Scaling Properties of Text Conditioning in Visual Generation
The paper claims diffusion loss improves predictably when prompts contain more structured language.
The authors measure that structure with two metrics, GPG and ED, and report scaling behavior across controlled training runs. They use the finding to build prompts with semantic and geometric annotations derived from images. They also train a prompter through supervised fine-tuning, cold-start, and verifier-gated on-policy distillation. The resulting system is reported to beat evaluated open-weight models on most benchmarks and match or surpass top closed-weight models on many of them. HF Daily Papers' note
The authors measure that structure with two metrics, GPG and ED, and report scaling behavior across controlled training runs. They use the finding to build prompts with semantic and geometric annotations derived from images. They also train a prompter through supervised fine-tuning, cold-start, and verifier-gated on-policy distillation. The resulting system is reported to beat evaluated open-weight models on most benchmarks and match or surpass top closed-weight models on many of them. HF Daily Papers' note
score 5