ChatGPT Images 2.0 Still Can’t Draw the Seven-legged Spider I Want

Whenever a new image generator comes out, I run the same test: “Please generate a spider silhouette missing its left front leg. Use an art deco style.”

And every time, the model fails.

Recently, Nano Banana/Gemini at least realized it failed, which feels like a huge step forward, but its spider still has 8 legs.

Chat GTP Images 2.0 has come the closest by far – the spider is missing a leg! It’s not missing the leg I asked for, but it’s actually missing a leg! You might not be shocked by this, but it’s because you haven’t spent nearly as much time trying to get various image models to pass this test (2025 testing).

Attempts to get it to remove the correct leg gave me a 5-legged spider with a tiny portion of the leg I wanted removed still floating in space. This is not the spider I wanted.

If I truly wanted to use an LLM image generator to create this spider, I’m sure it’d be possible. I’m not good at prompting to create images, and I don’t want to be. What I’m interested in with this test is the default experience when I ask the LLM to create an image I want.

Image generation with these tools reminds me a lot of both LLM-generated writing and code. It’s good at creating a plausible image, but creating exactly what I want is quite hard. A plausible-looking answer can be powerful in domains where validation is cheap so that the LLM can retry multiple times with feedback, and in domains where that sort of answer is good enough, but there are also plenty of domains where that sort of usage would be horrifying (which doesn’t seem to be stopping folks from trying).