SAEParate：面向概念分离扩散模型遗忘的解耦稀疏表示（NeurIPS 2026）
Disentangled Sparse Representations for Concept-Separated Diffusion Unlearning
AI 导读
SAEParate 是一种文本到图像扩散模型概念遗忘方法，通过概念感知对比学习在 SAE 潜空间中显式分离语义重叠概念，并借助 cross-attention 引导的空间聚合强调概念相关区域。该方法不改模型参数、以轻量方式抑制目标概念，还为编码器增加非线性变换以增强表达力。在 UnlearnCanvas 上达到 SOTA，尤其在联合风格-物体遗忘这一目标与非目标概念易互相干扰的困难场景中提升显著。
正文 · 原文
Abstract:Unlearning specific concepts in text-to-image diffusion models has become increasingly important for preventing undesirable content generation. Among prior approaches, sparse autoencoder (SAE)-based methods have attracted attention due to their ability to suppress target concepts through lightweight manipulation of latent features, without modifying model parameters. However, SAEs trained primarily for sparse reconstruction do not explicitly enforce concept-wise separation, allowing semantically related concepts to share latent features. To address this, we propose SAEParate, which explicitly separates semantically overlapping concepts in the SAE latent space through concept-aware contrastive learning, supported by cross-attention-guided spatial aggregation that emphasizes concept-relevant regions. In addition, we enhance the encoder with a nonlinear transformation to increase its expressive capacity under this separation objective, enabling more concept-discriminative latent representations. Experiments on UnlearnCanvas demonstrate state-of-the-art performance, with particularly strong gains in joint style-object unlearning, a challenging setting where existing methods suffer from severe interference between target and non-target concepts.
| Comments: | 53 pages, 25 figures. Accepted at NeurIPS 2026 |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computer Vision and Pattern Recognition (cs.CV) |
| Cite as: | arXiv:2605.12122 [cs.LG] |
| (or arXiv:2605.12122v3 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2605.12122 arXiv-issued DOI via DataCite |
Submission history
From: Hyeonjin Kim [view email]
[v1]
Tue, 12 May 2026 13:39:07 UTC (15,075 KB)
[v2]
Thu, 22 Oct 2026 09:02:42 UTC (17,884 KB)
[v3]
Mon, 26 Oct 2026 04:38:42 UTC (17,884 KB)
来源：arXiv cs.AI · arxiv.org