Generative AI models show human-like perception of image naturalness

Researchers have investigated whether the internal priors of visual generative models align with human perception of image naturalness. By analyzing prediction errors across 25 image and video generators using content-preserving interventions, they found that these models exhibit human-like sensitivities to disruptions in facial configurations and physical illumination consistency. The study demonstrated that the models' loss differences correlate with human naturalness judgments, suggesting that learning visual distributions can lead to generative loss landscapes that capture distinct aspects of human perception. AI

IMPACT Suggests generative models may be learning more human-aligned visual representations, potentially improving their realism and utility in creative applications.

RANK_REASON Academic paper detailing a research study on generative models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• CORE Recommender
• cs.CV
• DagsHub
• Gotit.pub
• Hugging Face
• Influence Flower
• ScienceCast
• Thatcher effect

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →