Abstract:Training Vision Transformers (ViTs) presents significant challenges, one of which is the emergence of artifacts in attention maps, hindering their interpretability. Darcet et al. (2024) investigated this phenomenon and attributed it to the need of ViTs to store global information beyond the [CLS] token. They proposed a novel solution involving the addition of empty input tokens, named registers, which successfully eliminate artifacts and improve the clarity of attention maps. In this work, we reproduce the findings of Darcet et al. (2024) and evaluate the generalizability of their claims across multiple models, including DINO, DINOv2, OpenCLIP, and DeiT3. While we confirm the validity of several of their key claims, our results reveal that some claims do not extend universally to other models. Additionally, we explore the impact of model size, extending their findings to smaller models. Finally, we untie terminology inconsistencies found in the original paper and explain their impact when generalizing to a wider range of models.
Abstract:Composed Image Retrieval (CIR) allows users to search for images by combining a reference image with a text prompt that describes desired modifications. While vision-language models like CLIP have popularized this task by embedding multiple modalities into a joint space, developers still lack tools that reveal how these multimodal prompts interact with embedding spaces and why small wording changes can dramatically alter the results. We present InfoCIR, a visual analytics system that closes this gap by coupling retrieval, explainability, and prompt engineering in a single, interactive dashboard. InfoCIR integrates a state-of-the-art CIR back-end (SEARLE arXiv:2303.15247) with a six-panel interface that (i) lets users compose image + text queries, (ii) projects the top-k results into a low-dimensional space using Uniform Manifold Approximation and Projection (UMAP) for spatial reasoning, (iii) overlays similarity-based saliency maps and gradient-derived token-attribution bars for local explanation, and (iv) employs an LLM-powered prompt enhancer that generates counterfactual variants and visualizes how these changes affect the ranking of user-selected target images. A modular architecture built on Plotly-Dash allows new models, datasets, and attribution methods to be plugged in with minimal effort. We argue that InfoCIR helps diagnose retrieval failures, guides prompt enhancement, and accelerates insight generation during model development. All source code allowing for a reproducible demo is available at https://github.com/giannhskp/InfoCIR.