内在分布匹配:提升少步文本到图像生成
原标题:Paper page - Intrinsic Distribution Matching
AI 摘要
该论文提出 Intrinsic Distribution Matching(IDM),在分布匹配蒸馏(DMD)基础上加入对生成器 EMA 特征的均值与逐坐标标准差的显式矩约束,无需外部特征编码器。研究在 Z-Image-Turbo 6B 和 Qwen-Image 20B 上以全参数和 LoRA 微调验证,2 次网络评估下在 LongText-Bench 中英文分别达到 94.25% 和 90.20%,为同预算对比方法中最高。消融实验显示矩损失带来稳定增益,生成器 EMA 特征与 DINOv2 编码器表现相当,代码已开源。
正文节选
Intrinsic Distribution Matching Abstract Intrinsic Distribution Matching improves few-step text-to-image generation by adding explicit moment constraints to distribution matching distillation, achieving state-of-the-art long-text rendering with minimal network evaluations. One-step and few-step text-to-image generation must accurately render the details specified in the prompt despite limited opportunities for refinement. Distribution matching distillation (DMD) uses score differences to bring t