Brief

Researchers have developed a new method called multi-level Floyd-Steinberg error-diffusion dithering to enhance the adversarial robustness of vision foundation models. This technique acts as an input transformation that disrupts adversarial attacks while maintaining the semantic content of the images. Tested across various tasks and model families, the dithering method, particularly with intermediate quantization and post-processing blur, demonstrated superior or comparable performance to existing baselines with less degradation on clean inputs. AI

NVIDIA's Pixel Diffusion Decoder (PiD) approach is being integrated into ComfyUI through custom nodes, enabling a combined decode and upscale process. This method treats latent-to-image decoding as conditional pixel diffusion, offering improved quality for higher resolutions. The experimental nodes support various NVIDIA checkpoints and include features for lower VRAM usage and text prompt assistance. AI

A new benchmark study explores how to best incorporate geographic location data into AI models for satellite image analysis. Researchers tested three methods—naive sin/cos, GeoCLIP, and SatCLIP—to encode latitude and longitude, finding that while naive sin/cos produced the most geographically coherent embeddings, SatCLIP offered a better balance for land-cover classification. The study used a DINOv2 vision model and the EuroSAT dataset to evaluate the effectiveness of these location encoders. AI

Meta AI's open-source computer vision model, DINOv2, is being used by Forest Research in the UK to improve the accuracy of tree canopy mapping. This collaboration aims to support the UK government's environmental goals, including increasing tree canopy cover and ensuring access to greenspaces. The DINOv2 model, trained on millions of satellite images, enables the detection of individual trees at a global scale, offering a more cost-effective and precise alternative to traditional monitoring methods like LiDAR. AI

Researchers have developed a three-stage framework to assess nursing student competency using egocentric video from simulation exercises. The system extracts action timelines and sequence-level features from video, then correlates these with instructor-rated competency. Surprisingly, higher recognition accuracy of actions in the video correlated negatively with student competency, suggesting that more skilled students perform diverse, less predictable actions that are harder for the AI to classify. AI

Researchers have developed UniRefiner, a framework designed to improve the spatial accuracy of Vision Transformer (ViT) models. This method teaches pre-trained ViTs to identify and discard irrelevant or spurious tokens that can degrade performance on spatially sensitive tasks. By using contrastive registers and a dual objective, UniRefiner refines diverse ViTs with minimal fine-tuning, leading to significant improvements in tasks like semantic segmentation. AI

Hugging Face has launched new features for PapersWithCode, a platform tracking AI state-of-the-art. The updates include support for multiple metrics on leaderboards, such as for Automatic Speech Recognition and Object Detection. The platform now also accommodates external papers beyond arXiv, automatically enriching them with relevant tags and data, and displays paper lineage to show follow-ups or predecessors. AI

Researchers have developed a new framework to measure the human interpretability of vision foundation models. This framework uses two protocols: localizability, which assesses an observer's ability to predict where a feature fires on an image, and nameability, which evaluates how accurately an observer can describe what a feature represents. When applied to six vision transformers, including DINOv2, DINOv3, CLIP, and SigLIP, the study found that foundation models are consistently less interpretable than supervised models, and this difference is not due to a capability tradeoff. AI