2 Jan 2026
Custom training can now reach an AI model's 'eyes'
- What was broken
- LLaVA is an AI model that can look at pictures. People could add small custom training (called LoRA) to it, but only to the part that handles text, not the part that reads images.
- What I changed
- I wired up the image-reading parts so that custom training applies to them too.
- Why it matters
- Teams can now teach the model their own kinds of pictures as well as their own words.
see the code, #31513 ↗ (opens in a new tab)
← every project