mortal pasture
Bug Report:
Steps to reproduce:
create an image of a woman walking on the beach, cinematic still, 1990 fashion style, film grain
create an image of a man meditating on a stone plinth overlooking Budapest, Dusk, film grain
Expected result:
Square outputs depicting the prompt
Actual result:
Widescreen outputs, despite not prompting for Aspect Ratio changes. Should default as Square.
Additional information
Browser: Safari 16.5.1
OS: Mac Ventura 13.4.1
additionally, images can come out rotated if you actually prompt for Aspect Ratio changes.
marsh trout
In the first example prompt it says "cinematic still". In real life, those are most likely in some sort of widescreen aspect ratio. (Let's ignore how anamorphic works for a moment.)
I do expect wide format when prompting for anything cinematic.
But I agree with you on wanting to have full control over the aspect ratio and not get sudden surprises.
marsh trout
additionally, images can come out rotated if you actually prompt for Aspect Rati...
my 9:16 images are most of the time rotated from the very start. 😦
mortal pasture
the content of your prompt should have no influence over the aspect ratio, I have this issue with various other prompts that do not contain "cinematic still" which was only in the very first prompt
maybe logically that could be conceived as a wide format term but it's still the case that no requests were made to influence AR which is contrary to how these tools work
chilly python
i can confirm that I am experiencing the same issue with most of my output
marsh trout
I never thought of this as an issue, really. ChatGPT writes its own prompts and adds things to vague ideas, so it's understandable that it chooses the aspect ratio on its own to fit the content if nothing is specified.
mortal pasture
I don't see how this could at all be an intended function.
It's not an issue that I clearly tell it 9:16 and get a rotated image? 😂
even prompts like this, where I specify vertical 9:16 rotate
just honestly makes it too inconvenient to use
dire sleet
To clarify, the tall images have an aspect ratio of 4:7 with dimensions of 1024 × 1792. DALLE-3 can be a bit odd sometimes with its generations but it seems as though they can be fixed! If this can be solved on the user's end, perhaps OpenAI will be able to resolve the issue in due time. The prompts for the erroneous and correct images are quite similar so I am quite perplexed as to what may caused this, nor what fixed it.
Erroneous (rotated landscape)
Tall photograph of an elderly Black business man in a sharp blue suit, balancing himself on a stone in a flowing river. The sun casts a warm glow on him as he blows a dandelion, its seeds floating gracefully in the breeze, surrounded by tall reeds on the riverbanks.
Correct (portrait)
Tall photograph portrait of an elderly Black business man in a sharp blue suit, standing in the middle of a flowing river. The sun casts a warm glow on him as he blows a dandelion, its seeds floating gracefully in the breeze, surrounded by tall reeds on the riverbanks.
ancient nymph
To clarify, the tall images have an aspect ratio of 4:7 with dimensions of 1024 ...
I tried "tall photograph portrait" like in your example and still got rotated images.
dire sleet
Like I said before, "DALLE-3 can be a bit odd sometimes with its generations"
My prompts do not guarantee efficacy, and they were only used to demonstrate
ancient nymph
Like I said before, "DALLE-3 can be a bit odd sometimes with its generations"
I'm starting to wonder if it's completely random - or somehow the AI decides some descriptive scenes are better horizontal vs vertical. I've written other simple prompts that just say poster sized and it makes them vertical without specifying pixels or what side should be longer than others.
dire sleet
If it's somewhere along the lines of "tall" then it will be fine
DALLE-3 has 3 pre-defined dimensions
square, tall, and wide
nothing else will be chose as of now
ancient nymph
square, tall, and wide
Just tried a brand new one asking for a tall poster and 3 out of the 4 are vertical and one is horizontal.
