DynamiCrafter

DynamiCrafter animates still images into short videos, using text prompts to guide the motion. Developed by researchers from CUHK and Tencent AI Lab and presented as an oral paper at ECCV 2024, it leverages pre-trained video diffusion priors to synthesize natural movement from a single open-domain image, and it also handles frame interpolation between two input images, looping video generation, and multi-scene storytelling. Released checkpoints cover three resolutions, 256x256, 320x512, and 576x1024, generating up to 16 frames with VRAM needs from roughly 12 GB to 18 GB, plus a dedicated interpolation model. The 1024 model has ranked first on the VBench image-to-video benchmark. Local use goes through Gradio interfaces or ComfyUI integration, with Docker deployment supported and hosted demos on Hugging Face, Replicate, and OpenXLab. Open weights make it accessible to researchers, content creators, and developers building image-to-video features.