Gpt-image-1 bias towards cropping with 1024x1024 aspect ratio

[ant1]

I’ve been using gpt-image-1 and noticed that, especially when it comes to rendering people or other tallish things (having a portrait-like aspect ratio), the model seems to have a bias towards cropping off the top of the head and the lower body, even if I add e.g. “full body shot, entire subject in frame”.

When I generate the same exact prompt on a portrait aspect ratio, the subject is not cropped off as expected. Does anybody have a prompting trick to get the model not to crop naturally rectangular content in square images?

e.g. for “firefighter” or “firefighter, entire body in view, object fit to content vertically”, I get the results attached

[ant1]

The square one

[polepole]

Prompt

Ultra‑realistic square photo, full‑body male firefighter, perfectly centered, wide framing (head & boots fully in frame, 5% margin, no cropping). Black boots, turnout pants+reflective tape, flame‑resistant coat, black gloves, helmet+chinstrap. Clean‑shaven, youthful, serious; short hair under helmet. Background: muted concrete firehouse wall, equipment racks softly blurred.
NO cropped, cut‑off head, cut‑off feet, out‑of‑frame, partial body, equipment.

[ant1]

Thanks!! The square image/wide angle and full body / head to toe seemed to do the trick

[jeffvpace]

@ant1 @polepole I hate to spoil the party, but it’s not that simple. If you look closley at the three fireman images, their heads are a bit to big for their bodies - they almost look like midgets. You can’t really prompt your way out of an aspect ratio problem.

There was a deep-dive thread on this awhile back:

Why GPT image API crops original? Please support flexible aspect ratios

You have control over the 32 bit RGBA being sent, and even masking to outfill into that (altho gpt-image-1 mask seems a big lie) You also have the ability to restore an aspect ratio by crop instead of delivering an expansion - or allow the selection of either. Since I seem to have at least a bit of imagination, I’ll apply that as a bit of AI prompting, 1000 tokens input, wanting some reusable code. No guarantee that it actually goes the optimized budget route as intended, because vettiing AI…

@j provided some Python code which does the trick.

[nmurrell011]

(post deleted by author)

[_j]

@jeffvpace

There was a deep-dive thread on this

That forum topic is about the edits endpoint.

This discusses generation itself.

The release notes written by OpenAI do indicate this “issue” of being too zoomed in when generating images, still there since release.

This is likely because there wasn’t separate post-training on each aspect ratio. The AI anticipates and predicts the length of the image, but then goes wrong when constrained. DALL-E 3 also has a similar issue, where you often get square images with truncation that anticipate expanding on either side that never comes when you use 1024x1024.

The strongest way to overcome this is to imagine that when you make an API request for a 1024x1024 image, that the AI has in mind a tall or wide image that it will create for you depending on the subject, for tall, at 1024x1536, where you then only get the center of it.

You have to prompt your way into having more space above and below the subject described in your prompt, more background field that needs to be displayed on either side with contents, or actual object in the “virtual” extra space above and below the person, to receive a final square that is even beyond requesting “zoomed out” or “from a distance” to obtain the goal.

I find the user input technique of requesting side-by-side duplicates works best for tall things.

Image size and framing:

A square image that is split into two halves side-by-side. One half on the left shows the person or character facing us in front profile as the primary image, and on the right half, the person is shown viewed in side profile.

Image contents of each pane:

A photorealistic studio fashion portrait of a tall, slender young woman with fair skin and long, straight platinum blonde hair, parted neatly and flowing smoothly down her back. She has soft, symmetrical facial features, bright eyes, and a warm, natural smile. Her makeup is subtle but polished—light foundation, soft pink lip gloss, gentle eye shadow, and well-defined brows. She wears a fitted, sleeveless mini dress made of rich burgundy satin that reflects light with a smooth sheen. The dress has a clean, elegant cut with a modest V-neckline and tailored waist, flaring slightly at the hem to flatter her silhouette.

She stands confidently on high-heeled black ankle-strap sandals that complement the outfit’s elegance. Accessories are minimal and tasteful: small gold stud earrings and a delicate gold bracelet. The lighting is bright, diffuse, and shadow-free, typical of a high-end fashion catalog shoot. The background is a seamless, clean white sweep with no distractions, ensuring the subject and outfit remain the visual focus. The overall aesthetic is stylish, modern, and professional—ideal for a luxury lookbook or seasonal fashion editorial.

At least she’s not beheaded by the API, but the AI ran out of token output space:

Adding:

Caution: you must zoom out in each image, as if viewed from a distance, and predict the extended height and space needed to show people head-to-toe.

[jeffvpace]

Caution: you must zoom out in each image, as if viewed from a distance, and predict the extended height and space needed to show people head-to-toe.

So you can prompt it! A lot easier than your Python code… However, one must always remember the prompt when you need to use it.