baaivision/NOVA
NOVA is an autoregressive model for generating images and videos without vector quantization, achieving continuous-valued generation through a diffusion-based approach.
Velocity · 7d
+1.2
★ / day
Trend
→steady
star history
NOVA (NO-nQuantized Video Autoregressive Model) enables autoregressive image and video generation without discrete vector quantization. Instead, it uses continuous-valued predictions combined with diffusion-based training to generate high-quality visual content. The model handles both text-to-image and text-to-video generation tasks, with inference demos available through HuggingFace spaces.