4 ms·
It ought to be possible to guide a diffusion-based model to make several panels about a single, consistent, AI-invented character. It can already generate seve
by manimino 4y ago
It ought to be possible to guide a diffusion-based model to make several panels about a single, consistent, AI-invented character.
It can already generate several independent images containing known characters that are in the training data (Obama, Pikachu, etc.)
- jcims 4y agoI'm a total layperson in this area but there used to be a concept of 'fine tuning' with language models (may still be). My understanding of those is that you could inject additional training data that overlays(?) the existing model in some way to help direct the output. In this case it seems that you could fine-tune the generative image model with previous frames to provide that continuity. After all that's what we do when we read the panel, we instantly store the previous one in memory so that we can actually recognize the difference in the next panel.