Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

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.

Sweta Mahajan, Sukrut Rao, Jiahao Xie, Alexander Koller, Bernt Schiele• 2026

Related benchmarks

TaskDatasetResultRank
Image RetrievalMS-COCO
R@136
174
Image-to-Text RetrievalDOCCI
R@126.06
66
Text-to-Image RetrievalDOCCI
Recall@19.26
66
Image-to-Text RetrievalMS COCO short-caption (test)
R@137.88
10
Image-to-Text RetrievalFlickr30k short-caption (test)
Recall@165.48
10
Text-to-Image RetrievalMS COCO short-caption (test)
Recall@125.04
10
Text-to-Image RetrievalFlickr30k short-caption (test)
Recall@147.14
10
RetrievalCOCO
R@136.46
4
RetrievalRoCOCO
Rand-voca R@126.32
4
Robust Image RetrievalRoCOCO Rand-voca
R@126.14
2
Showing 10 of 13 rows

Other info

Follow for update