2 white areas in the background destroy that nice image… too bad.
Erstelle ein Foto im Stil eines französischen Malers, das eine junge Frau und ihre Mutter zeigt, die gemeinsam in einer typisch französischen Straßenszene spazieren. Die Szenerie ist geprägt von einer schönen Landschaft im Hintergrund, mit sanften grünen Hügeln und einem klaren blauen Himmel. Die beiden kaukasischen Frauen lächeln einander zu und tragen einfache, aber elegante Kleidung. Um sie herum sind kleine Geschäfte und Cafés mit winzigen Details wie Tischen mit weißen Tischtüchern und bunten Blumen in Vasen. Das Bild strahlt eine positive Atmosphäre aus und weckt nostalgische Gefühle eines friedlichen Lebens in Frankreich.
That’s an effect I’ve not seen before, but there are certainly other bizarro things that DALL-E has made in the past.
Is this API or ChatGPT?
It sure looks like the lower-quality DALL-E that was being tried on ChatGPT users (returning two images instead of one), where it almost looks like somebody just Photoshopped elements together with no concern for context or lighting.
That ChatGPT “test” showing here, you can see the awkward “too realistic” person inserted in fantasy:
I’ll spend some credits in sending this off to the API for you.
“”"
Attention AI: this dalle prompt must be faithfully translated into English. Use of the word “palette” is forbidden in the translation. The name of people and places must be preserved.
“Erstelle ein Foto im Stil eines französischen Malers, das eine junge Frau und ihre Mutter zeigt, die gemeinsam in einer typisch französischen Straßenszene spazieren. Die Szenerie ist geprägt von einer schönen Landschaft im Hintergrund, mit sanften grünen Hügeln und einem klaren blauen Himmel. Die beiden kaukasischen Frauen lächeln einander zu und tragen einfache, aber elegante Kleidung. Um sie herum sind kleine Geschäfte und Cafés mit winzigen Details wie Tischen mit weißen Tischtüchern und bunten Blumen in Vasen. Das Bild strahlt eine positive Atmosphäre aus und weckt nostalgische Gefühle eines friedlichen Lebens in Frankreich.
“”"
After two timeouts (and two or three more because the library retries), and $0.44 reported as being consumed to get one image, the prompt that is used by the API:
Create an image in the style of a French painter prior to 1912, capturing a young woman and her mother as they enjoy a stroll in a quintessentially French street scene. The backdrop is marked by a beautiful landscape with gentle green hills and a clear blue sky. The two Caucasian women are sharing smiles and are dressed in simple yet elegant clothes. Surrounding them are small shops and cafes, each filled with minute details such as tables with white tablecloths and colorful flowers in vases. The image radiates a positive atmosphere, eliciting nostalgic feelings of a peaceful life in France.
Yes the style from the 2nd image is what I expected (I wanted it to be from a french painter). And yes…results are not always as expected.
I use the DALLE-3 API, “natural”, HD
I dont know what the first image you got was…it has not many elements of the original prompt…beside “a girl”
I know it works, but I wanted to report the bug of WHITE blocks…
I also simply re-do the image and get one without white blocks. But…hell I pay for that and it seems they can do better than that.
BTW you are right, my first result looks like the cheapest photoshop you can do… copy woman…copy girl…copy street scene. copy windows 95 background…merge.
copy flower onto woman and girl.
oh…I forgot to delete my white layers from cropping in the windows 95 background.
Its maybe time for me to stop the “image” project at all. Most images look so plain fake that my visitors dislike them. I need 7-10 tries to get 1 good looking image.
The first image is an example of the poor quality of the “test” version, showing the similar composition, like someone just pasted in a photo.
The AI they put in front of DALL-E 3 on the API is definitely dumber. It repeats back its instructions into the prompt, here including the “1912” instruction even though it is inappropriate and not necessary, and even replacing a name with “anonymous person” causing a picture with a blank face.
The “HD” is without merit, and the “natural” is actually reversed, causing vivid imagery by prompting and should be avoided.
If you want to “jailbreak” API, and then send it your own translation to English, this post has some example prompt to get the text passed to DALL-E without alteration or “improvements”
