Scaling Properties of Text Conditioning in Visual Generation

Researchers Zilong Chen and colleagues at ByteDance Seed study why simply making text prompts longer does not reliably improve text-to-image generation. Across 15 controlled caption configurations, they show that image-grounded information—not token count—predicts diffusion training quality. Their typed JSON structured prompt represents global scene context, object attributes, bounding boxes, depth, and relationships; information is measured with Grounded Perplexity Gain, or GPG, and Effective Detailness, or ED. On the BAGEL backbone, converged diffusion loss correlates with GPG at -0.984 and with ED at -0.971, while both measures rank caption configurations consistently. The system then combines structured supervision for higher diffusability with a Qwen3.5 prompter trained through supervised fine-tuning, cold-start distillation, and verifier-gated on-policy self-distillation, or OPSD. With Qwen-Image as the diffuser, the resulting ByteDance Seed system reaches 72.5 GenEval2 geometric mean and 85.2 on CoReBench, exceeding the matched natural-language Qwen-Image control, which scores 56.2 and 76.1. Zero-shot prompter scaling also transfers: thinking-mode GenEval++ rises from 46.4 percent with a 0.8B Qwen3.5 model to 86.8 percent with 397B. On broader comparisons, the system surpasses open-weight models and matches or exceeds closed systems such as OpenAI’s GPT-Image-1 and Google’s Nano Banana on most evaluations. An inference-time refine–render–judge loop provides additional gains, but performance saturates after a few rounds, indicating that training the prompt interface is more valuable than merely adding inference compute.