A surreal collage portrait fails in one predictable way: the repeats and the pasted-on objects land wherever they land, and the result reads as a pile rather than a picture. What separates the arranged version from the accidental one is a small number of rules about count, line, one oversized object, and one surface, and each of them can be written into a request and checked on the result. This article makes one such portrait in CapCut at 16:9, a cover you can crop a square avatar out of, from a single generated photo, and measures three versions of it to show which rule does what.
Start from the crop
The cover is 2560 by 1440. The avatar most people will cut from it is a square, and on many services a circle inside that square. So the layout question is settled before the first repeat is placed: one copy of the face has to sit in the middle of the frame with the square around it, and whatever else the collage does happens to the sides. The finished cover below has the square and the circle drawn on it.
The finished cover with the 1440 by 1440 avatar crop and the circle a round avatar keeps. The middle copy and its green eyes sit inside both; the neighbors are what the cover adds.
This order matters because it decides where the oversized object goes: on the copy that will become the avatar, not on a neighbor that the crop removes. It also decides the count. An even number of copies has no middle to crop around; an odd number does, and three is the smallest odd number that still reads as repetition rather than a pair.
Repeat the face three times, on one line
The source is a plain head-and-shoulders portrait on a light gray background, generated for this article rather than photographed, so that the face repeated below belongs to no one. It was attached to the Video Studio composer and converted in Image mode. The prompt writes the arrangement as numbers and geometry rather than as a mood.
Turn this portrait into a surreal paper collage on a 16:9 cover. Cut the head and shoulders out along a rough white paper edge and repeat that same cut-out three times in a straight row across the middle of the frame, evenly spaced, all the same size, with the eyes of all three on one horizontal line. Paste one oversized pair of glossy green paper cut-out eyes over the middle face only, about twice the size of its real eyes. Behind everything a flat mustard-yellow ground with no photo background left, a visible paper grain and light halftone dots over the whole image so all the pieces share one texture, and a small drop shadow under each cut-out. No other objects, no text, no logos.
To see what the numbers buy, the same portrait was converted a second time with the arrangement sentence replaced by its opposite: "repeat that cut-out seven times, scattered across the frame at random positions, random sizes and random tilts, some overlapping", with green eyes over several faces and everything else unchanged.
Three on a line against seven scattered, from the same portrait and the same ground. Both are collages; only one is arranged.
On the ordered cover the three face centers, found by matching the source face against the result, sit 732 and 729 pixels apart, and their heights differ by 3 pixels across a 1440-pixel frame: the eyes are on one line to within the width of an eyelash. On the scattered cover there are seven copies at seven sizes and seven tilts, the green eyes became mask-shaped and landed on every face, and there is no line to measure against because none was asked for. It is a legitimate look, the photo-dump kind, but it cannot be cropped to an avatar without losing most of itself, and it does not get more ordered by adding more copies. The rule, then, is a count and a line: three, evenly spaced, eyes level.
The surreal collage portrait needs one oversized thing, on the middle copy only
The surreal part of a surreal collage portrait is a single object at the wrong scale, and the prompt limits it in three ways: one object, one copy, one size relation. "One oversized pair of glossy green paper cut-out eyes over the middle face only, about twice the size of its real eyes" is that limit written out. On the finished cover the two green discs measure 235 and 231 pixels across against a real eye about 90 pixels wide in the same image, so the result came back at roughly two and a half times rather than two; in this run the ratio was treated as a guide, not as a dimension. The discs sit over the middle copy and nowhere else, which is what keeps the two outer copies legible as the same face and the middle one as the event.
Put the object anywhere else and the crop from the first section breaks: an avatar cut from a neighbor would show an ordinary portrait with a torn edge, and the cover would lose the reason its middle is the middle. The scattered version shows the other failure, where the object is on every copy and so is on none in particular.
One ground and one grain hold it together
Three photos of the same person do not become a collage by being cut out. They become one when they share a surface, and the prompt buys that with two phrases: a flat single-color ground with "no photo background left", and "a visible paper grain and light halftone dots over the whole image so all the pieces share one texture". The third conversion removed both, keeping each cut-out's own photo background and asking for no grain, to see what those phrases were doing.
With and without the shared surface. On the right the three copies read as three framed photographs of one woman, not as one picture.
Measured on the files, 86 percent of the finished cover is within a narrow band of the mustard hue; on the version without the ground, 18 percent of the frame shares any single color. A texture figure that rises with fine grain reads 2.1 on the finished cover, 1.0 on the version without grain and 0.6 on the untouched source portrait, so the grain the prompt asked for is there and is roughly what separates paper from photograph at full size. The no-ground version also did something the prompt did not ask for: a white paper strip runs across all three faces at eye height, and the green eyes came back as a single mask with drawn irises rather than two pasted discs. Without a ground to sit on, the collage grammar started inventing its own connective tissue. One color and one grain are cheaper than that.
The face you are allowed to repeat
The portrait here was generated on 8 September 2026 in Video Studio's Image mode through the CapCut AI image generator flow, at Seedream 4.5, 16:9 and 2K, from a text description, and came back at 2560 by 1440 with the face centered and room on both sides, which is the framing the row needs. The first request for it returned a card that read "Couldn't load content" with the download button disabled; the same prompt was sent again and came back normally. Each collage was then made by dropping that portrait onto the home composer, setting the mode menu to Image, confirming the settings row read Seedream 4.5, 16:9, 2K, and sending the prompt above.
The composer in Image mode with the portrait attached. The row has to read 16:9, or the cover comes back in the portrait's shape.
The cost line before the portrait and before the collage, both 1 credit at Image mode, Seedream 4.5, 16:9 and 2K on 8 September 2026.
Before each of these images the chat printed "Generating 1 item will consume 1 credit", at Image mode with Seedream 4.5, 16:9 and 2K; that line is the only credit figure used here, other models, resolutions and modes are priced differently, and video is priced separately. With a real photo, the face you repeat three times is a real person's, multiplied: use your own, or one whose subject has agreed to be a collage, and treat the cover as you would the photo. The method does not care whose face it is; the people in it do.
Cut the avatar from the cover
The last step is the crop the first section planned for: a 1440 by 1440 square centered on the middle copy, which in this file runs from 565 to 2005 pixels across the full width. Inside it are the middle face, both green discs, the torn paper edge around them and a margin of ground; the neighbors show as a sliver of paper at each side, which tells a viewer of the square that there is a wider picture, and a round avatar trims even that away. Make the cover first and cut the avatar from it, not the other way around, because the avatar cut from the cover carries the ground and grain that make it read as a collage even at a thumbnail's size.
What this article does not cover: background removal, inpainting individual pieces, upscaling for print, and any layout done in design software. The three rules above were applied by wording alone and checked on the downloaded files, and that is where the next cover should start: a count, a line, one object on the middle copy, one ground.
Written 8 September 2026. The portrait and all three collages are generated images from CapCut made for this article; no real person, product or brand appears. Interface labels, the cost line and pixel sizes reflect a single session on that date and the configuration named above, and may change. Face positions were found by matching the source face against each result; color share and the texture figure were computed on the downloaded files.