A paper on Hugging Face measures for the first time how diffusion models (those that generate images) scale relative to the text you give them. Generation quality improves predictably when the prompt contains more structured language, regardless of word count. Researchers defined two metrics to quantify this structure (GPG and ED) and found regular relationships: diffusion loss drops linearly with the first and follows a power law with the second.

If you write prompts to generate images, the results suggest that organizing text with semantic and geometric annotations improves output quality in measurable ways. Starting from this insight, researchers built a system that outperforms all open-weight models on compositional, reasoning, and world-knowledge benchmarks, and matches or exceeds proprietary models on most evaluations.

If you want to try it: the code is public on GitHub, with the project described on the paper’s official page.

In detail

Diffusion models generate images from text. Until now, no one had systematically measured how generation quality changes as prompts get longer. The reason is technical: diffusion loss, the metric measuring how well the model reconstructs an image, doesn’t scale with the number of tokens in a natural language prompt. More words don’t mean better loss, so the relationship seemed non-existent.

This paper shifts perspective. Instead of counting tokens, researchers measure the amount of structured language in the prompt using two complementary metrics: GPG (a “white-box” measure based on language likelihood) and ED (a “black-box” measure based on attributes). They find that loss converges predictably: linear with GPG, power law with ED.

What does “structured” mean in practice? A structured prompt includes semantic annotations (what’s in the image) and geometric ones (where it’s positioned in space). Instead of “a cat on a table with a cup,” a structured prompt explicitly specifies positions, relationships, and attributes.

From here, researchers improve the system on two fronts. Diffusability: they construct structured prompts from annotations derived from the images themselves. Promptability: they train a “prompter” that transforms simple descriptions into structured prompts, using supervised fine-tuning, cold-start, and verifier-gated on-policy distillation. On this last technique, on-policy distillation is the same thread we followed with β-OPSD, where the paper identified the parameter that made distillation stable for reasoning models.

The result: the system outperforms all evaluated open-weight models on nearly every compositional, reasoning, and world-knowledge benchmark, and matches or exceeds the strongest proprietary models on most evaluations.

How much to trust this. Results come from the abstract and project page. The paper was published on July 31, 2026 and uploaded to Hugging Face on August 3, with 23 upvotes. The GitHub repo has 2 stars: code is public but adoption is still early. The GPG and ED metrics are defined by the researchers themselves, and the claim of surpassing “all open-weight models” should be read carefully until the community replicates the results.

For those using AI in their work, the practical angle is clear. If you generate images with prompts, experiment with text structure: separate semantic and geometric description, use explicit annotations. The principle that organizing context matters more than its quantity is the same underlying our playbook on getting the context right first time, where the order and structure of what you give the model makes the difference between a useful answer and one that needs fixing.