Summary
- ChatGPT’s image-generation capabilities create unintentionally hilarious cartoons with mangled text.
- The new GPT-4o model is much better at generating images with clear text but isn’t used for scheduled tasks.
- Lack of updates to scheduled tasks means the dream of perfect cartoons is still on hold.
People are going nuts for ChatGPT’s new image-generation capabilities, creating everything from images of themselves in the style of Studio Ghibli to images of other people in the style of Studio Ghibli. Incredibly, ChatGPT can even make images in other styles, too.
Sarcasm aside, I was very excited when the new features were announced, as I thought it would mean an upgrade in quality for my favorite scheduled task. Sadly, I was wrong.
ChatGPT Makes a Cartoon Every Morning (and They’re Hilariously Bad)
When ChatGPT first announced it was adding a feature that allows you to create scheduled tasks, I immediately set about creating a few. Some were practical, such as a scheduled task that each morning sends me a list of long-term tasks I need to complete, which I can ask it to remove once I’ve completed them.
My favorite scheduled task was something sillier, however. I discovered that you could set up a scheduled task to generate an image using DALL-E based on your description. After a few tries, I managed to set up a task that sends me an original whimsical cartoon every morning.
These cartoons have been an endless source of hilarity, although most of it is unintentional. While there is the occasional mildly amusing idea, most of the laughs come from the fact that the ideas are usually just plain weird.
What makes them even funnier is the text.
The DALL-E image generation that ChatGPT was using to generate the images is fine at creating pictures but really struggles with text. There are almost always superfluous letters or mangled writing that make the cartoons even more unintentionally funny.
I Thought My Mangled Text Days Were Over
ChatGPT has now introduced a new image-generation model that replaces DALL-E, and it’s far superior. Not only can it generate impressive photorealistic images with excellent instruction adherence, but it is also able to reproduce your exact text (almost) all the time.
I was looking forward to seeing what my cartoons would look like with better imagery and text that you can read. However, when my first one came through, it was like all the previous versions with mangled text and run-of-the-mill image quality.
This was not what I was expecting. The images I was creating manually in ChatGPT were excellent, so why weren’t my cartoons coming out as well?
Scheduled Tasks Still Use DALL-E
I tried asking ChatGPT directly to generate a cartoon in the same style as my previous versions, and after a few tweaks to ensure it didn’t violate the content policies, I got a cartoon with better imagery and perfect text, without a mangled letter in sight. What was going on?
It turns out that, for some reason, ChatGPT’s scheduled tasks still rely on DALL-E to create images. This is despite the fact that ChatGPT Tasks use the GPT-4o model, and generating images in a standard GPT-4o chat will now always use the superior 4o Image Generation model.
The DALL-E-generated images even have text beneath them that reads “Made with the old version of image generation. New images coming soon.” The new images are already here, however, as long as you’re not creating a scheduled task.
I’m not sure why this is the case. If scheduled tasks use GPT-4o, and GPT-4o uses 4o Image Generation, then you would think that scheduled tasks would automatically use 4o Image Generation, too. Currently, however, this isn’t the case.
There’s Another Scheduled Task I Want to Create
The cartoons weren’t the only scheduled task for which I wanted to generate images. One of the first scheduled tasks I tried was to get ChatGPT to search for today’s weather forecast and then use that information to generate an image that encapsulated the day’s weather conditions. That way, I’d get a quick visual summary of the day’s weather every morning.
However, I found that the images almost always included text, despite my best efforts to stop it from happening. As you can probably guess, that text always ended up mangled, making the image mostly useless. I had to give up on the idea at the time, but I was hoping that I’d be able to make it work now that the new image generation is here. Unfortunately, my dream of beautiful weather forecast images is still on hold.
My Perfect Cartoons Will Have to Wait a Little Longer
OpenAI released support for scheduled tasks back in January, but at the time of writing the model is still labeled as “GPT-4o with scheduled tasks (beta)” in the app. Nothing obvious has changed since the feature was released, and it feels a little like it’s been forgotten about, with releases such as o3-mini, GPT-4.5, and 4o Image Generation all coming this year.
Hopefully, in the coming months, scheduled tasks will move out of beta and gain some more features. It would seem fairly trivial to update the image-generation model that’s used, so I’m keeping my fingers crossed that one day I’ll be able to generate my daily whimsical cartoon with legible text. Until then, I’ll enjoy my cursed cartoons and wait patiently for the day when I’ll finally be able to read the punchlines.
