Fang, Yu, Duan et al., CUHK, HKU, BUAA and Alibaba (2025) — ARXIV-2509.09680
Key takeaways
- The dataset: 6M images generated by FLUX with 20M English and Chinese descriptions, grouped by six characteristics, each with a step-by-step “Generation Chain-of-Thought” description. 15,000 A100 GPU-days to curate
- The benchmark: PRISM-Bench, seven tracks including a long-text track built from the GCoT prompts, scored for alignment and aesthetics by VLM judges. 19 models evaluated
Standing in the anthology
Filed from #163, which lists the dataset, this paper, and
Supra2-IMG, a small text-to-image model trained on it and released as a
model card with no paper. The model is not filed, because there is no
document to cite.
Skimmed, abstract only. No NOTE and no practice. Distilling one model’s outputs into training data, and judging with VLMs, both bear on practices the record may want, but this reading cannot support one.