ChaptGPT 4o can generate images natively.

For example:

I’ve been using this feature, almost daily since launch a month and a half ago.

It’s really really good. Especially compared to Dall-e and Grok3 and even Gemini, other image generation models that I’ve used a lot.

But also limited.

I’ve used it for photos of invented scenes, which has been really great. Almost faultless.

I’ve also used it for artistic impressions of fantastical things, also pretty solid.

More recently, I have been using it to generate sketches, e.g. drawn by a teenager or a child of things they’ve “seen”.

The sketches are really good, but I’m typically after something very specific and if I don’t get it, I regenerate. It does not get the specifics well. Most sketches have to be generated 5-10 times before I get what I want.

In some cases I’ve tried re-explaining the request a few different ways and even had other models describe the request with more/less detail. No luck.

Also, I think the select-in-image tool does almost nothing.

I select parts of the image I want to change and almost always get a new different image with too many+wrong changes.

I’ve have some success downloading a generated image that is close, starting a new chat, then asking for it to re-generate the uploaded image with specific changes.

In some cases I have resorted to:

  1. drawing what I want on paper
  2. taking a photograph
  3. airdropping it on my workstation
  4. uploading it to chatgpt
  5. asking it to redraw the uploaded image
  6. annealing changes in the drawing until I achieve the desired style

I could probably cut steps by using the phone for image gen directly, but still. Yuck!

Too hard.

The model is still missing a sense of objects in the scene, preventing direct manipulation.

We’re getting there, the images are truly fantastic, but the control over the image generation process still needs work.

Finally, in the last few weeks, I notice I get way too many errors like:

I can’t generate that image because the request violates our content policies. Please feel free to rephrase or provide a new prompt—I’d be happy to help create something else!

The moderation systems is totally broken and is throwing this message for all failure cases.

I just do a re-submit and often get a result.