Two requests for the same 14 August street differed by one phrase. With “traditional attire” the woman's face took 15.0% and 15.3% of the picture's width in two runs, and 42 and 58 local features matched the selfie. With a named long tunic it took 6.4% to 7.1% in six runs on the same street, and 13 to 23 matched. In these runs the sentence that names the outfit also moved the camera back, and the likeness went with it. This article makes a 14 August Independence Day photo from one AI-generated selfie in CapCut and measures, picture by picture, what each part of the request moved: the outfit words, the framing, the light, the scene and the flag.
Of the 65 scene pictures made with the selfie attached, 10 had a face of 12% of the width or more, and in all of them the match with the selfie was 42 or higher; in the other 55 it ran from 9 to 43, with a median of 24. A light direction written into the request moved the brighter side of the background the way I asked in 13 of 14 runs and the brighter cheek in 10 of 14. One sentence asking for a soft white light from the front brought the night cheek color to within 1.6 to 8.5 of the selfie's in four runs of four. Recolor turned the dark green tunic to a tan in two runs of two when the target was cream, and left it green in two of two when the target was near white. The five wall flags with one crescent were 1.95 to 2.18 times as long as they were high, where the written description gives 1.5. Upscale to 2K removed the colored specks around the eyes in ten pictures of ten.
Before you start. The woman in these pictures is AI-generated, and no real person appears in any of them. A portrait puts a face in front of an audience, so use a photo of yourself, or of someone who agreed to be in it, and label the result as AI-generated where your platform asks for that. The national flag appears only as part of a street scene; this article measures how it was drawn and does not change it, and the lines drawn over some figures are measurement lines, not text. The outfit is a plain long tunic and a scarf described by cut and color. This article does not explain dress or customs, does not cover territory, military or political subjects, and does not cover video or print.
One selfie, five request sentences and, at bottom right, a request with no selfie. Each picture is the first run of its sentence; the selfie tile is cropped to 16:9.
One 14 August sentence in four slots, and the selfie it needs
The pictures came from Image mode of the AI image generator. In the composer I clicked + and chose Upload from device to attach the selfie, which appears as a chip at the start of the request. The settings line read Seedream 4.5, 16:9 and 2K, and after each send the chat showed “Generating 1 item will consume 1 credit.”
The selfie is AI-generated. I made it earlier on the same day in the Design Studio chat from this prompt and downloaded it as a PNG of 1,152 × 2,048 pixels; a JPEG copy of it was attached to each request:
A casual front-camera selfie of a woman in her late twenties with shoulder-length dark hair, a slight smile, wearing a plain white t-shirt, in front of a plain light gray wall. Half-length framing from the head to the waist, head in the center, sharp focus, soft even light. No text, no logos.
Each request was one sentence built from the same four slots. The first two sentences and the last two stay the same; the bracketed parts are what I changed:
Keep the same woman from the selfie: her face, hair and expression. Make a 16:9 photograph of her at a 14 August Independence Day celebration. [She stands in the center of the frame, half-length, facing the camera,] wearing [a long straight cotton tunic in deep green with matching trousers and a white scarf over one shoulder]. Behind her: [a street at night, with strings of warm lights, green-and-white bunting and a few small Pakistani flags across the street, and blurred people celebrating in the distance]. [The main light comes from the left edge of the picture, so the left side of her face is brighter than the right side.] Green and white decorations only. No text anywhere.
The light test needed selfies lit from the side. I made one in Image mode with the 9:16 chip by attaching the selfie and sending the request below, and flipped the result left to right for the other side:
Keep the same woman, the same pose, the same expression, her hair and the same white t-shirt as in the photo. Change only the light: strong direct sunlight from a window at the left edge of the picture lights the left side of her face and the left side of the wall, and the right side of her face and the right side of the wall are in deep soft shadow. Same camera position.
Naming the garment moved the camera back, and the likeness went with it
I measured the face as its width in the picture, and the likeness as a match between the face in the picture and the face in the selfie: local features are matched and the score is how many agree. The measure was tested before use. The selfie scaled so that its face is 100 to 400 pixels wide scored 94 to 135, the selfie flipped left to right scored 7 and two other generated women scored 0 and 2. The faces in these pictures are about 130 to 390 pixels wide, so a low match at a small face is not the measure running out of reach. The same photo sets a ceiling of about 100 or more, and no run came near it.
The first two requests differed in the garment phrase and nothing else. The generic words drew a close, half-length portrait in an embroidered cream or white tunic with a dark green shawl, which I had not asked for, in four of four runs. The named garment drew a plain tunic, and on the daytime street the camera stood far enough back to show her nearly full length, trousers included.
Across the 65 runs, face width and match rose together (correlation 0.84). A framing phrase reached a face of 12% or more in 6 of 12 runs: the close-up wording in 5 of 6, the medium close-up wording in 1 of 2, and the head-and-shoulders and waist-up wordings in 0 of 2 runs each. In the pictures where it worked, the outfit showed from the shoulders to the chest and the trousers dropped out of the frame.
