DEADiff: 一种具有解耦表征的高效风格化扩散模型
DEADiff: An Efficient Stylization Diffusion Model with Disentangled Representations
AI总结：
针对现有风格迁移方法损害文本可控性的问题，提出DEADiff模型，通过风格语义解耦机制和非重建学习，实现风格化效果与文本可控性的最优平衡。
摘要：
基于扩散的文本生成图像模型在迁移参考风格方面具有巨大潜力。然而，当前基于编码器的方法在迁移风格时会严重损害文本生成图像模型的文本可控性。本文提出DEADiff模型，通过以下两种策略解决该问题：1）参考图像风格与语义的解耦机制。首先由受不同文本描述引导的Q-Formers提取解耦后的特征表征，随后将这些特征注入互斥的交叉注意力层子集，以实现更好的解耦效果。2）非重建式学习方法。采用成对图像而非相同目标训练Q-Formers，其中参考图像与真实图像具有相同的风格或语义。定量和定性实验结果均表明，DEADiff取得了最优的视觉风格化效果，在文本生成图像模型固有的文本可控性与参考图像的风格相似度之间实现了最佳平衡。项目页面为https://tianhao-qi.github.io/DEADiff/。
英文摘要：
The diffusion-based text-to-image model harbors immense potential in transferring reference style. However, current encoder-based approaches significantly impair the text controllability of text-to-image models while transferring styles. In this paper, we introduce DEADiff to address this issue using the following two strategies: 1) a mechanism to decouple the style and semantics of reference images. The decoupled feature representations are first extracted by Q-Formers which are instructed by different text descriptions. Then they are injected into mutually exclusive subsets of cross-attention layers for better disentanglement. 2) A non-reconstructive learning method. The Q-Formers are trained using paired images rather than the identical target, in which the reference image and the ground-truth image are with the same style or semantics. We show that DEADiff attains the best visual stylization results and optimal balance between the text controllability inherent in the text-to-image model and style similarity to the reference image, as demonstrated both quantitatively and qualitatively. Our project page is https://tianhao-qi.github.io/DEADiff/.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
