Joint Image-Video Training

Key Insight

When you fine-tune an image model on video alone, its still-image quality quietly collapses — the rich appearance knowledge it started with drifts away because video datasets are smaller and more compressed than image datasets. Joint image-video training prevents this by mixing the two in every batch — here 90% still images (each treated as a one-frame "video") and 10% real clips — so the model keeps practicing crisp single-frame generation while it learns motion. This project co-trains your inflated model both ways and compares it against a video-only run, so you can see directly how much sharper and more data-efficient the joint recipe is. It works without any change to the architecture because a still image is simply the T=1 special case of a video: the very same layers process both.