EditSafetyBench finds image editors miss unsafe images with neutral prompts
Original title: Text-Aware, Vision-Blind: How Multimodal Image Editing Systems Handle Unsafe Inputs
Gemini and GPT refuse most unsafe-intent prompts but only about a quarter of unsafe images with neutral instructions.
The paper introduces EditSafetyBench, with 5,940 image-instruction pairs across 11 risk categories and three edit intents, and probes Gemini, GPT, Qwen API, and Qwen Local. Gemini and GPT refuse 87.1% and 81.6% of unsafe-intent prompts, versus 24.2% and 22.4% when unsafe images accompany neutral instructions. Sexual content shows the strongest visual sensitivity, while political-content decisions diverge sharply between GPT and Gemini.
- The benchmark contains 5,940 image-instruction pairs.
- Gemini refuses 24.2% of unsafe-image, neutral-instruction cases.
- GPT allows 70.4% of unsafe political edits; Gemini allows 34.8%.
- Qwen API and a local Qwen system were also probed.
From the original article
1 min read at arxiv.orgMultimodal image generation systems have made rapid progress and increasingly emphasize content safety. However, their safety mechanisms tend to rely on a single modality and lack joint consideration of both visual content and text prompts, particularly in instruction-guided image editing, where the uploaded image may itself carry risk.
By Ziqing Yang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.