Beyond “Is It AI?”
An image can look real and still get the facts wrong. EviScene investigates what it depicts—and checks it against evidence.
Agentic Image Factuality Assessment · 2026
A finished museum, hosting an opening ceremony.
The museum was still under construction at assessment time.
Start with an imageNo user-provided textual claim required.
Investigate the factsConnect visible details to external evidence.
Explain the verdictA compact 9B policy, trained to investigate.
THE QUESTION
A genuine photograph can be misleading. A generated image can depict a factually consistent scene. Origin and factuality are different questions.
Identify the people, objects, events, text, and context that can be checked.
Find sources and inspect whether their details actually match the depicted situation.
Supported Refuted
With an explanation tied to the evidence.
THE EVISCENE AGENT
A multimodal ReAct policy chooses what to inspect, where to search, and when the evidence is sufficient to conclude.
Read the scene and its details. Surface specific, checkable cues rather than relying on visual plausibility.
Search the web and retrieve related images to identify relevant events, entities, and sources.
Visit source pages and compare reference images. Test dates, identities, and details against the input.
Resolve open questions, then produce a verdict supported by the collected evidence.
An investigation loop, not a fixed checklist. The policy can revisit observations as new evidence arrives.
THE BENCHMARK
A benchmark designed to test factuality across both web-sourced and controlled images—not just whether an image was generated.
What happened, and between whom?
Do the visible details hold up?
Is this the right time, place, and source?
8,647 training images · 1,527 test images, with paired metadata.
TRAINING A COMPACT INVESTIGATOR
Supervised fine-tuning teaches a Qwen3.5-9B policy to carry out evidence-seeking, tool-using investigations.
Privileged on-policy self-distillation targets weak decisions with training-time hints, while preserving successful behavior.
The student acts from the original, unhinted context.
MAIN RESULTS
EviScene combines classification quality with evidence-grounded success. All values below follow the current manuscript snapshot.
| Policy | BAcc ↑ | Macro F1 ↑ | SESR ↑ |
|---|---|---|---|
| GPT-5.5 | 78.13 | 72.06 | 50.49 |
| Gemini 3.7 Flash | 81.57 | 76.70 | 42.89 |
| Gemini 3.1 Pro | 62.42 | 64.17 | 32.68 |
| Qwen3.5-397B | 71.72 | 73.76 | 43.29 |
| Qwen3.5-9B · Base | 72.34 | 74.57 | 42.89 |
| Qwen3.5-9B · SFT | 78.81 | 75.69 | 50.49 |
| EviScene · SFT + PSD 9B | 79.05 | 74.67 | 52.26 |
All rows use the agent setting. EviScene has the highest SESR among these policies.
| Qwen3.5-9B setting | BAcc ↑ | Macro F1 ↑ | SESR ↑ |
|---|---|---|---|
| Direct prediction | 65.70 | 57.26 | 1.70 |
| Base agent | 72.34 | 74.57 | 42.89 |
| After SFT | 78.81 | 75.69 | 50.49 |
| After SFT + PSD | 79.05 | 74.67 | 52.26 |
From the base agent to SFT + PSD: +6.71 points in BAcc and +9.37 points in SESR.
| Tool configuration | BAcc ↑ | SESR ↑ | Δ SESR |
|---|---|---|---|
| All tools · EviScene | 79.05 | 52.26 | — |
| Without web search | 74.37 | 35.89 | −16.37 |
| Without image retrieval | 74.52 | 40.01 | −12.25 |
| Without evidence inspection | 76.55 | 46.04 | −6.22 |
Evidence inspection comprises source-page visits and reference-image comparison. Web search has the largest SESR drop when removed.
BAcc Balanced accuracy. SESR Strict evidence-supported success rate: a correct verdict must also meet the evidence-quality criteria. Scores are percentages; differences are percentage points.
SEE THE INVESTIGATION
Four qualitative cases from the paper. Read the visible cue, the retrieved evidence, and the reason for the final assessment.
Supported
The “DisplayHDR True Black 1400” label looks questionable without context. The investigation finds VESA documentation introducing that certification tier in July 2026.
Refuted
The poster shows 14 October 2025–1 February 2026. The retrieved schedule instead gives 14 October 2026–1 February 2027.
Refuted
The image depicts an opening ceremony at a completed museum. Project records show berlin modern was still under construction at the paper’s assessment time.
Supported
The investigation identifies the venue and event. France’s Ministry of Culture confirms the launch at the Moulin Rouge on 21 January 2025.
EXPLORE THE PROJECT
Start with the paper, explore the implementation, and find the checkpoints and EviLens in one Hugging Face project.
Method, benchmark, experiments, and additional analyses. The paper and original figures are in English.
Read PDFAgent runtime, tool integrations, SFT and PSD training, and evaluation.
GitHub repositorySFT and original PSD checkpoints, EviLens images and metadata, code, and the paper. See the repository for release availability.
Hugging Face projectREFERENCE
@misc{wang2026evisceneisaiagentic,
title={EviScene: Beyond "Is It AI?" for Agentic Image Factuality Assessment},
author={ZiAn Wang and Ruiqi Liu and Zhengbo Zhang and Yi Han and
Jialiang Shen and Ruoxin Chen and Manni Cui and Ximo Zhu and
Jin Jiang and Mingzhe Liu and Linhao Wang and Wenhao Wang and Shu Wu},
year={2026},
eprint={2610.14609},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2610.14609},
}