HKU, DAMO Academy (Alibaba), Hupan Lab · Text-to-Video

LayerFlow

LayerFlow is a unified diffusion transformer model for layer-aware video generation that decomposes videos into transparent foreground, clean background, and blended scenes using per-layer prompts. It supports multi-layer generation, decomposition, and conditional layer synthesis through a multi-stage training strategy combining low-quality video data with high-quality layered image annotations.