OpenAI announced a long-awaited upgrade to ChatGPT's image generation capabilities, which haven't seen any improvements in over a year.
I'll be first to say that ChatGPT's previous image-generation capabilities were awful! I can' tell you how many times I almost cancelled my Plus subscription over frustrations with its image tools, which couldn't follow simple instructions, were incapable of including text in images (or excluding text if requested), and had no way of building upon previous designs.
Now ChatGPT users can leverage the company's GPT-4o model, which up until now, has only been able to generate and edit text. Altman said that GPT-4o image generation is live in both ChatGPT and Sora, its video-generation product, for Pro, Plus, and free plan users, with Enterprise and Edu access coming soon.
I was able to access it on my $20/month Plus plan, however, advanced image-generation requires a 7 minute cooldown period if you use it too much. I can relate…
GPT-4o with image output “thinks” longer than the image-generation of DALL-E 3, the model it effectively replaces, to create more accurate and detailed images. The output is worth the wait in my opinion.
Other improvements include:
- Accurate text rendering, precisely following prompts that leverage 4o's knowledge base.
- Ability to upload images to transform or use as visual inspiration.
- Character consistency, allowing users to create multiple images of the same character in different positions or scenes, such as when generating comics.
- The ability to upload and restyle images.
- The option to create images with transparent backgrounds, which is helpful when creating logos, badges, or elements for use in other image applications.
- Ability to handle up to 20 objects at a time per image-generation.
- The ability to make code-generated images, infographics, product instructions, and other visual guides that combine text and images.
Although the image-generation itself is significantly better than before, it's still like pulling teeth to get ChatGPT to produce images that don't “violate its content policy” — even though my requests never do.
For example, I uploaded a product photo from our website of a model wearing our leggings and sports bra and asked ChatGPT to create a lifestyle photo by placing the model on top of a mountain scene, close up on the model with the landscape slightly out of focus in the background.
First ChatGPT told me this was request was a policy violation and didn't make the image. Then I slightly adjusted the request, indicating that it was our company's product photo, and asked how I could change the prompt to move forward. It told me it couldn't and that I was in violation of its content policy.
Then I wrote a whole new version of the prompt, and ChatGPT told me I had to wait 3 more minutes because I hit a rate limit. I asked, “How did I hit a rate limit? You haven't generated a single image yet for me.” To which it replied, “Shut up, I make the rules” and posted this meme.
Finally I got ChatGPT to create the image, but instead of using the photo I uploaded, it recreated both the woman and our apparel (removing our name and logo from the clothing). Next I explicitly told it to use the exact image of the woman from my uploaded photo, and at most to adjust the lighting and levels to match the background it created — and it still screwed up. Several more attempts (and 45 minutes later), and ChatGPT told me to go find my own mountain image.
So it seems to be good at creating original image concepts from scratch, but not quite there yet when it comes to editing real product images.
It's only been released for a few days, but in what ways are you already using ChatGPT's new image-generation? Hit reply and let me know.