Scaling Properties of Text Conditioning in Visual Generation

TL;DR - An empirical study showing that converged diffusion loss in text-to-image generation scales with the amount of "structured language" in prompts rather than raw token count, and uses that insight to build a system that beats open-weight models and rivals closed-weight ones.

• Diffusion loss doesn't scale with prompt token count, so the authors quantify prompt structure with two metrics: a white-box likelihood measure (GPG) and a black-box attribute measure (ED).
• Across controlled training runs, converged diffusion loss falls roughly linearly with GPG and follows a power law with ED.
• "Diffusability" is improved by constructing structured prompts with semantic and geometric annotations derived from images.
• "Promptability" is improved via a trained prompter using supervised fine-tuning, cold-start, and verifier-gated on-policy distillation; the system leads on compositional, reasoning, and world-knowledge benchmarks.