Left: the first run of six wordings. Right: each run with a selfie attached, with the width of the face on the horizontal axis and the match with the selfie on the vertical one.
The choice for a template is therefore which of two things the picture is for. If it is for her face, put a close-up phrase in and keep the garment words short; the outfit shows from the shoulders up. If it is for the outfit, expect a face of about 7% of the width and, in these runs, a match of about 24.
The selfie's light came along, and a written direction moved the scene more than the face
I asked for a main light from the left edge, from the right edge or no light words, with three selfies: the even one, the one lit from the left and the one lit from the right. The numbers are the left cheek's lightness (L*) minus the right cheek's, read in a pair of windows on the face, and the same difference between the left and the right margin of the background, outside a band around the face. Positive means the left side is brighter. The cheek measure was tested on planted pictures: the selfie flipped left to right read −12.9, where the mirror of the original's +13.7 is −13.7; darkening the right half of the picture to 0.8, 0.6 and 0.4 of its values read 25.1, 36.5 and 48.6; and a warm color cast over the whole picture left it at 13.7.
I counted a side as moving the way I asked when its number fell outside the range of the runs for the same selfie with no light words, in the direction asked. On that count the background moved in 13 of the 14 runs where I asked for a direction, and the face moved in 10 of 14. The face followed the words most closely with the even selfie, in five of six runs. When the words opposed the selfie's own light, the face crossed to the other side once in four runs: the selfie lit from the right stayed lit from the right in both runs that asked for the left. With no selfie attached, the background followed the words in six of six runs while the right cheek was brighter in all six, so the cheek numbers for those runs do not follow the words, and I did not find out why.
The three selfies and one night-street result for each request, with the left-minus-right cheek lightness under each face.
For a template that takes other people's selfies, this puts the first requirement on the selfie. An evenly lit, frontal one let the words steer the face in these runs, and a strongly side-lit one kept its side.
A soft white light from the front pulled the night face back toward the selfie
At night with no light words, the cheeks came back darker and yellower than in the selfie. The selfie's cheeks read L* 64.9 and b* 21.5; in four runs without light words they read L* 55.5 to 56.9 and b* 26.5 to 27.0, a color distance from the selfie (CIE76 ΔE) of 10.1 to 11.5. I added one sentence, either “A soft warm light from the front lights her face.” or “A soft white light from the front lights her face.”, and ran each four times.
The white light gave cheeks of L* 57.6 to 67.8 and b* 20.0 to 24.5, at a distance of 1.6 to 8.5 from the selfie. The warm light gave L* 56.3 to 64.0 and b* 27.5 to 60.5, at 11.1 to 42.5; in the first warm run the face turned orange. The color measure was tested on a planted picture: a warm cast added to the selfie read ΔE 15.2. The sentences changed the color and lightness of the face, and they did not raise the likeness: the match was 22 to 31 with either light sentence and 28 to 43 without one.
The first three of four runs per column, with the cheek color and its distance from the selfie's.
The night pictures were not noisier than the day ones. My noise estimate read a median of 0.65 over 36 daytime-street pictures and 0.67 over 43 night-street pictures. It was tested on a picture with Gaussian noise of 0, 2, 5 and 10 levels added and read 0.80, 1.71, 3.46 and 6.45.
Day, night and a plain backdrop all left the face small
The scene slot changed the picture's character and not the framing. With the same named garment and “half-length”, the face took 6.4% to 7.1% of the width on the daytime street in six runs, 7.5% to 8.7% at night in four and 6.8% on the plain deep green backdrop in two. The matches were 13 to 23, 28 to 43 and 18 and 34, and they follow the face widths in the scatter above.
The backdrop is the one scene where the result carried more than the request named. The request said “a plain deep green studio backdrop” and also “Green and white decorations only”, and both pictures added flowers or garlands. I did not run the backdrop request without the decorations sentence.
Colors written in words held, and Recolor changed the tunic and little else
Three colorways were asked for in words: a deep green tunic with a white scarf, a white tunic with a deep green scarf, and a cream tunic with a green hem and cuffs and a light green scarf. Each was run twice on the daytime street, and all six pictures came back in the colors named.
The other way is to keep a picture you like and change one color. In the Design Studio canvas I selected the picture, clicked Recolor, chose Select area, clicked the tunic to place a numbered pin, typed a hex code into the pin's field and clicked Recolor. A cream target (E8DFC8) turned the dark green tunic to a tan in both runs, with the tunic window 60 and 62 in ΔE from the original and the scarf window 1.4 and 1.5; the tan was 20 to 23 from the target, darker than the cream I typed. A near-white target (F4F4F0) left the tunic as it was in both runs (ΔE 1.6 and 2.1) and warmed the scarf (7.1 and 6.2). For each target one pin sat on the torso and one near the hem, with the same result. In all four runs the face changed by 3.2 to 3.4 in mean ΔE and the part of the picture outside the woman by 2.2.
