What it is
Measures how caption information, not length, scales diffusion training, then uses structured JSON prompts plus a trained LLM prompter to improve text-to-image generation. Instead of longer prose prompts, it uses image-grounded structured JSON prompts whose informativeness predicts diffusion loss. A separate LLM prompter learns to map user text into this schema, improving both training-side supervision and inference-time prompting.
Key results
- Diffusion loss vs GPG fits linearly with r=0.984 across controlled runs
- SP reconstructions improve to DINOv3 0.65 vs 0.51 for long NL captions
- SP reconstructions cut LPIPS to 0.59 vs 0.68 for long NL captions
Shows that image-grounded prompt information, not token count, is the variable that scales text conditioning for diffusion models.