Go to ICLR 2026 Conference homepageMaskInversion: Localized Embeddings via Optimization of Explainability Maps
Keywords: vision encoder, localized embedding, CLIP
TL;DR: This paper introduces MaskInversion, a method that creates a localized representation for a specific image region by optimizing an embedding so that a pre-trained model's focus aligns with a given mask, all without retraining the model itself.
Abstract: Vision-language foundation models such as CLIP have achieved tremendous results in global vision-language alignment, but still show some limitations in creating representations for specific image regions.
To address this problem, we propose MaskInversion, a method that leverages the feature representations of pre-trained foundation models, such as CLIP, to generate a context-aware embedding for a query image region specified by a mask at test time.
MaskInversion starts with initializing an embedding token and compares its explainability map, derived from the pretrained model, to the query mask.
The embedding token is then subsequently refined to approximate the query region by minimizing the discrepancy between its explainability map and the query mask. During this process, only the embedding vector is updated, while the underlying foundation model is kept frozen
allowing to use MaskInversion with any pre-trained model.
As deriving the explainability map involves computing its gradient, which can be expensive, we propose a gradient decomposition strategy that simplifies this computation.
The learned region representation can be used for a broad range of tasks, including open-vocabulary class retrieval, referring expression comprehension, as well as for localized captioning and image generation. We evaluate the proposed method on all those tasks on several datasets such as PascalVOC, MSCOCO, RefCOCO, and OpenImagesV7 and show its capabilities compared to other SOTA approaches.
Primary Area: transfer learning, meta learning, and lifelong learning
Submission Number: 7808
Loading