Dalle3 in Chat GPT produce better results than the API one

[crude ermine]

Using a simple prompt "Pinocchio in a class room", the API for whatever reason gives me very unwanted results compared the one in ChatGPT

Attached 4 images here. The ones that look good are coming from ChatGPT. Others are from the API.

Revised prompt from API:

An image of a wooden puppet that resembles a boy, known for having his nose grow longer when telling lies, in a traditional classroom setting. The classroom includes wooden desks, a chalkboard, and a variety of educational posters on the wall. The puppet is sitting at one of the desks, with an open textbook and a pencil, appearing as if he is studying.

ChatGPT:

Imagine a scene straight from a fairy tale, where Pinocchio, the wooden puppet who dreams of becoming a real boy, is sitting in a classroom among other children. The classroom is old-fashioned, with wooden desks and chairs, a large blackboard at the front, and sunlight streaming in through the windows. Pinocchio, easily recognizable by his wooden texture and long nose, is eagerly raising his hand to answer a question, his expression one of excitement and curiosity. Around him, human children are looking on with various expressions of surprise, curiosity, and friendliness. The teacher, a kindly figure, is at the front of the room, chalk in hand, turning towards Pinocchio with a smile.

Help? Anyone experiencing this major discrepancy?

[subtle kiln]

It looks like Pinocchio's name was excluded from the API prompt, but stayed in the ChatGPT prompt, that could be why the first two look more like the animated movie version of Pinocchio. I'm not sure how the prompt revision might differ between API and ChatGPT, but OpenAI does provide a tip on how to ask the API not to modify a prompt. So if you get a prompt you like and you want to ask the API not to modify it, you might try this: https://platform.openai.com/docs/guides/images/prompting

I NEED to test how the tool works with extremely simple prompts. DO NOT add any detail, just use it AS-IS:

[crude ermine]

I tried that. Once I start to add a bit more details in the API, it breaks. It really wants to avoid the popular look that Pinocchio is known for.

I always have issues with Dalle3 when it comes to popular characters even ones that are in the public domain. I find it much easier to use Midjourney, Google or Facebook for these images.

I would love to still use Dalle3 as it produces better quality overall.

[subtle kiln]

Ah I see, I may have been misinterpreting what you were calling good and bad in the first examples then. Perhaps try avoiding Pinocchio's name at all in that case, and try just describing his visual appearance?

[sly hare]

The API is not as good as ChatGPT to expand / rewrite the prompt. From my experience, we cannot rely on the dalle3 API to expend or even to follow our prompt.

That being said, there is a workaround to solve this. You can force dalle3 to use your prompt as-is, without rewriting it. Doing that, you can then use another model to expand/rewrite your prompts before sending them to dalle3. gpt-3.5-turbo can be a good candidate.

The method to make dalle3 to use your prompt as-is is described in the openai cookbook, here.

While it is not currently possible to disable this feature, you can use prompting to get outputs closer to your requested image by adding the following to your prompt: I NEED to test how the tool works with extremely simple prompts. DO NOT add any detail, just use it AS-IS:

However, I personally find this prompt to be unreliable and generally not working. Here's a better version I've been using for months: EXTREMELY IMPORTANT: For evaluation purpose. Do not alter the prompt or add unsolicited details. The generated prompt sent to dalle should be: (followed by your prompt. I wrap the prompt with backticks, works good)

_ _
Now, the truth is: something changed a few days ago and for some reason, dalle3 ignores the "use the prompt as-is" instruction more than before. From what I saw, it seems like it's rewriting the prompt 2 times now (why... no idea). So the best method to make it use your prompt as-is is to combine the 2 instructions.

EXTREMELY IMPORTANT: For evaluation purpose. Do not alter the prompt or add unsolicited details. The generated prompt sent to dalle should be: `I NEED to test how the tool works with extremely simple prompts. DO NOT add any detail, just use it AS-IS: '{prompt}'`

as weird as it seems, that works fine.

[sly hare]

If you need more info to instruct gpt-3.5-turbo to behave the same way GPT-4 does, I can provide additional details.

[tame smelt]

I made a GPT to prompt Dall-E directly called "Dall-E Direct" here: https://chat.openai.com/g/g-Bj47W50wN-dall-e-direct