The artist Karen X. Cheng shows in a small demo video which creative possibilities are already possible with the help of various free AI tools. The clip shows her taking a few steps while her AI-generated clothing transforms several times into a different style - a time-lapse fashion show, so to speak.
She generated the new clothes per specific prompts using the in-painting function in DALL-E 2. This algorithm is designed for single frames, not sequences of frames - a video made on it has no consistency - the individual frames differ too much, there are also no sequences of multiple frames with the same outfit in motion.
The solution brought the AI tool EBSynth to Style Transfer, with the help of which changes to a single picture can be transferred to a whole video sequence, for example paintings can be animated. This made it possible, based on an image of a new outfit, to dynamically superimpose it correctly over the old clothing across several frames, entirely according to movement.
The final tool used was DAIN, a special AI model for frame interpolation originally created for stop-motion animation. With it, Karen Cheng was able to interpolate intermediate frames to make the change from one outfit to another smooth, and also realize a slow-motion effect.
EBSynth InputVideo - Keyframe and Result
The whole thing is just a small experiment, but it shows how the combinations of several AI tools can create interesting effects that would have been possible without these tools, too, but would have been significantly more complex to realize. From simply creating fitted clothing for a person at the touch of a button after only a brief description, to transferring that change to multiple moving images, to a smooth transformation from one outfit to another - all things that would have taken a lot of time manually and are now easily repeatable in variations at any time thanks to the AI workflow.
We also came across another interesting tool during our research that automatically creates animations between two images specified via text prompts using Stable Diffusion, as seen in the following clip: