Black Forest Labs unveils FLUX 3, a unified multimodal foundation model that jointly learns from images, videos, and audio. It excels in video generation with native audio up to 20 seconds, outperforming competitors in early evaluations. The model also offers improved image generation and action prediction capabilities, with gradual rollout of APIs, private weights, and open-source access.

Codex Slides is an open-source AI slide studio integrated into Codex, turning prompts, repos, or files into beautiful presentation decks. It generates 10+ slides in ~4-5 minutes with fast parallel rendering, offers 45 templates, 73 community styles, and 24 guided scenarios, exports to PPTX/PDF, and requires zero extra API keys.

On July 24, 2026, ETH Zurich, EPFL, and CSCS released Apertus 1.5, the latest version of one of the world's largest fully open language models. New capabilities include multimodal understanding (images and audio), improved reasoning, instruction following, and tool use. The model is already deployed in government translation, medical AI, and journalism. The project aims to build sovereign AI infrastructure as an open alternative to proprietary models.

In this tutorial, we build a complete workflow for running Baidu’s Unlimited-OCR model on document images and multi-page PDFs. From configuring the GPU environment to comparing high-detail tiled Gundam inference and faster Base modes, you'll learn how to process dense layouts, tables, and cross-page content in a reproducible, end-to-end pipeline.

MIT is spending over $3 million to install more than 500 AI-powered surveillance cameras across campus, capable of real-time face and object detection, as part of a multi-phase security expansion. The project, which began in November 2025, has sparked privacy concerns among students, especially amid broader debates about surveillance on university campuses.

At the AI Summit in San Francisco, South Korean President Jae Myung Lee and top business leaders meet with NVIDIA to advance Korea's AI infrastructure. Key announcements include a joint AI lab with KAIST and expanded collaboration with SK Group.

cloro.dev, the leading AI UI scraping platform, introduces a new recurring free tier offering 500 credits every month. The platform extracts structured data from ChatGPT, Perplexity, Grok, Gemini, Google Search, Google News, Copilot, and AI Overview via a single API. It has crossed 200 customers, mostly large enterprises.

Black Forest Labs launches FLUX 3, a unified multimodal model covering image, video, audio, and action prediction. FLUX-mimic, built on FLUX 3, enables video-action models for robotics. Also covers open data release The Stack v3, distillation debate, audio/TTS systems, agent infrastructure, and OpenAI product updates.

LOCK TF IN is a satirical AI fitness companion that acts as a no-excuses drill sergeant, roasting your excuses to keep you accountable. It features a streak engine, training log, themes, and a privacy-first free tier with on-device AI. Pro subscription unlocks cloud AI, advanced analytics, and more.

FLUX 3 is a new AI visual model on Textideo that supports text-to-image and image-to-video generation, with advanced prompt understanding and high-quality output. Designed for creators, marketers, designers, and businesses, it can be used for product photography, character design, social media content, marketing visuals, and more. FLUX 3 helps users create realistic visuals quickly using simple natural language prompts, reducing costs and improving efficiency.

This research explores using a VLA with a Mixture-of-Experts (MoE) action head to learn compositional robot policies end-to-end from demonstrations, without predefined task decomposition. The experts are reused across tasks and correspond to distinct low-level behaviors, with the router implicitly performing high-level sequencing, achieving performance on par with monolithic baselines while demonstrating interpretable specialization.

This paper presents a real-time RGB-D perception pipeline for automating hydraulic impact hammers in mining. It combines instance segmentation with point cloud processing, runs at ~10 Hz on embedded hardware with ~675 ms latency, and generates rock-breaking poses and a robot-free 3D workspace representation. Experimental results in a scaled scenario show suitability for real-time autonomous operation.

This work tests a neuro-inspired self-supervised learning framework for trajectory planning that uses forward and inverse models as internal supervision. Experiments show feasibility but reveal a tendency to exploit the learning signal, leading to proposed mitigation strategies.

With the increasing use of UAVs in military operations, the demand for defense systems against UAV attacks has grown. This paper proposes a method for ground target protection using autonomous UAV swarms in GPS- and communication-denied environments. The approach relies solely on onboard sensors and relative measurements for target tracking and swarm coordination, employing a decentralized encirclement technique that adapts to target motion. Experiments with real robots validated its effectiveness in detecting, encircling, and intercepting hostile UAVs.

A new framework called FELT synthesizes per-finger pressure tactile images from RGB observations, reducing the need for tactile sensor data. It uses a frozen visual encoder and lightweight query decoder to predict tactile signals in one forward pass, improving policy success on contact-rich tasks.

