ali-vilab/VGen

Supports multiple controllable generation modes—image-to-video synthesis, motion-guided composition, and human-feedback refinement—via cascaded diffusion architectures with latency-optimized variants like VideoLCM. Includes unified training infrastructure for joint image-video learning, LoRA fine-tuning, and inference acceleration tools, targeting researchers building custom video diffusion applications.

3,152 stars. No commits in the last 6 months.