A new research paper published on arXiv explores a bias in text-to-image models where generated clocks frequently display 10:10, mirroring a common practice in advertising. The study tested various prompting methods on 52 models across Magnific and Higgsfield platforms to overcome this bias. Results indicate that while text prompts offer some improvement, adding a drawn reference dial significantly reduces the 10:10 bias and increases overall accuracy in generating correct clock times. AI
IMPACT Highlights a specific failure mode in text-to-image generation, suggesting prompt engineering and reference images as solutions for improved accuracy.
RANK_REASON Research paper detailing a specific bias in text-to-image models and methods to mitigate it. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →