I had similar bad results with gpt 4o/5 when they got these image generation capabilities.
Don’t know what’s the reason: my bad prompting or these models being tuned to work with photos/illustrations only