I have encountered issues when using the GPT-4-O model in ChatGPT and the DALL· E GPTs for image generation, specifically with the 9:16 vertical format. Despite making various adjustments and modifications to the prompts, the generated images rarely appear in the correct 9:16 vertical ratio. This issue occurs both in the image generation feature of ChatGPT and on the DALL· E platform.
This problem has impacted my experience, so I would like to know if this is a known issue or if there are any recommended solutions that could help me generate images that accurately fit the 9:16 vertical ratio.
For Example:
Beautiful woman,Japanese actress,off shoulder,city background,
looking us and smiling,upper half of the body,incredibly detailed
face,incredibly detailed beautiful eye,focus on face,masterpiece,
high quality,best quality,highly detalled,insanely detailed,4K,9:16
output:
The issue was fixed when I tried generating again today; all my images were generated at 9:16. I used to experience this issue too months back. Better still, use Sora; the aspect ratio is there.
You’re right! The issue with DALL-E 3 producing rotated images when requesting portrait/vertical images is not an issue any more with the replacement image generator using gpt-4o technology. Which now uses a different size, now 1024x1536 instead of 1024x1792.
Therefore, to be more precise, images were not generated at 9:16, as one can see from the choices in Sora.
Have you found solutions to this aspect ratio problem? I have not been able to generate 16:9 or 9:16 for quite a while now and it has virtually ruined ChatGPT image generation for me because I rely on these aspect ratios for my YouTube channel.
Hey there! Great question. To get that perfect 16:9 frame for HDTV, you’ll want to crop 160 pixels off the total height of your image (so from 1024 px down to 864 px), usually by taking 80 pixels off the top and 80 off the bottom. That’ll leave you with an image that’s 1536 × 864 pixels—which is a spot-on 16:9 ratio that fills an HDTV screen nicely without any distortion.
If you want to match a standard HDTV format like 1920x1080 (1080p), you can resize your image or use another provider’s AI upscaler.
And here’s a handy tip: to make sure no one’s head gets chopped off during cropping, try requesting things like “zoom out,” “seen from a distance,” or “pad with extra background content at the top and bottom of the image.” This helps keep the important stuff centered vertically, so your main subjects stay safely in the frame!
Let me know if that helps, as the new image model only has the single wide output size.
Thanks a lot for taking the time to reply with these tips. Actually, I have tried them extensively with varying results. It’s really hit and miss. Mostly miss as the generator often places elements within the crop zone, among other errors.
The new image model’s inability to accommodate 16:9 and 9:16 has been a hammer blow to my YouTube video creation. It doesn’t appear that Open AI plans to re-enable these aspect ratios.
I haven’t tried this one [“pad with extra background content at the top and bottom of the image.” ] so I will give it a go.
If you are willing to pay another $0.04 for dubious quality: You can use DALL-E 2 on the API edits endpoint to out-fill beyond the edge of a 1536x1024 image, 1820 px using its 1024x1024 input as a “window” of the area where you need more image created, by making it transparent.
Needed is a minimal code function as if for a library that:
takes in a base64-encoded png, jpg, or webp file, your return,
verifies it is in the 1536x1024 aspect and size required,
converts it to 32-bit image (thinking of png as a destination),
adds the needed up-size to the destination width with transparency mask that is full transparent where width is added,
uses a function that slices out the left or the right 1024x1024 with its transparency and gets an outfill:
subfunctions: uses the dall-e-2 AI model on the OpenAI edits endpoint with the OpenAI Python SDK (openai library) to receive back it’s non-transparent image with new 1024x1024 generation.
subfunction: overlays that result over the RGB with full opacity being used in A, blending the internal edge by a width parameter for fading back to the original center image
has an optional upsize function if a standard HDTV 16:9 resolution is needed.
Delivers a final base64 encoding 24 bit PNG that makes the entire “expand” function a drop-in on an existing base64 string.
Thanks a lot for the advice. I’m not technically advanced enough to be able to tackle such work arounds, but I do appreciate you taking the time to share this.
The topic got a bump of unknown purpose or demonstration.
However, in ChatGPT, as of a week or two ago, you will be able to request an aspect ratio (not a large size) and have it fulfilled (after a long wait with no progress indication):
