AI

GRACE: Generation-aware latent compression for efficient video generation

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 03:18:25 UTC.

Briefings

A two-stage framework called GRACE reduces video generation token count by 8x and latency by 11.1x while maintaining generation quality on VBench. The method compresses a pretrained video autoencoder using a frozen base latent and learns a residual latent aligned with the pretrained DiT's feature space. This approach avoids retraining the DiT from scratch, though residual learning still shifts the compressed latent away from the original distribution. The study was conducted on Wan2.1-I2V-14B models at 480x832x81 resolution.

Key points

• GRACE reduces Wan2.1-I2V-14B token count by 8x and latency by 11.1x at 480x832x81 resolution.
• The framework uses a frozen base latent and learns a residual latent aligned with the pretrained DiT's feature space.

Context

Video diffusion models like DiT require large token counts, but highly compressed autoencoders degrade reconstruction quality. GRACE addresses this by compressing the autoencoder while preserving compatibility with the pretrained DiT through a two-stage approach.

Limitations

The study was conducted on Wan2.1-I2V-14B models at 480x832x81 resolution. The evidence does not specify the training data distribution or the specific video generation tasks tested, limiting generalizability to other models or resolutions.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 04/11/2026, 13:31:28 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiccikoahegwlo5zpn5lxgoahl6jzwniuwmuk55nmxm7bmiaoq6zva