Post-training is crucial for improving text-to-image models, yet existing pipelines typically rely on manually curated training data and fixed training distr…
机构:Microsoft
来源:arXiv 2610.18097 | AI4Papers 论文推荐平台