---
title: "ETH Zurich's SynthPix Removes the Biggest Bottleneck in AI Fluid Simulation"
subtitle: "A JAX-based synthetic Particle Image Velocimetry generator streams GPU-rendered image pairs on-the-fly, cutting terabyte storage costs and unlocking data-hungry flow estimators."
url: "https://alphasignal.ai/news/eth-zurich-s-synthpix-removes-the-biggest-bottleneck-in-ai-fluid-simulation"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-04-13T00:00:00.000Z"
updated: "2026-08-24T03:01:42.503Z"
type: "Paper"
topics:
  - "Data"
  - "Gpus"
subtopics:
  - "Synthetic Data"
  - "Quantization"
upvotes: 1749
source_url: "https://arxiv.org/abs/2512.09664"
comments: 0
truncated: true
---

# ETH Zurich's SynthPix Removes the Biggest Bottleneck in AI Fluid Simulation

> A JAX-based synthetic Particle Image Velocimetry generator streams GPU-rendered image pairs on-the-fly, cutting terabyte storage costs and unlocking data-hungry flow estimators.

Measuring how fluids move usually involves seeding a flow with tiny reflective particles, taking two closely spaced photographs, and computing how each particle shifted between frames. This technique, called Particle Image Velocimetry (PIV), is a workhorse in labs, wind tunnels, microfluidic chips, and even bridge-mounted cameras watching rivers. Modern PIV research has hit a wall that has nothing to do with fluid dynamics, though: the deep learning models people now want to train on synthetic PIV imagery are choking on the sheer size and rigidity of pre-rendered datasets.

A team at ETH Zurich has released [SynthPix](https://github.com/antonioterpin/synthpix), an open-source Python package that generates synthetic PIV images in JAX, designed for hardware accelerators and direct integration into learning and benchmarking pipelines. It supports the same configuration parameters as existing tools while achieving a throughput several orders of magnitude higher in image-pair generation per second.

## Why offline datasets stopped scaling

The problem is combinatorial. Robustness studies need to sweep over particle density, particle size, illumination unevenness, noise, blur, and inter-frame timing, all against varied flow fields. Modern simulation collections already span multi-terabyte regimes before any PIV rendering, as illustrated by The Well (15 terabytes across 16 datasets) and the Johns Hopkins Turbulence Database (well over 200 terabytes).

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [SynthPix: A lightspeed PIV image generator (arXiv)](https://arxiv.org/html/2512.09664)
- [SynthPix repository](https://github.com/antonioterpin/synthpix)
- [Flow Gym (arXiv)](https://arxiv.org/html/2512.20642v1)
- [SynthPix abstract listing](https://arxiv.org/abs/2512.09664)