💡 This paper studies the scaling properties of text conditioning in visual generation, which has rarely been measured due to the difficulty of scaling diffusion loss with the number of tokens in natural language prompts. The authors surprisingly find that the converged diffusion loss scales with the amount of structured language in the prompt. To quantify structured language, they adapt two complementary measures: a white-box like likelihood metric and a black-box attribute metric.

The authors use these metrics to analyze the relationship between the converged diffusion loss and the amount of structured language in the prompt. They find that the converged diffusion loss decreases approximately linearly with the white-box metric and follows a power law with the black-box metric across controlled training runs.

Guided by these scaling properties, the authors improve the diffusability of visual generation models by constructing structured prompts with semantic and geometric annotations derived from images. They also improve promptability by training a prompter through supervised fine-tuning, cold-start, and verifier-gated on-policy distillation.

The resulting system outperforms all evaluated open-weight models on nearly every compositional, reasoning, and world knowledge benchmark, while matching or surpassing the strongest closed-weight models on most evaluations. The paper's contributions include providing a better understanding of the scaling properties of text conditioning in visual generation and developing methods to improve the diffusability and promptability of visual generation models.