Researchers propose the Capability-Aware Traversability (CAT) framework, which embeds robot physical limits directly into the spatial feature space using an interactive annotation pipeline and Spatially-Adaptive Denormalization (SPADE) blocks. CAT achieves significant improvements in traversability estimation, with AUROC up by 11.0% and AUPRC up by 15.8% on benchmark datasets, and demonstrates real-time embodiment-aware obstacle avoidance at 4.8 Hz on legged and wheeled platforms.

This work proposes a learning-based framework for safe and scalable multi-drone cooperative payload transport. It introduces a minimal 2D abstraction preserving task-relevant coupling, and trains a distributed policy via Discrete Graph Control Barrier Function Proximal Policy Optimization (DGPPO) with domain randomization, enabling zero-shot sim-to-real transfer. Real-world experiments demonstrate generalization across team sizes and dynamic environments.

Researchers present RAINBOTTM, a low-cost liquid-handling robot built from a consumer 3D printer with a digital twin for remote supervision, costing under $1300, drastically lowering the barrier to lab automation.

PhysCoRe is a physics-corrected residual world model that couples a differentiable Material Point Method (MPM) simulator with two feed-forward neural networks to infer per-particle elasticity from visual observations and learn corrections for simulator biases, significantly improving prediction accuracy for deformable objects under robotic manipulation and enabling online material identification with confidence-guided exploration.

This study introduces HERMES, a heterogeneous edge-relational graph neural network with SSM-informed multi-head attention, achieving AUC-ROC of 0.9898 on traffic conflict prediction at signalized intersections, outperforming Transformer baselines and demonstrating strong transferability.

This study evaluates ISO/IEC 29794-6 iris quality metrics on VR-acquired data, finding some metrics (e.g., margin adequacy) fail; addresses off-axis gaze, non-uniform illumination, and specular reflection using generative models; and achieves ~11% EER reduction via multimodal fusion of iris and periocular.

3D-GIMP presents a novel hybrid paradigm for high-fidelity object removal in 3D Gaussian Splatting. Instead of diffusing every view, it inpaints a single key reference view and propagates textures via a 3D-aware PatchMatch algorithm, avoiding multi-view inconsistencies and hallucination drift. Experiments show competitive inpainting quality with faster rendering and better view consistency.

This paper identifies 'definition blindness' in current open-world video anomaly detection (OWVAD) evaluation, where models become insensitive to the user-specified definition of abnormality. The authors introduce three new metrics (DC-Disc, DC-DetΔ, DC-SelΔ) and a definition-contrastive scoring rule (DeCoS) that significantly improve definition following.

U-CFR is a novel inference-time framework for interactive segmentation that autonomously self-corrects after each user interaction. It introduces a boundary-aware uncertainty score fusing segmentation uncertainty, contour gradients, and explicit edge predictions to guide internal pseudo-clicks. A dual-head network with segmentation and edge heads enables cascade refinement steps that progressively improve masks. Experiments show over 10% click reduction on challenging datasets like Berkeley.

A new framework for privacy-preserving lung CT slice generation combines Optimal Transport Conditional Flow Matching with geometric filtering, achieving PPS 0.549 and FID 0.3290, but patient re-identification remains a challenge.

This paper introduces Spatially Grounded Concept Bottleneck Models (SG-CBM), which use coarse lesion delineations as weak supervision to encourage anatomically plausible concept evidence in breast ultrasound diagnosis. The model defines two clinically relevant zones from lesion masks: in-lesion region for morphology and posterior acoustic band for posterior phenomena. Results show improved diagnostic AUROC and concept macro-AUROC with better spatial alignment. A Train-corrupt/Test-clean stress test highlights the impact of supervision quality on model trustworthiness.

ODeform introduces a novel extension of Neural Ordinary Differential Equations to model continuous 4D dynamics of deformable objects in 3D space. It transforms 3D point clouds and physical conditions into a unified latent space and solves ODEs over time for continuous deformation flows, eliminating discrete time steps while maintaining efficiency. Experiments show improved motion prediction accuracy on unseen physical parameters and successful transfer to real 3D objects. The code and data will be made public.

Axolotl3D is a multi-modal, occlusion-aware 3D generation model that jointly conditions on images, visibility masks, camera parameters, and a partial point cloud, enabling faithful shape completion. It achieves state-of-the-art performance on Toys4K and OmniObject3D under both clean and occluded settings.

The paper introduces Masked Topology Modeling (MTM), a self-supervised pretraining task that leverages the face-adjacency graph unique to B-Reps. By masking edges and predicting their convexity and curve type, combined with MoCo contrastive learning and BFS-connected region reconstruction, the method achieves strong performance on several benchmarks after pretraining on ABC and a new procedural dataset.

RealVDeblur is an efficient generative framework that addresses real-world video deblurring by synthesizing realistic training data via a physically grounded blur pipeline using 3D Gaussian Splatting, and leveraging a video diffusion prior with one-step distillation for robust inference on long videos.