What's this about?
GRACE compresses the intermediate video representation seen by a diffusion transformer while preserving compatibility with the pretrained generator. The authors report 8× fewer tokens and 11.1× lower latency for one Wan2.1 configuration.
What exactly is a latent?
It is a compressed numerical representation of a video, like a compact working sketch. Fewer latent tokens make generation cheaper, but excessive compression can erase details or present the generator with unfamiliar inputs.
Why does this matter?
It attacks video generation's runtime without retraining a huge model from scratch. Results are author-reported for specific settings; wider quality, hardware and resolution tests are needed.