Dreamix: Video Diffusion Models are General Video Editors
Dreamix
Few-shot finetuning for personalized video editing
Main idea: Mixed Video-Image Finetuning
- Finetune Imagen Video (Ho et al., 2022) which is a strong video foundation model
- Finetuned to generate individual frames (bypassing temporal attentions) & video
P163
Inference Overview
- Corrupt the input video by downsampling and add noise
- Apply the finetuned video diffusion model to denoise and upscale