GenVC: First video diffusion model cuts bitrate by 71% with single-step decoding

Generative video compression has long relied on diffusion models originally built for text-to-image generation, adapted post-hoc for codec tasks. A new paper from researchers at multiple institutions introduces GenVC, a video diffusion model trained from scratch specifically for compression — a first in the field. Operating directly in pixel space with a global-to-local hierarchical structure, GenVC recovers fine spatiotemporal details from heavily compressed representations. The model uses distribution matching distillation (DMD) for single-step inference, but naive DMD produced motion-stalled reconstructions due to teacher guidance failures. To solve this, the authors propose Adaptive Score Distillation (ASD), which gates DMD updates based on alignment with ground-truth directions, breaking the feedback loop that caused motion artifacts.

GenVC delivers state-of-the-art perceptual quality at ultra-low bitrates. On standard benchmarks, it achieves 62.5% bitrate savings at matched LPIPS and 71.3% at matched FID compared to the GLVC baseline. Remarkably, the model uses just 478 million parameters — far fewer than prior billion-scale pretrained backbones — and decodes 1080p video in a single step at 15.1 frames per second on an A100 GPU. This makes GenVC not only more efficient in storage and bandwidth but also practical for real-time or near-real-time applications. The technique opens the door to compression-first diffusion models that prioritize motion coherence and perceptual fidelity, potentially reshaping video streaming, conferencing, and storage.