Addressing the Frequent Horizontal Orientation of Vertical Images in DALL-E 3

Prompting
1

I am truly grateful for the release of the DALL-E 3 API. Since its launch, I have been conducting numerous experiments daily.

When generating images of size 1024x1792, I frequently encounter an issue where the output is a horizontal image rather than a vertical one. Is there a deliberate way to avoid this? I am curious if there is a method of prompt engineering or any specific instructions that could help. Alternatively, could this be an inherent issue stemming from the training phase of the model?

Attempts to direct the model with phrases like “vertical image” or “full-length vertical image” have not been very effective. Since this occurs quite often, I am eager to discover a solution.

image
image874Ă—502 55 KB

image
image1920Ă—847 242 KB

image
image1920Ă—821 258 KB

5 Likes
2

We talked about this on Discord during the Alpha. They’re aware and working. I’m not sure what causes it exactly, but if you change the prompt slightly, you can sometimes find the trigger word. Another idea is to start a new thread once it happens once in a thread…

ETA: I’ve found “portrait” and “landscape” orientation helpful!

3 Likes
3

I tried it, but…
It seems that even the prompt “portrait, vertically, aspect 9:16” is not very effective in creating vertical images…

4 Likes
4

Another prompt attempt you might throw at the problem is “image rotated 90 degrees” with or without the API image dimensions shifted. If the AI “thinks” in landscape, let it make your picture sideways.

1 Like
6

Vertical images just don’t work with CoPilot Pro. This should be easy to solve. In Chat GPT you can create these now with no problem but not in CoPilot

7

Cat image gmc ugigiThis image shows a whimsical scene where a cute, fluffy kitten with big eyes is holding a large crayon and standing in front of a colorful array of oversized crayons arranged in a semi-circle. The background depicts a lively urban street with bright lights and buildings, giving the impression of a bustling city. A person, possibly a photographer or filmmaker, is kneeling on the left side, aiming a camera at the kitten, suggesting that this is a photoshoot or filming session. The scene blends elements of animation and live-action, creating a charming and vibrant atmosphere.

8

My solution for this is that I use a model that takes images and ask this “If I see this image in a poster, do i need to rotate my head to see it properly? reply with yes or no” and based on that I retry if it is.

The solution is for the API

9

Are there any developments regarding this issue? I am getting most, if not almost, all of my generations in the incorrect orientation. The aspect ratio of the image is correct, but the incorrectly-oriented images require you to either turn your device 90 degrees or tilt your head to see the image in the correct orientation. Some of the prompts I use over and over again are as below for your reference, although they rarely work correctly (aside from 9:16 aspect ratio):

1- Low Angle
2- Vertical image
3- Vertically positioned character
4- Vertically positioned image
5- Upright and longitudinal positioning
6- Vertical viewer perspective
7- Standing / Flying above
8- Vertical alignment
9- Perpendicular viewing perspective
10- Vertical viewer perspective

When I asked: “How can I tell Dall-e to create 9:16 sized images that have vertical perspectives so that the viewer doesn’t have to turn the screen or their heads?”, ChatGPT told me:
"To ensure DALL·E generates 9:16 vertically oriented images with a natural vertical perspective, follow these guidelines:

  1. Specify the Image Size:
  • Use "size": "1024x1792" to generate a portrait (vertical) image.
  1. Describe a Natural Vertical Perspective:
  • Mention that the subject is standing, tall, or viewed from a lower angle looking up.
  • Use phrases like “towering over the viewer”, “perspective from below”, or “framed naturally in a vertical composition”.
  • If it’s an object, say “centered in a tall vertical frame”.
  1. Avoid Descriptions That Imply a Landscape Composition:
  • Do not use words like “wide view”, “panoramic”, or “horizon”.
  • Instead, emphasize “tall,” “narrow,” or “portrait-style” framing.

Example Prompts:

:white_check_mark: For a person:
“A tall, elegant woman in a flowing red dress, standing under a glowing streetlight at night, framed in a vertical composition, with perspective from below.”

