Grok's New Image Model Inherits GPT's Noise Artifacts via Synthetic Data Inbreeding

mark_k · x · 2026-08-08

Users have noticed that Grok Image 2.0 is exhibiting the infamous 'noise artifact' issue previously seen in GPT-Image-2. The author points out that this isn't due to shared underlying technology, but rather because modern image models are increasingly trained on each other's synthetic outputs (an 'ouroboros' effect), leading to the cross-contamination and inheritance of specific flaws.

Original post →

More from Fun

Fun channel →