Dreamix

Dreamix is a video diffusion model that enables text-based motion and appearance editing of real-world videos. By combining low-resolution spatio-temporal information from the original video with newly synthesized high-resolution content, Dreamix aligns the output with guiding text prompts. It employs a novel mixed finetuning approach to enhance motion editability and supports applications such as image animation and subject-driven video generation. Extensive experiments have demonstrated Dreamix's superior performance in producing temporally consistent and high-fidelity edited videos.