I find that AI image to prompt like gemma3 and ollama have trouble describing images when they're just a little different.