media creative tools
Video Diffusion Lost Eight Times the Tokens, Not the Picture
GRACE compressed a pretrained autoencoder while aligning its new latent to the frozen diffusion transformer's feature space.
Summary
GRACE compressed a pretrained autoencoder while aligning its new latent to the frozen diffusion transformer's feature space.
Compressing a video autoencoder reduces the token load seen by a diffusion transformer, but the new latent can drift away from what that transformer learned. GRACE keeps a frozen base latent, learns a residual for information lost under heavier compression and aligns the compressed representation inside the frozen transformer's feature space. Lightweight adaptation and asymmetric denoising then process the base ahead of the residual. On Wan2.1-I2V-14B at the reported resolution and duration, the method cut token count eightfold and latency 11.1-fold while matching the pretrained pipeline's VBench generation quality. The comparison is tied to that pipeline and benchmark, yet it shows why reconstruction quality alone is the wrong target when a latent exists to serve a generator.
Why it matters
GRACE compressed a pretrained autoencoder while aligning its new latent to the frozen diffusion transformer's feature space.
Limits and context
No additional limitation was separately recorded.
Key claims
GRACE compressed a pretrained autoencoder while aligning its new latent to the frozen diffusion transformer's feature space.
Evidence: source-2026-10-08-009
Sources
- arXiv preprint 2610.10524arXiv · primary research
Corrections
No corrections have been recorded for this story.