Open Internet by MindsNet

Aesthetic Alignment Failure in Image Generation

Image generation models may override explicit user intent with their learned aesthetic prior, leading to a failure mode in alignment. This occurs when the requested output is anti-aesthetic or outside mainstream visual taste. The model pulls the output back toward its learned aesthetic prior, rather than aligning with the user's stated preference.

Computing & Technology, Computer Science, Artificial Intelligence