One Patch is All You Need: Joint Surface Material Reconstruction and Classification from Minimal Visual Cues

The introduction of DinoLizer, a model based on DINOv2, aims to enhance the localization of manipulated regions in generative inpainting. By utilizing a pretrained DINOv2 model on the B-Free dataset, it incorporates a linear classification head to predict manipulations at a granular patch resolution, employing a sliding-window strategy for larger images. This method shows superior performance compared to existing local manipulation detectors across various datasets.

LLaVA-UHD v3 has been introduced as a new multi-modal large language model (MLLM) that utilizes Progressive Visual Compression (PVC) for efficient native-resolution encoding, enhancing visual understanding capabilities while addressing computational overhead. This model integrates refined patch embedding and windowed token compression to optimize performance in vision-language tasks.

A new automated histopathologic assessment framework for Hirschsprung Disease has been developed using a multi-stage Vision Transformer approach. This framework effectively segments the muscularis propria, delineates the myenteric plexus, and identifies ganglion cells, achieving a Dice coefficient of 89.9% and a Plexus Inclusion Rate of 100% across 30 whole-slide images with expert annotations.

A recent study has introduced a modular, on-site solution for sustainable nutrient management in agriculture, utilizing lightweight anomaly detection techniques to optimize nutrient consumption and enhance crop growth. The approach employs a tiered pipeline for status estimation and anomaly detection, integrating multispectral imaging and an autoencoder for early warnings during nutrient depletion experiments.

Recent advancements in the optimization of Vision Transformers (ViTs) have been achieved through the integration of Decorrelated Backpropagation (DBP) into Masked Autoencoder (MAE) pre-training, resulting in a 21.1% reduction in wall-clock time and a 21.4% decrease in carbon emissions during training on datasets like ImageNet-1K and ADE20K.

A new architecture named HCFSSNet has been introduced, combining Convolutional Neural Networks (CNNs) with a Vision Frequency State Space block to enhance learned image compression (LIC). This hybrid approach captures local high-frequency details while effectively modeling long-range low-frequency information, addressing limitations seen in traditional methods.

A new method for classifying glioblastoma subregions using a contrastive learning-based encoder has been developed, achieving notable performance metrics in the BraTS-Path 2025 Challenge. The model, which fine-tunes a pre-trained Vision Transformer, secured second place with an MCC of 0.6509 and an F1-score of 0.5330 on the final test set.

CAMformer has been introduced as a novel accelerator that reinterprets attention mechanisms in Transformers as associative memory operations, utilizing a Binary Attention Content Addressable Memory (BA-CAM) to enhance energy efficiency and throughput while maintaining accuracy. This innovation addresses the scalability challenges faced by traditional Transformers due to the quadratic cost of attention computations.