Scaling Properties of Text Conditioning in Visual Generation

ByteDance Seed

Research official 2 src. ~1 min

Studies how diffusion loss scales with the amount of structured language in text prompts, introducing white-box (GPG) and black-box (ED) metrics to quantify prompt structure, and shows converged loss decreases roughly linearly with GPG and follows a power law with ED.

Why it matters

Fills a gap in scaling-law research for visual generation, where prompt-token scaling had rarely been measured; the resulting system is competitive with both open- and closed-weight models on compositional and reasoning benchmarks.

Importance: 2/5

Notable scaling-law study for visual generation, below the 100-upvote HF Daily bump threshold.

Sources