TEVI: Text-Conditioned Editing of Visual Representations via Sparse Autoencoders for Improved Vision-Language Alignment
About
Vision-language models such as CLIP are highly useful for diverse tasks due to their shared image-text embedding space. Despite this, the image and text embeddings are often poorly aligned, affecting downstream performance. Recent work has shown that this can be attributed to an information imbalance: images contain more information than their captions describe. In this work, we propose TEVI, a framework that uses captions as a signal for what to retain from image embeddings. Specifically, we use sparse autoencoders to disentangle image embeddings and train a masking module to selectively reconstruct the embedding based on a given caption. In a controlled setup with synthetic captions, we show that TEVI is effective at preserving caption-described attributes while discarding others. By applying TEVI to CLIP models trained on natural images, we further achieve improved retrieval performance across coarse-grained short-caption (MS COCO, Flickr) and fine-grained long-caption (IIW, DOCCI) benchmarks, with stronger gains on richer captions, and improved robustness on the RoCOCO benchmark.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Image Retrieval | MS-COCO | R@136 | 174 | |
| Image-to-Text Retrieval | DOCCI | R@126.06 | 66 | |
| Text-to-Image Retrieval | DOCCI | Recall@19.26 | 66 | |
| Image-to-Text Retrieval | MS COCO short-caption (test) | R@137.88 | 10 | |
| Image-to-Text Retrieval | Flickr30k short-caption (test) | Recall@165.48 | 10 | |
| Text-to-Image Retrieval | MS COCO short-caption (test) | Recall@125.04 | 10 | |
| Text-to-Image Retrieval | Flickr30k short-caption (test) | Recall@147.14 | 10 | |
| Retrieval | COCO | R@136.46 | 4 | |
| Retrieval | RoCOCO | Rand-voca R@126.32 | 4 | |
| Robust Image Retrieval | RoCOCO Rand-voca | R@126.14 | 2 |