视觉生成中文本条件化的缩放特性研究

Longer prompts are not what image generators need.

Text-to-image models seem less constrained by prompt length than by how clearly the prompt exposes the scene.

This paper finds that text conditioning scales with image-grounded information, not token count.

Across open-weight models, simply extending natural-language captions eventually made outputs worse than each model’s shortest-caption result.

The authors replace prose with a structured prompt that separates the scene, individual objects, bounding boxes, depth, attributes, and relationships into named fields.

The shift here is: prompt engineering for visual generation should optimize how explicitly visual variables are represented, then train the prompter to fill that representation well.

The biggest prompt-engineering gain may come from how visual content is organized before it reaches the image model.

– arxiv. org/abs/2607.29679

Title: "Scaling Properties of Text Conditioning in Visual Generation"