Scaling Properties of Text Conditioning in Visual Generation
2026-07-31 • Computer Vision and Pattern Recognition
Computer Vision and Pattern Recognition
AI summaryⓘ
The authors study how the complexity of language prompts affects the performance of image generation models based on diffusion. They find that the better the structured language in the prompt, measured by two different methods, the lower the model’s final error becomes. Using these insights, they create improved prompts with detailed annotations and train a special prompt generator. Their approach leads to better model outputs on tests requiring understanding, reasoning, and world knowledge compared to other models with open or closed weights.
diffusion modelstext conditioningprompt engineeringscaling lawslikelihood metricsemantic annotationsgeometric annotationssupervised fine-tuningon-policy distillationcompositional reasoning
Authors
Zilong Chen, Chaorui Deng, Kunchang Li, Hongyi Yuan, Haoqi Fan
Abstract
We study empirical scaling properties for text conditioning in visual generation. Such properties have rarely been measured because diffusion loss does not scale with the number of tokens in natural-language prompts. Surprisingly, we find that the converged diffusion loss scales with the amount of structured language in the prompt. To quantify structured language, we adapt two complementary measures: a white-box likelihood metric (GPG) and a black-box attribute metric (ED). Across controlled training runs, the converged diffusion loss decreases approximately linearly with GPG and follows a power law with ED. Guided by these scaling properties, we improve \emph{diffusability} by constructing structured prompts with semantic and geometric annotations derived from images, and improve \emph{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.