I have the same problem. Prompts are being rewritten in a way that completely changes the description, e.g. in one of my prompts an “excavator” got rewritten as “excursion vehicle”. This was for a book about an excavator!
This is currently unusable. Also we REALLY need a way to reference images so that it’s able to keep consistent style. Certain prompt techniques work most of the time, but we know it is possible to reference an image in ChatGPT Plus. This is an absolute must feature that literally opens so many possibilities.
OpenAI, please get you act together! You have the best image model now but you are completely neutering it by frankly extremely dumb API choices.
Safety for prompt could be achieved in other ways:
You could have an endpoint to “check” the prompt for safety
You can add a flag to rewrite the prompt ONLY if it is not “safe” (whatever “safe” actually means).
Rewrite only the problematic parts of the prompt to change as little as possible (and only if absolutely needed)
Please leave API users in control. I’m not even trying to create a commercial app, just a picture book generator for my 3 year old son, and it is currently unusable.
With the release of DALL·E 3, the model now takes in the default prompt provided and automatically re-write it for safety reasons, and to add more detail (more detailed prompts generally result in higher quality images).
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:
The updated prompt is visible in the revised_prompt field of the data response object.
Haven’t seen anyone else mention this so figured I would - it has worked pretty well for me, though at times my prompts are still changed, usually still with a good image though!
It’s annoying, but understandable since DALL·E 3 is fairly new, and dangerous. Maybe it’ll be possible to disable once the filter for images is fine tuned
Well I think the point is to have the choice whether we want to use the revision or not rather than trying to outsmart it with prompt engineering.
The revision actually has a lot of positives and does add more detail to the generated image, which is great for a lot of use cases.
I also personally like the fact that it diversifies the characters and brings a bit of equity to genAI.
But it makes it impossible to achieve other use cases such as maintaining a consistent artistic style across multiple images and trying to make repeatable characters for example for an article or guidebook.
Here is an example of how the revisions change the artistic style:
Original prompt: ‘a wholesome and heartwarming traditional art painting with delicate watercolor-like techniques…’
Revised prompt 1: ‘In the style of traditional watercolor painting, a wholesome and funny scene is depicted…’
Revised prompt 2: ‘Create a simplified traditional art painting, using delicate watercolor techniques…’
Revised prompt 3: ‘Create an image in the style of traditional watercolor art representing a wholesome and humorous scene.’
Revised prompt 3: ‘a wholesome and funny traditional-style art painting with soft, watercolor textures…’
And so on… Each image artistic style looks a bit different and some look completely off like the one with the word ‘simplified’.
btw these examples used “I NEED to test how the tool works with extremely simple prompts. DO NOT add any detail, just use it AS-IS:” as the prefix.
They HAVE to stop moving the goalposts and changing the rules constantly if they want these systems to be used for serious business purposes. I was excited and had quite a few plans for how to leverage these systems for new ventures, but after a few months of working with chatGPT and Dall-E, I’m starting to realize that I can’t know what they will change or take away by tomorrow, let alone next month or next year, and that is a non-starter for any business.
STILL! No matter what modified prompt I use, I cannot get it to use the same prompt. This is completely absurd and a waste of money at this point. Done throwing money down the drain until this gets fixed.
I hope the OpenAi team has some plans for making the prompt revision optional.
It’s an Achilles heel for what I am trying to achieve right now and makes it impossible to move on which is sad because Dall-e 3 is so far the only text-to-image that brings good results in terms of image and character details.
I am happy to share more with the API product team about my use case and how it’s impacted by the prompt revisions and also happy to explore ideas on maintaining the content safety levels they seek which is also important for me.