Animated-ART: Multi-Layer Transparent Video Generation

Layered image design is fundamental to professional creative workflows, and the generation of layered images has attracted significant research interest. In this work, we extend layered image generation to the video domain, introducing the novel task of multi-layer transparent video generation-synthesizing multiple transparent dynamic layers that compose into coherent video sequences. We first construct the Transparent Dynamic Layer (TDL) dataset, specifically designed for training and evaluating models on animated RGBA layer sequences. Building on this dataset, we propose the Diffusion Multiple Dynamic Layers (DMDL) model, which generates variable-resolution transparent dynamic layers from static layer inputs, text prompts, and motion region layouts. DMDL supports both single-layer and multi-layer animated generation. Our approach makes two key technical contributions: (i) a latent multiple dynamic-layer diffusion model with staticto-dynamic generation capability, and (ii) a transparent dynamic-layer autoencoder. The diffusion model incorporates layer-aware spatial-temporal 4D-RoPE positional embeddings, enabling cross-layer interaction of visual tokens across layer, spatial, and temporal dimensions. Our autoencoder employs a specialized ViT-based decoder that leverages a two-stage, layout-conditional 3D-RoPE strategy to reconstruct transparent dynamic layers, effectively handling the varying levels of temporal upsampling inherent in the process. Extensive experiments validate our method's effectiveness, establishing a strong baseline for dynamic layer generation.