Publications

Large Language Models (LLMs) have revolutionized language modeling through autoregressive generation, enabling strong performance across a wide range of tasks. Recently, Diffusion Language Models (DLMs) have emerged as an alternative paradigm that generates text through iterative denoising rather than next-token prediction, allowing parallel refinement of entire sequences. While numerous diffusion-based architectures have been proposed, differences in evaluation protocols, datasets, inference budgets, and generation hyperparameters make it difficult to compare their capabilities and understand the trade-offs they offer. In this work, we present a systematic experimental analysis of modern DLMs. Specifically, we evaluate eight state-of-the-art DLMs across eight benchmarks spanning reasoning, coding, translation, knowledge, and structured problem solving, while explicitly considering both generation quality and computational efficiency. Beyond downstream evaluation, we analyze the impact of key inference-time factors, including denoising steps, context length, block size, and parallel unmasking strategies, and complement large-scale experiments with controlled comparisons of smaller models trained under identical conditions. Our analysis highlights the strengths and limitations of diffusion-based language modeling across different tasks, architectures, and inference budgets. We show that the behavior of DLMs is strongly influenced by generation-time design choices, leading to distinct trade-offs between performance and computational efficiency. Overall, our study provides practical insights into the capabilities and deployment characteristics of contemporary DLMs.

Progress in dental computer vision is limited by the absence of large-scale multimodal datasets that jointly capture 3D intraoral geometry and 2D appearance across diverse clinical settings. Existing resources are typically unimodal, which hinders robust cross-modal learning and generalization. We assemble and release a multi-center dataset of 1,000 patients comprising 2,000 registered upper/lower intraoral scans, 5,000 paired intraoral photographs, and 2,403 clinician-authored reports. This combination links detailed 3D dental geometry with complementary 2D evidence, supporting occlusal and orthodontic analysis. Moreover, to enable scalable and privacy-preserving acquisition and annotation across distributed centers, we introduce an open platform that supports multimodal ingestion and structured labeling. Experiments indicate that state-of-the-art multimodal models fail to generate clinically faithful reports, motivating geometry-aware adaptation. We therefore propose IOS-Qwen, which fuses a PointTransformer 3D encoder with Qwen3-VL to generate structured, point-cloud-conditioned reports. Together, the dataset, the platform, and the baselines establish a foundation for multimodal dental AI research. Code is publicly released (https://github.com/AImageLab-zip/IOS-Report)

Answer grounding in visual question answering over documents remains an open challenge: most existing benchmarks lack grounding annotations or provide annotations of limited quality, and the development of answer grounding datasets is still bounded to expensive manual effort. We introduce DocAttriBench (DAB), an improved large-scale benchmark for fine-grained, element-level source attribution in Document VQA, grounding each answer to specific layout elements such as text blocks, tables, and images. To build DAB, we propose a Mask-based Perplexity-Derived Attribution method (MAPPET) that leverages document layout and language modeling to assign each answer to its most informative document element. MAPPET measures the increase in perplexity when masking candidate elements, attributing the answer to the element that contributes most to model confidence. Applying MAPPET to multiple existing Document VQA datasets yields DAB, which contains 237k documents and 296k question-answer pairs, where each answer is grounded to the corresponding layout elements. We benchmark a range of Multimodal Large Language Models (MLLMs) with grounding capabilities on DAB, evaluating answer accuracy, attribution accuracy, and overall answer quality. Our results show that while larger models tend to achieve higher answer accuracy, even the most capable models often fail to correctly localize the supporting elements. DAB thus provides a scalable benchmark for developing grounded, verifiable, and trustworthy Document VQA models.

Federated continual learning (FCL) enables models to learn new tasks across multiple distributed clients, protecting privacy and without forgetting previously acquired knowledge. However, current methods face challenges balancing performance, privacy preservation, and communication efficiency. We introduce a Distributed Online LoRA for Federated INcremental learning methodDOLFIN, a novel approach combining Vision Transformers with low-rank adapters designed to efficiently and stably learn new tasks in federated environments. Our method leverages LoRA for minimal communication overhead and incorporates Dual Gradient Projection Memory (DualGPM) to prevent forgetting. Evaluated on CIFAR-100, ImageNet-R, ImageNet-A, and CUB-200 under two Dirichlet heterogeneity settings,DOLFINconsistently surpasses six strong baselines in final average accuracy while matching their memory footprint. Orthogonal low-rank adapters offer an effective and scalable solution for privacy-preserving continual learning in federated settings.

