyle a still beautifully and a video badly. Every frame draws its own noise, invents its own detail, and the result boils. The usual answers bolt cross-frame attention on afterwards: AnimateDiff and WAN both work, and both cost 20 GB and most of an hour for a hundred-second clip.
Pixelvid removes the cause instead. A diffusion model styles a few dozen keyframes once, offline. A small convolutional network learns the mapping from (frame, line art) to a palette index per art-grid pixel. That network then runs over every frame with no sampler, no noise tensor and no seed anywhere in the path.
Nothing is random, so nothing flickers. Temporal coherence is a property of the function rather than something bought with VRAM.
The full documentation is in the README.