SynthPix: A lightspeed PIV image generator

A research team from ETH Zurich, including Antonio Terpin and Raffaello D'Andrea, has released SynthPix, a high-performance synthetic image generator for Particle Image Velocimetry (PIV). Built using Google's JAX framework, the tool is engineered for massive parallelism on modern accelerators like GPUs and TPUs. It generates realistic pairs of particle-seeded flow images from user-defined velocity fields, with full control over experimental parameters such as particle seeding density, image size, illumination non-uniformity, noise, and motion blur. This allows researchers to create perfectly ground-truthed datasets for developing and benchmarking computer vision algorithms that measure fluid flows.

Unlike traditional offline dataset generation that creates massive static files, SynthPix is architected to stream synthetic images directly into machine learning training loops on-demand. This 'lightspeed' generation capability is a breakthrough for data-hungry techniques like deep learning for PIV and enables previously impractical closed-loop procedures. Researchers can now perform adaptive sampling, where an AI model requests specific types of challenging flow data during training, or engage in acquisition/parameter co-design to optimize real-world experiment setups. The tool supports diverse scenarios, from controlled lab flows to environmental applications like riverine velocimetry, and allows for rapid sweeps over 'nuisance factors' to systematically evaluate algorithm robustness.