DynamiCrafter

Why we included this project

DynamiCrafter animates open-domain images by conditioning a video diffusion model on a still frame and a text prompt, so any everyday photo can be brought to life with motion you describe. It is the implementation of an ECCV 2024 oral paper, and the repo includes pretrained checkpoints (the 576x1024 version currently ranks first on the VBench image-to-video leaderboard), command-line inference, a Gradio interface for trying things out, and full training and fine-tuning code. On top of plain image-to-video, it also provides frame interpolation and looping models, which are handy for creating seamless transitions and loop assets. That makes it a practical pick for researchers and engineers who want to reproduce, adapt, or extend a strong open image animation method.