Massimiliano Mancini

From Weights to Concepts: Data-Free Interpretability of CLIP via Singular Vector Decomposition

Large Multimodal Models as General In-Context Classifiers

SEM: Sparse Embedding Modulation for Post-Hoc Debiasing of Vision-Language Models

Safe Vision-Language Models via Unsafe Weights Manipulation

On Large Multimodal Models as Open-World Image Classifiers

Automatic benchmarking of large multimodal models via iterative experiment programming

Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers

Compositional Caching for Training-free Open-vocabulary Attribute Detection

Not Only Text: Exploring Compositionality of Visual Representations in Vision-Language Models

3D Part Segmentation via Geometric Aggregation of 2D Visual Features

GradBias: Unveiling Word Influence on Bias in Text-to-Image Generative Models

Group-robust Machine Unlearning

Unlearning Personal Data from a Single Image

Harnessing Large Language Models for Training-free Video Anomaly Detection

MULTIFLOW: Shifting Towards Task-Agnostic Vision-Language Pruning

OpenBias: Open-set Bias Detection in Text-to-Image Generative Models

Less is more: Summarizing Patch Tokens for efficient Multi-Label Class-Incremental Learning

Frustratingly Easy Test-Time Adaptation of Vision-Language Models

Image-free Classifier Injection for Zero-Shot Classification

Iterative Superquadric Recomposition of 3D Objects from Multiple Views

On the Effectiveness of LayerNorm Tuning for Continual Learning in Vision Transformers

PDiscoNet: Semantically consistent part discovery for fine-grained recognition

ProbVLM: Probabilistic Adapter for Frozen Vison-Language Models

Transitivity Recovering Decompositions: Interpretable and Robust Fine-Grained Relationships

Vocabulary-free Image Classification