Fraunhofer’s RealOrRender spots deepfakes by recreating them

Fraunhofer IOSB has introduced RealOrRender, a deepfake detector that tries to recreate an image before deciding whether it was generated by AI. The system reportedly reaches up to 91% accuracy while showing which parts of an image influenced its verdict.

How RealOrRender detects synthetic images

RealOrRender combines two analysis methods. A deep-learning classifier examines the image directly, while a second model measures reconstruction error: the difference between the original and a new version generated by an AI system.

The underlying logic is unusual. If the system can reproduce an image particularly well, that may indicate the image was itself created by another generative model. Fraunhofer’s reported baseline accuracy is about 85%, rising above 91% in some cases.

The project also uses explainable AI to make its decisions easier to inspect. It marks suspicious regions with heat maps and image segments, highlighting textures, artifacts, and other features associated with generated images rather than returning only a binary result.

Research demonstrator, not a commercial product

RealOrRender was developed in Germany by Fraunhofer IOSB and funded by the German Federal Office for Information Security (BSI). Fraunhofer describes it as a research project and demonstrator running from September 2024 through May 2026. There is currently no announced commercial release, US pricing, or EU product pricing.

Recommended reading

Samsung’s zHBM concept promises 8x HBM5 speed

Fraunhofer has also not specified whether the technology will become a product or SaaS service, offer an API or SDK, support on-premises deployment, or release its source code or models. That makes the reported accuracy more notable as a research result than as a directly deployable alternative to existing detection services.

For comparison, Reality Defender’s RealScan lists a Business plan at $399 with annual billing, including 1,000 scans per month. Truepic Vision lists a Growth plan at $1,450 per month, or $17,400 per year billed monthly, while Sensity AI offers cloud and on-premises deployments through sales-led pricing. RealOrRender’s explainable output could be valuable for investigations, but until Fraunhofer clarifies distribution and licensing, there is no way to buy or integrate it commercially.

The project arrives as deepfake abuse continues to grow. A UK government investigation in 2023 identified 276,149 explicit deepfakes on one specialist site; 96% depicted women. AI “undressing” sites attracted 24 million visits in September 2023, prompting UK authorities to discuss stronger protections for children. In that context, RealOrRender’s combination of detection and visual evidence addresses a practical weakness of systems that provide a score without explaining it.