This is one of the few image jobs with a published yardstick, so the yardstick comes first. Identity photographs are judged on three things that can be counted rather than argued about: how much of the frame the head occupies, whether the background is a single even color, and — if you are making more than one — whether they look like they came from the same session. Those are the criteria this article measures, in that order.
Four portraits were generated from a single prompt through the AI image generator, on Seedream 4.5 at 16:9 and 2K, 1 credit each, plus one two-person frame. The prompt asked for the standard thing.
A passport-style headshot of an adult woman in a plain dark top, facing the camera straight on with a neutral expression, eyes open, shoulders square to the lens, against a plain even mid-blue background. Flat frontal studio lighting with no shadow falling on the background. Nothing else in the frame. No text, no logos, no printed words and no brand marks anywhere in the frame. 16:9 landscape.
Each result was cropped to 3:4 around the subject, which is the shape this kind of picture is used at, and then measured.
Head size is the part that moves
Head height was taken from the silhouette against the plain background: the top of the head is the highest subject pixel, and the neck is the narrowest row between the crown and the shoulders, found from the width of the silhouette row by row. The distance between them, divided by the height of the 3:4 crop, is the number below. Doing it from the outline rather than from face detection keeps it reliable on a plain background.
The same prompt, run four times. Head height as a share of the crop.
Two of four land in the band, one is just under it, and one is a long way under. The spread from smallest to largest is 17.8 percentage points, from a prompt that did not change between runs. Nothing in the wording asked for a head size, which is the point: the words “passport-style headshot” do not carry a number, so the framing is redrawn on each run.
The fix is not a better adjective. It is to stop asking the generator for the framing and take it yourself: generate wider than you need, then crop to the head size you want. A 16:9 frame has room to crop a 3:4 out of, and the head fraction becomes something you set with a crop box instead of something you hope for. This is also why generating at 16:9 rather than at the final shape is worth doing even when the final shape is tall.
The blue is even, and it is not the same blue twice
Background uniformity — the thing people mean by a clean studio backdrop — came out well. Within each frame, the spread of lightness across the background sits between 0.77 and 2.24, which is flat enough that no banding or vignetting is visible. On that measure the generator does the job.
Left: variation inside each frame, which is small. Right: the blue each run settled on, which is not consistent.
Three runs cluster tightly around b* −21, a muted teal-leaning blue. The fourth is at −41, which is close to twice as saturated and reads as a different color sitting next to the others. The phrase “plain even mid-blue” is doing less work than it appears to: it reliably produces an even blue and does not pin down which one.
For a single portrait that does not matter. For the job people actually want this for — a set, one per family member, posted together — it decides the result, because the mismatch only becomes visible once the pictures are side by side, which is after all of them have been made.
What this leaves out, and why
Two parts of the obvious brief are missing here on purpose.
The first is the parent-and-child version, which is a popular form of this and which is not demonstrated in this article. Generating document-style portraits of children is a combination worth declining: the output is already not valid for any official use, and a synthetic ID-style picture of a minor has uses nobody wants to supply a recipe for. The technique below works the same whoever is in frame; the demonstration stays with adults.
The second is an upscaling step. Descriptions of this workflow routinely include one, and in the web workspace there is no image upscaler to include: the tools page lists eight items — text to speech, voice changer, batch edit, Video Studio, remove background, resize video, video stabilization and super slow motion — and none of them enlarges a still. Enlargement exists inside the Design Studio canvas as a tool on a generated image, which is a different place from where instructions usually send you. Generating at 2K to begin with sidesteps the question.
Make the set in one frame
Everything above points at one change, and it is a change to the unit of work rather than to the wording.
The background inconsistency is between frames rather than inside one. The head-size lottery is redrawn per frame. Both problems are properties of running the generator more than once — so run it once. Ask for everyone in a single wide frame, standing in a row against the one background, lit by the one light, and then cut the individual portraits out of it afterwards with a crop box. The blue is then identical across the set by construction rather than by luck, and the head sizes are set by where you put the crop rather than by what the model decided that time.
A passport-style photograph of two adults standing side by side, both facing the camera straight on with a neutral expression, eyes open, shoulders square to the lens, against a plain even mid-blue background. Flat frontal studio lighting with no shadow falling on the background. Nothing else in the frame. No text, no logos, no printed words and no brand marks anywhere in the frame. 16:9 landscape.
One frame, one background, one light. The two portraits cut from this will match each other, which four separate runs did not.
One caveat from measuring it: the two-person frame's background was the least uniform of the five, at 4.24 against 0.77 to 2.24 for the singles. That is a single observation rather than a pattern, and the figure is still low in absolute terms, but it is the direction to watch if you push to four or six people in a row.
What this costs
Five credits. Two further attempts aborted before sending, when a check confirmed the aspect ratio had not been applied, and cost nothing. The figures are what the interface showed for this configuration on 16 September 2026; the panel states the cost before you send.
Written 16 September 2026. Every image was generated in CapCut for this article; no photograph of a real person was used. Results are 2560 by 1440 PNGs from Seedream 4.5 at 16:9 and 2K, each cropped to a 3:4 window of full frame height, centered on the subject, before measurement. The background is the set of pixels whose CIELAB b* falls below −18, cleaned with a 9-pixel opening; the subject is the largest filled component of the remainder. Head height runs from the highest subject pixel to the narrowest silhouette row found between 12 and 55 percent of the way down the subject, which is the neck; the resulting figure is that distance over the crop height. Face detection was deliberately not used, since a plain background makes the outline the more reliable signal. Within-frame background spread is the standard deviation of L* over background pixels inside the crop. A centring measure was computed and then discarded before reporting: the crop is centered on the subject by construction, so the number was zero in every case and described the method rather than the picture. The 50 to 69 percent band is quoted as a common requirement for orientation only and is not the standard of any particular authority; acceptance rules differ by country and none of them accept generated images in any case. The two-person condition was run once, so its background figure is a single observation.