AI Creates More Realistic Multi-Person Images

Researchers at Cornell have developed a new AI approach for generating more realistic images involving multiple people interacting with one another. While existing image-generation models can create convincing individual people, they often struggle to represent complex interactions accurately. The Cornell method addresses this through iterative pose-image generation, progressively constructing a scene one person at a time. Each predicted pose helps guide the generation of subsequent people, allowing the system to better capture spatial relationships and interactions without requiring users to manually specify poses. The approach uses the FLUX image-generation model as its foundation and combines pose detection with a multimodal large language model to organise descriptions, poses and spatial regions for each person.

The researchers also introduced DrawWaldoWorlds, a benchmark designed to evaluate whether image-generation systems correctly represent not only multiple individuals but also their roles and relationships, essentially testing whether the model understands who does what to whom. Experiments showed that the new approach produced more faithful multi-person scenes than existing methods. In a user study involving 20 participants, images generated using the Cornell method were preferred roughly two-to-one over images produced by two versions of FLUX. The researchers argue that automatically incorporating pose information could make generative AI considerably better at depicting complex social activities, sports, group scenes and other situations where realistic human interaction is essential.

More information:

https://news.cornell.edu/stories/2026/07/strike-pose-creating-more-realistic-multi-person-images

Ultra-Low-Power Face Detection Chip

Nvidia researchers have developed an ultra–low-power, always-on face detection system-on-chip (SoC) capable of identifying human faces in under a millisecond, addressing a key challenge in continuous computer vision: energy consumption. Traditional vision systems can require around 10 watts, which is too high for constant operation, but this chip uses less than 5 milliwatts while maintaining about 99% detection accuracy. It achieves this by activating only briefly (processing each frame in microseconds) and remaining fully powered for just a small fraction of time, enabling efficient real-time performance.

The system’s efficiency comes from a specialized architecture called Alpha-Vision, which combines a lightweight CPU, a deep-learning accelerator, and local SRAM memory to avoid costly data transfers. By storing data locally and using a race-to-sleep strategy (quickly completing computations and then entering low-power mode) it minimizes energy use even further. This design enables practical applications such as laptops that automatically turn screens on/off based on user presence, as well as always-on vision in robotics, drones, and autonomous vehicles, where continuous sensing must not drain power.

More information:

https://spectrum.ieee.org/face-recognition-nvidia-chip-soc

IEEE Access 2026 Article

Recently, I co-authored an open-access journal paper that was published at IEEE Access. The paper is entitled “An Augmented Reality System With an Offline LSTM-Based Fault Recognition Model for Sewer Pipeline Inspection”. The paper introduces XR5.0, a novel framework that combines artificial intelligence with extended reality (XR) technologies to support the vision of Industry 5.0, where advanced digital systems are designed around human needs and capabilities. The research proposes a human-centric XR paradigm that integrates immersive environments with AI to enhance collaboration between workers and intelligent machines.

A key component of the approach is the use of human-centred digital twins, which create digital representations of users to enable XR systems to adapt training, guidance, and information delivery according to individual skills, context, and tasks. The framework also integrates advanced AI techniques (including explainable AI, generative AI, active learning, and neurosymbolic AI) to provide real-time decision support and personalized learning within immersive environments. These capabilities enable practical applications such as industrial training, remote maintenance, assembly guidance, and product design simulations.

More information:

https://ieeexplore.ieee.org/abstract/document/11363212

PNG Updated

The World Wide Web Consortium (W3C) just released an update to the Portable Network Graphics (PNG) image format this week. The W3C introduced the initial specifications for this file format in 1996, with the group formalizing the 2003 standard in 2004.

The third edition of the PNG specification introduces three new features: support for High Dynamic Range (HDR) images, animation through Animated PNGs (APNGs), and Exif data storage.

More information:

https://www.tomshardware.com/software/png-has-been-updated-for-the-first-time-in-22-years-new-spec-supports-hdr-and-animation

EdgeFace Face Recognition Model for Edge Devices

EdgeFace is a novel face recognition model optimized for use on devices with limited processing power and storage. This lightweight face recognition network is inspired by the hybrid architecture of EdgeNeXt, which combines the strengths of convolutional neural networks (CNNs) and transformers to perform accurate face recognition while conserving computational resources.

However, with the introduction of newer models like vision transformers (ViTs), the technology shows a promising approach to enhancing face recognition by effectively capturing long-range interactions. Researchers have extended the existing EdgeNeXt architecture tailored for face recognition on next-generation edge devices.

More information:

https://www.biometricupdate.com/202402/researchers-develop-edgeface-a-face-recognition-model-for-resource-constrained-edge-devices