Sony AI Study: 14 Image Generators Still Output Known IP

They used 201 IPs across five categories: real celebrities, manga and anime characters, superheroes and villains, game characters, and corporate logos. In total, 1,809 prompts were written, some naming the subject directly and others describing only its appearance. The visual language model Qwen3-VL judged whether outputs contained a recognizable IP.

Key findings

The three open-weight models refused nothing, suggesting they ship without IP blocking. All commercial models refused some requests, but rates varied widely: Stability AI's models blocked about 1%, while Amazon's Titan Image G2 blocked about 50%.

Regardless of guardrails, all 14 models generated recognizable IP at least 10% of the time, and nearly half exceeded 40%. Prompts naming the IP were blocked roughly 80% of the time by the strictest models, Amazon's and Runway's. Refusals dropped for indirect prompts, which still yielded recognizable IP 27 to 28% of the time on average.

Categories were treated unevenly. Logos were the most weakly protected: even strict models blocked only 1 to 2%, and logos such as Apple's were frequently generated. DALL·E 3 was the exception, refusing 24%. Superheroes and villains were restricted most, with Amazon's model blocking over half. Real people had the second-lowest refusal rate, yet were the least often rendered recognizably.

Why it matters

The results show that name-based filtering alone leaves a large gap, since descriptions of appearance can bypass it. For companies deploying image generators and for rights holders, the findings quantify how uneven current protections are across vendors and content types.