I2VGen-XL

I2VGen-XL is an image-to-video synthesis method from Alibaba's Tongyi Lab, published within the VGen video generation codebase. It animates a still image into high-definition video through cascaded diffusion: a base stage establishes semantics and motion coherent with the input image, and a refinement stage sharpens detail and raises resolution, optionally guided by a text caption. Trained on WebVid-10M and LAION-400M, the released models are intended for research and non-commercial use, and the authors note weaker results on anime-style images and black backgrounds due to training data coverage. Inference is available through the VGen repository along with a local Gradio demo, plus hosted versions on ModelScope, Hugging Face, and Replicate. Researchers exploring image conditioning in video diffusion use it both as a baseline and as a component of VGen, which bundles several related Tongyi Lab video generation models in one codebase.