Discussion: Why Model Artifacts Look Like MSE Blur, Not Natural Ugliness

kalomaze · x · 2026-08-21

Countering the defense that "some natural images are ugly" explains model flaws, the author argues:

Blur Pathology: Ugly samples generated by models typically exhibit MSE blur, not natural-looking ugliness.
Root Cause: This is a pathology of limited-rank function approximators that do not care about the limit case of score matching.
Conclusion: The issue cannot be attributed merely to the presence of ugly images in training data; it is a specific artifact of model architecture and training objectives.

More from Multimodal