GRACE Cuts Wan2.1-14B Video Latency 11.1× at Matched VBench

GRACE, a video-diffusion compression method posted to Hugging Face Papers, reduces the token count of Wan2.1-I2V-14B by nearly 8× and cuts its latency by 11.1× at 480×832 with 81 frames while matching the pretrained pipeline's VBench score. The paper notes the work "was done while the first three authors were interns at Kakao Corp."

The method keeps the frozen pretrained encoder's base latent and learns a residual latent for information lost under stronger compression. The autoencoder, the paper writes, is aligned "with the pretrained latent in the feature space of the frozen DiT, so that the autoencoder is optimized for generation rather than reconstruction alone." A second stage adapts the DiT with lightweight fine-tuning and asymmetric denoising, where "the base latent is denoised ahead of the residual latent" in a shared forward pass.

The headline VBench parity is cleaner than the reconstruction numbers. On VBench-I2V, GRACE totals 87.90 against the pretrained Wan2.1-14B's 87.92. On Panda-70M at 256×256×81, PSNR falls from 35.15 to 32.63 and rFVD jumps from 1.13 to 13.53. The retrofit cost sits at 38.5 H200 GPU-days (8.5 for the autoencoder, 30 for the DiT), and at 736×1280×81 the authors report 218.8 seconds against 3396.8 for the pretrained baseline on a single A100, a 15.5× speedup. It lands in a dense run of inference-efficiency work we have been tracking this quarter.