I’ve recently created my own picture creation app powered by the DALLE-2 API. But to be honest the quality of the generated pictures are really not good.
So, how do i came to this statement.
Microsoft Bing Image Creator is also powered by DALLE-2. When i compare the results of an exact same prompt from my API DALLE-2 app to Bing Image Creator then i have to say there are worlds between the quality.
The Bing pictures are much more better than the API results.
So, the question is, am I doing anything wrong? Do i have to add something in the prompt? In API developer documentation i couldn’t find anything.
The prompt:
a dog running in a garden, black and white, high quality, high resolution, clear image, white background, ad ink outlines
With Bing Image Creator → exact the result i expected
With the API → the image is blurred and has no clear lines
Maybe someone got the same results and have an idea if there is something i can do about and how.
I think you would need to look into doing some model training on specific image sets, or possible use it with a GPT4 plugin.
I do not have much experience with images but seen some info in youtube videos.
Goodluck!
Hey Oliver, I had noticed the same things. Images seem to be far lower quality at the moment using the API than I have had previously. As you did, I have also done some tests against the Bing image creator - and, while I appreciate it might include some newer capabilities - the results are drastically different through all of my tests. It is rather strange. I feel better knowing someone else might be seeing the same thing so it’s not only in my head
I’m noticing a world of difference too, between API vs Bing and API vs Dall-E web app by Open AI. Both Bing and Dall-E web app are superior in quality. All same prompts at 1024x1024 on API:
A photorealistic full-body, front view, product shot of a backpack with teddy bear design, front pockets, brown , professional design, 4k, white background
Yes same here what we notice that the quality of the output in DALL-E Web UI is better than the API. It is possible that they use more light weight model on the API, but I would say the quality should not be impacted.
Do you know how we can escalate this question?
