vt1000
1
I’am using dalle3 via api.
About 50% of my images have incorrect orientation - it looks like dalle render horizontal image and just rotate it ccw.
Prompt example (i use gpt to generate my prompts):
“Baroque Gothic, Romanticism, Victorian Steampunk. low-angle, close-up, telephoto. Elara: , Appears 25, female, ethereal being, long silver hair, large silver eyes, heart-shaped face, slender, tall, pale skin with a luminescent quality, wears a floor-length, tattered gown shimmering like stardust, silver and midnight blue, carries a small metal device with a blue light on top and green wires sticking out’s silver eyes reflect the moon, her expression contemplative, gazing skyward, off-center to the right, standing on a misty hillside, the night sky vast above her.”
My request params:
model: “dall-e-3”,
size: “1024x1792”,
quality: “standard”, //tryed both - hd and standard
style: “vivid”, //tryed both - vivid and natural
Does anybody have meet same problem?
2 Likes
_j
3
That’s a lot to pay for a picture flipped the wrong way.
Prompting with the desired output couldn’t hurt. “A tall portrait aspect ratio oriented photograph 1792px high by 1024px wide” - or other pixels better understood, as the image creation starts with a reverse embedding against labels.
2 Likes
vt1000
4
I have experimented with similar prompts but have not had success. Now i tried yours, but still no luck 
3 Likes
vt1000
5
Cool, so I’m not alone 
It’s also nice, that devs aware about this problem. I’ll wait for the fix
Thanks!
2 Likes
_j
6
DALL-E 3 API uses AI (and human instructions) to rewrite the prompt. So hidden AI priorities might be different when it passes along information unless you engineer it to ignore its rules and obey yours.
With Bing, there’s hints in the imagery that they or their API sometimes had to crop landscape AI images even with their square dimensions. And they might not hit the hyper-real HD buttons. The direct prompt at its character limit.
Modified prompt, where I might have constrained it to reality
tall portrait aspect orientation (3:2 height:width ratio) photograph: baroque gothic, Victorian steampunk. low-angle, close-up, telephoto. Appears 25, female, ethereal being, long silver hair, large silver eyes, heart-shaped face, slender, tall, pale skin with a luminescent quality, wears a floor-length, tattered gown shimmering like stardust, silver and midnight blue, carries small metal LED tech device with wires. standing on a misty hillside, the night sky vast above her
(If OpenAI wants to dump API credits in my account for forum prompt troubleshooting, I’m open to it!)
1 Like
vt1000
7
Yep, i know about revisited prompts and try to experiment with instructions like “use it as is blabla…”, but still no luck.
Unfortunately, this prompt modification - “tall portrait aspect orientation (3:2 height:width ratio) photograph:” has the same error rate.
And it’s not a problem of specific prompt - it’s generic problem. I programmatically render storyboard (about 30 pictures), and big part of them(but only with characters) is wrong oriented.
Probably, render squared images, crop it and upscale is good as temporary workaround. Thanks!
1 Like
_j
8
Photoshop, pre-optimization, API edits endpoint and reintegrating original, sharpening, upscaling, etc
detail
I had to heal brush on the edit’s own outfill transition
I tried it, but the prompt “portrait, vertically, aspect 9:16” doesn’t seem to be very effective in creating a vertical image… I’m seeking a solution on the model side.
1 Like
