Training Krea 2: What matters in generative model training — Sangwu Lee, Krea.ai
- Source
- youtube.com
- Author
- AI Engineer
- Date
Context
What an image caption leaves out can shape what an image generator learns. The published description of Krea's training talk gives a concrete example: captions for paintings photographed on walls can omit the surrounding wall and frame. The model then learns an unwanted association between a painting and that presentation.
The account describes data preparation aimed at controlling such effects, including removing duplicate images and using learned filters to identify visual artifacts. These choices explain why training-data quality includes accurate descriptions and unwanted correlations, as well as image quality. The example is the speaker's explanation of the training problem, rather than a quantified experiment isolating its effect.
Checking sign-in…
Loading comments…






