Hey everyone. I’m playing around with generating images using DALL-E 3 via the API, and I’ve noticed that the image quality for the same prompts is much worse than ChatGPT. For example, see the images below:
Prompt:
A vector illustration of an individual holding a car key in one hand and a pile of dollar bills in the other hand, against a backdrop of a sleek car and document featuring loan terms, all predominantly in a light blue color scheme on a white background… The dimensions need to be 1792x1024
What is going wrong here? Anyone else having the same issue?
UPDATE:
It looks the API I was using (via Pipedream) is using DALL-E 2 instead of DALL-E 3 (despite calling it DALL-E 3). I used the same process on Zapier and here are the results:
Depends. In your description, there are many randoms which may cause different output. E.g. “individual” may be anything, same as the car and so on.
When you are downloading an image from ChatGPT, the file name is what I assume is an actual prompt. Try to use the same prompt in DALL-E 3 API and see if it’s much worse.
In my experience, I see that it’s very different, and ChatGPT images are generally better, no matter what I do. And in terms of a prompt I do a lot. I am quite convinced, that they have an additional know-how middle layer in ChatGPT, which either modifies the prompt or tweaks the engine (e.g. different number of steps and so on).
Hey everyone, thanks for your input! Even though there are minor differences between ChatGPT and the DALL-E 3 API, the actual reason for the huge difference I reported above was that Pipedream is NOT using DALL-E 3, but probably DALL-E 2 instead (despite saying it’s DALL-E 3).
I have used the very same prompt via Zapier, and the results look normal, similar to ChatGPT, as you can see below:
