SynthPix: A lightspeed PIV images generator

TL;DR

<3-5 sentence high-level summary> SynthPix addresses the data bottleneck in learning-based PIV by delivering a GPU-accelerated, batched synthetic piv image generator implemented in JAX, with flexible configuration and seamless integration into modern ML pipelines. It presents a detailed particle-and-image formation model, efficient rasterization, and robust data-loading strategies, all designed to maximize throughput on accelerators. The paper validates SynthPix through throughput benchmarks, cross-tool comparisons, and estimator-based accuracy tests, demonstrating orders-of-magnitude improvements while preserving realism of the generated data. By standardizing synthetic piv data generation and emphasizing reproducibility, SynthPix enables scalable training of data-hungry methods and supports real-time flow-control research in fluid dynamics.

Abstract

We describe SynthPix, a synthetic image generator for Particle Image Velocimetry (PIV) with a focus on performance and parallelism on accelerators, implemented in JAX. SynthPix supports the same configuration parameters as existing tools but achieves a throughput several orders of magnitude higher in image-pair generation per second. SynthPix was developed to enable the training of data-hungry reinforcement learning methods for flow estimation and for reducing the iteration times during the development of fast flow estimation methods used in recent active fluids control studies with real-time PIV feedback. We believe SynthPix to be useful for the fluid dynamics community, and in this paper we describe the main ideas behind this software package.

Figures (5)

• Figure 1: The SynthPix pipeline (top) follows established PIV image-generation techniques but is optimized for performance on hardware accelerators. Below, we compare SynthPix’s throughput against existing synthetic particle-image generators (B: batch size; CPU/GPU: execution device) and summarize the key characteristics of the SynthPix codebase; see \ref{['sec:results']}. The rendered images are zoomed in and white on black (instead of black on white) here and throughout the paper for visualization purposes.
• Figure 2: Using SynthPix to instantiate the image generator, get a new batch of piv images, the corresponding groundtruth flows and the sampled parameters.
• Figure 3: Zoomed-in images generated with SynthPix (left) and PIVlab stamhuis2014pivlab (right). For each method we report images of resolution $256\times256$ and $512\times512$.
• Figure 4: Flows (left), original (middle) and Synthpix-generated (right) images from two datasets of cai2019deep.
• Figure 5: Results of the ablation studies in \ref{['sec:results:scaling-laws']}. Each plot shows how the throughput (in image pairs per second) changes with a single hyperparameter. The filled circles mark the mean over all the batches (we collect $1000$ batches for each hyperparameter). The black vertical error bars extend between the minimum and maximum observed values. The orange bands represent the mean $\pm$ one standard deviation (light) and the inter-quartile range (Q1‚ÄìQ3, dark).

Theorems & Definitions (2)

• Remark 2.1
• Remark 2.2