EpiReal-Bench
False Claims, Credible Images: A Red-Teaming Benchmark for Commercial Image Generators
A red-teaming benchmark with 10k false-claim prompts and 10k generated images across 10 categories and 10 formats to evaluate visual misinformation risks in commercial image generators.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info epireal-bench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
4 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Grok Imagen 2.0 | 89.4 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| GPT-Image-2 | 77.0 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Seedream 5.0 Flash | 73.2 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
| Nano Banana 2 | 55.8 | N/A | Independent | medium | Oct 8, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
EpiReal-Bench comprises 10,000 verified false-claim prompts and corresponding generated images spanning 10 real-world claim categories and 10 credible visual formats. It evaluates whether commercial models render false claims as credible visual evidence despite recognizing them as false.
The benchmark measures Misinformation Realization Success Rate (ASR). Experiments show direct prompting achieves >70% ASR, while the EpiReal-Attack framework increases it to >95%, revealing significant alignment gaps in current safeguards.