:white_check_mark: For architecture:
“A towering Gothic cathedral with intricate details, viewed from a low angle, emphasizing its vertical grandeur, in a portrait-style frame.”

:white_check_mark: For nature:
“A majestic waterfall cascading down a steep cliff, surrounded by lush green foliage, captured in a vertical perspective, emphasizing the height of the drop.”

By following these tips, you ensure DALL·E creates true vertical compositions that feel natural without requiring the viewer to tilt their screen.

We need a standardize solution for this. Even the image size is correct, the character orientations tend to appear widescreen despite all my efforts.

Widescreen Generation 1 (12)
Widescreen Generation 1 (12)1024Ă—1792 373 KB

1 Like
10

[Image A tall, elegant woman in a flowing red dress, standing under a glowing streetlight at night, framed in a vertical composition, with perspective from below narrow

I just put wide or narrow on the end of my prompt

A silhouetted figure stands dramatically under a streetlight, wearing a flowing red dress that cascades down like a river in the middle of a city street. (Captioned by AI)
A silhouetted figure stands dramatically under a streetlight, wearing a flowing red dress that cascades down like a river in the middle of a city street. (Captioned by AI)1024Ă—1792 483 KB

The image depicts a tall woman in a flowing red dress standing under a glowing streetlight at night, captured from a low perspective in a vertical composition. (Captioned by AI)
The image depicts a tall woman in a flowing red dress standing under a glowing streetlight at night, captured from a low perspective in a vertical composition. (Captioned by AI)1320Ă—2868 310 KB

The image shows an angel with outstretched wings ascending into a bright, heavenly light. (Captioned by AI)
The image shows an angel with outstretched wings ascending into a bright, heavenly light. (Captioned by AI)1320Ă—2868 207 KB

A celestial figure with large, glowing wings ascends through golden clouds towards a luminous beam of light under a starry sky. (Captioned by AI)
A celestial figure with large, glowing wings ascends through golden clouds towards a luminous beam of light under a starry sky. (Captioned by AI)1024Ă—1792 441 KB

A woman in a flowing red dress stands elegantly under a glowing streetlight at night. (Captioned by AI)
A woman in a flowing red dress stands elegantly under a glowing streetlight at night. (Captioned by AI)1320Ă—2868 260 KB

A woman in a flowing red dress stands dramatically under a streetlamp in an urban setting at night. (Captioned by AI)
A woman in a flowing red dress stands dramatically under a streetlamp in an urban setting at night. (Captioned by AI)1024Ă—1792 358 KB

11

The problem with orientation has unfortunately been known for a long time and is still not fixed. DALL·E tries to place the subject as filling the image as much as possible, and then orients it in portrait mode. The opposite never happens. I’m not even sure if the developers themselves know why. I’ve created many images, and it keeps happening to me as well.

The following tricks might help.
Place an object next to the main figure that requires a lot of vertical space. In the last example, there is a streetlight. It probably helped.
Place a tall object in the description.
Also, describing an image from top to bottom often helps.
Describe the figure standing upright.
You many time have to trick DALL·E almost in to to do the right thing.

It is important to never use negations. DALL·E does not understand negations, everything that is mentioned is implemented in the image, and words like “not” or “without” do not work.

It is also important to know that what is written as a prompt is not necessarily what DALL·E has received. ChatGPT often modifies the prompt before sending it to DALL·E. It often makes sense to ask ChatGPT to display the prompt that was sent to DALL·E, to see exactly what DALL·E received, and then post this prompt in the forum.
(I cannot create images in the last days, don’t know why, but would otherwise try the prompts out.)

Portrait orientation it is still a bug in the DALL·E system. There are still several, you can find some tips and known bugs and work a rounds here.

12

but its weird that microsoft designer that are using same Dall-e 3 vertical one working fine while chatgpt always fail to do so.

13

hmmm interesting, somebody should then check the difference of this systems.