TEVI: Text-Conditioned Editing of Visual Representations via Sparse Autoencoders for Improved Vision-Language Alignment
For practitioners using CLIP models, TEVI offers a method to improve alignment and retrieval performance without retraining, with gains on both coarse- and fine-grained benchmarks.
TEVI uses sparse autoencoders and a masking module to selectively retain caption-described information from CLIP image embeddings, improving vision-language alignment. It achieves improved retrieval performance on MS COCO, Flickr, IIW, DOCCI, and RoCOCO benchmarks, with stronger gains on richer captions.
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.