Etai Sella

LLM interrogations reveal diffusion model's internal tokens capture scene semantics surprisingly early, even without explicit prompts, and correlate with generation quality.