Recent advances in Virtual Try-On (VTON) and Virtual Try-Off (VTOFF) have greatly improved photo-realistic fashion synthesis and garment reconstruction. However, existing datasets remain static, lacking instruction-driven editing for controllable and interactive fashion generation. In this work, we introduce the Dress Editing Dataset (Dress-ED), the first large-scale benchmark that unifies VTON, VTOFF, and text-guided garment editing within a single framework. Each sample in Dress-ED includes an in-shop garment image, the corresponding person image wearing the garment, their edited counterparts, and a natural-language instruction of the desired modification. Built through a fully automated multimodal pipeline that integrates MLLM-based garment understanding, diffusion-based editing, and LLM-guided verification, Dress-ED comprises over 146k verified quadruplets spanning three garment categories and seven edit types, including both appearance (e.g., color, pattern, material) and structural (e.g., sleeve length, neckline) modifications. Based on this benchmark, we further propose a unified multimodal diffusion framework that jointly reasons over linguistic instructions and visual garment cues, serving as a strong baseline for instruction-driven VTON and VTOFF. Dataset and code available at this link: https://github.com/aimagelab/Dress-ED.

Modern Deep Neural Networks struggle to retain knowledge in streaming data environments, often leading to forgetting during incremental training. Most Continual Learning (CL) approaches address this issue by rehearsing past data – stored in a replay buffer – while acquiring new knowledge. However, in practical scenarios, noisy labels can contaminate the replay buffer, undermining performance. This work builds upon the previous “May the Forgetting Be with You”, designed to tackle Continual Learning with Noisy Labels (CLN). By leveraging the distinct learning dynamics between correctly and incorrectly labeled examples, the method induces targeted forgetting to identify and filter out noisy labels. We propose EARL, which improves on its predecessor by introducing i) a detailed analysis of the learning dynamics occurring in the presence of noise, ii) a robust analysis under more realistic noise conditions, iii) an evaluation of performance using pre-trained backbones and modern prompt-based CL baselines, iv) a detailed study on the influence of different sampling strategies, v) experiments on Natural Language Processing (NLP) benchmarks. This work unravels the motivations and findings of the previous research, shedding light on the effectiveness of its components in achieving high performance and minimizing forgetting.

Editing multiple elements of an image in a single forward pass is a practical alternative to multi-turn image manipulation, offering improved efficiency and potentially better harmonization. However, when several instructions target different regions, semantic interference often leads to attribute leakage and poor edit disentanglement, especially as the number of edits increases. In this work, we propose MICE (Multi-Instance Concurrent Editing), a training-free strategy for scalable multi-instance image editing with Multimodal Diffusion Transformers. MICE modifies the additive bias of joint attention to regulate interactions between instance-specific edit instructions, latent, and context tokens identified via user-provided segmentation masks. Specifically, MICE allows intra-instance attention, penalizes interactions between neighboring region tokens, and suppresses unrelated cross-instance attention. As a result, our method enforces attribute binding while preserving global visual consistency. We evaluate MICE on LoMOE-Bench and introduce MICE-Bench, a more challenging benchmark with an average of 8.5 concurrent edits per image. The experiments demonstrate that our approach outperforms strong baselines and recent competitors in terms of visual quality preservation and faithfulness to the editing instructions.

Visual Autoregressive (VAR) models face a severe "Memory Wall" on edge devices due to large model size and substantial KV-cache requirements. In this work, we analyze the Infinity VAR family (2B and 8B) and propose a compression pipeline for deployment on constrained NVIDIA Jetson systems. We diagnose critical bottlenecks: activation outliers reaching 353x the median and channel-skewed cache variance. To address this, we propose a hybrid pipeline combining SVDQuant—to structurally decouple weight outliers—and Asymmetric Per-Channel KV8 quantization. Our approach reduces the Infinity-8B footprint by 64% (37.1GB →13.3GB), fitting it on the mid-range Orin NX with a 4.1x speedup over Flux.1-dev (W4A4), while achieving superior aesthetic alignment (ImageReward 1.13 vs 0.935). Crucially, we also unlock entry-level feasibility for the Infinity-2B, compressing it from 16.0 to 7.71 GB to enable deployment on the Orin Nano. These results establish a new efficiency standard for high-fidelity generative AI at the edge. The code is available at https://github.com/Henvezz95/deepcompressor.

Facial age estimation has shown notable progress under controlled conditions. However, in unconstrained real-world environments, accurate age estimation remains challenging. This difficulty becomes more severe when facial images contain partial occlusions, as these obstructions hide important age-related information. Moreover, there is no publicly available occluded age estimation dataset to improve performance in real-world scenarios. To overcome this issue, we propose and publicly release three new datasets, FG-NET-O8, APPA-REAL-O8, and MORPH-O8, derived from existing benchmarks. These datasets contain eight types of realistic occlusions, providing a comprehensive testbed for age estimation under occlusions. These occlusions are generated using multiple diffusion-based methods, including Stable Diffusion Realistic Vision, Blended Latent Diffusion, and Fooocus, while preserving the facial identity of each subject. We also design and conduct a human survey to evaluate the quality of the generated occlusions. Furthermore, we test five state-of-the-art age estimation approaches to analyze the impact of real-world occlusions on age estimation performance. Experimental results demonstrate that all approaches exhibit severe performance degradation for nearly all occlusion types across all three datasets.