Concept Brittleness

Overview概览

Text-to-image diffusion models usually follow prompts well, yet we found an overlooked failure: a concept such as “ink sketch” works for most objects but fails again and again on a few — Flux draws grapes as an ink sketch but keeps painting a cake in full color. These failures persist across random seeds, revealing systematic blind spots in how the model has learned the concept rather than sampling noise. We train sparse autoencoders at every denoising step, compare successful and failed generations in this sparse space where concepts become separable, locate the missing or weakened concept dimensions, and repair them at inference time by nudging features toward a concept prototype built from reliable samples, with no retraining. Across multiple diffusion backbones this markedly improves concept consistency, text fidelity and repair success, and the earlier the intervention, the stronger the correction.

文生图模型通常能很好地按提示词作画，但我们发现了一种被忽视的失败：同一个概念（如“水墨素描”）用在大多数物体上都没问题，却在少数物体上反复失败——Flux 能把“葡萄”画成水墨素描，“蛋糕”却总是彩色的。这类失败换随机种子也照样出现，说明模型对概念的掌握存在系统性盲区，而不是偶然的采样噪声。我们在去噪的每一步训练稀疏自编码器，在这个概念更易区分的稀疏空间里对比成功与失败的生成，找出缺失或减弱的概念维度，再在推理时把特征向由可靠样本构建的概念原型轻量插值，无需重新训练即可修复。在多种主流扩散模型上，概念一致性、文本忠实度和修复成功率都显著提升，而且越早干预、纠正效果越好。

Highlights要点

1. A concept keeps failing on a few objects whatever the seed: a systematic blind spot, not sampling noise.同一个概念在少数物体上反复失败，换随机种子也一样：这是系统性盲区，而不是采样噪声。
2. Sparse autoencoders at every denoising step locate the missing or weakened concept dimensions.在去噪的每一步训练稀疏自编码器，定位缺失或减弱的概念维度。
3. No retraining: concept consistency and repair rates rise markedly, and the earlier the intervention, the stronger the correction.无需重新训练，概念一致性和修复成功率显著提升，而且越早干预、纠正效果越好。

Citation引用

@inproceedings{yuan2026concept,
  title     = {Interpreting Object-Dependent Concept Brittleness in Text-to-Image Diffusion Models},
  author    = {Yifan Yuan and Xiangyu Liu and Hongming Shan and Yu Han and Yu Jiang and Hao Tan and Junping Zhang and Linlin Shen},
  booktitle = {Proceedings of the ACM International Conference on Multimedia},
  year      = {2026},
  url       = {https://arxiv.org/abs/2609.09909}
}