Top: the colorways written in words, two runs each. Bottom: Recolor on the first picture of the top row, with the color difference measured in a window on the tunic and one on the scarf.
The two routes trade differently. A rewritten request is listed at 1 credit and redraws the whole picture, the face included; a Recolor run was listed at 16 credits in the Design Studio canvas on October 8, 2026 and kept the face and the street. If the picture is good and you want one color changed, Recolor kept the rest in these runs, provided the target was not near white. For a near-white target, rewrite the request.
The large flags came out too long, with the band too narrow
The flag is the one object in these pictures with a written design: a dark green field with a white band at the hoist end that is one quarter of the flag's length, a white crescent and a five-pointed star in the middle of the green, and a length of 3 to 2 against the width. I checked that description on October 8, 2026 in the English Wikipedia article on the flag and in a transcription of the Constituent Assembly's resolution on pakistani.org; neither is the standard itself, so I used it as a reference and not as a rule.
I asked for a large flag on a wall in nine pictures (six by day, three at night). In the six daytime pictures I measured each flag's length and height and the share of its length taken by the white band. The measure was tested on planted flags of known proportion: it read 1.50, 1.52, 2.00, 2.00, 1.50 and 1.80 for flags planted at 1.50, 1.50 with wavy edges, 2.00, 2.00, 1.50 and 1.80, and bands of 0.249, 0.258, 0.25, 0.20, 0.299 and 0.219 for bands planted at 0.25, 0.25, 0.25, 0.20, 0.30 and 0.22. Five flags with a single crescent came out 1.95 to 2.18 times as long as they were high, with the band at 20.4% to 22.1% of the length. The sixth had two crescents and was 2.91 to 1 with a band at 18.5%. In one of the five her head sat in the crescent. By eye the crescent opens toward the fly end in the five, with the star beside it. At night the white band sat against a white wall and I could not separate it, so those three are not measured. I did not measure the small flags on the bunting.
The six daytime wall flags with the measured edges. Blue: the band's outer edge. Red: the start and end of the green field.
In these pictures the large flag was not a reliable reference for the flag. If a post needs the flag to be right, check the band against one quarter of the length before posting; the length against the height and the band against the length are easy to read off a picture with a ruler.
Upscale cleared the specks around the eyes and left the face alone
Faces of about 7% of the width carried colored specks around the eyes, purple and green points a few pixels across. I counted pixels whose color differed from their neighbors by more than 9 in the a* and b* channels while their lightness differed by less than 6, in the band across the eyes. On planted pictures with 40 three-pixel specks added, the count rose by 228 to 261. The counts fell into two groups: 8 of the 65 scene pictures had 0 to 30 such pixels, and the other 57 had 61 to 425. The 8 were the four made with the generic outfit words (face about 15%) and four wall-flag pictures with a face of 5.0% to 7.5%.
In the Design Studio canvas I put pictures on the canvas with the upload icon in the bottom toolbar, selected one and clicked Upscale. The menu lists 2K, 4K and 8K, and I ran 2K and 4K and not 8K. The 2K result of the first daytime picture was 3,840 × 2,160 pixels and the 4K result 7,680 × 4,320, with the PNG growing from 6.1 MB to 7.6 MB and 25.3 MB. I then selected five pictures and ran 2K on all of them at once, and later four more; the download of a selection came as one ZIP. In all ten pictures the specks were gone after 2K (0 to 2 such pixels), and the 4K result of the first picture had none either.
The face stayed the same face. The match between the face in the original and in the upscaled copy, shrunk back to the original's size, was 72 to 145 over the ten pictures, and the mean color difference over the face box was 2.7 to 4.4. The edges were not sharper: on the first picture the steepness of edges at the original's scale read 98 to 103 against 100 for the original on the face, the bunting and the tunic, and 92 to 98 on the road. On planted pictures the same measure read 87 and 73 for blurring at 1 and 2 pixels and 140 and 184 for sharpening at two strengths. The small “Ai” mark in the top-left corner was still there, larger.
The same region of one picture in the original, after Upscale to 2K and after Upscale to 4K.
The request and its cost line, the Upscale menu, the Recolor panel and rows of the credit details.
A scene picture plus the selfie as a second picture did not carry the face
A template could be a picture of the scene and outfit made once, with the selfie added afterwards. I made scene pictures with no selfie attached, then attached each one together with the selfie in a new request:
Put the woman from the second picture into the first picture. Keep everything in the first picture: the street, the bunting and flags, the people, the light, her green tunic, her white scarf and her pose. Give her the face, hair and expression of the woman in the second picture.
I ran this for six scene pictures: two at night and two by day with the half-length wording (a face of 5.9% to 6.8%), and two by day with the close-up wording (a face of 10.6% and 10.9%). In four of the six the match with the selfie was 2 to 4, as it was in the scene pictures, so the face did not change. In the two close-up pairs it was 42 at a face of 10.9% and 16 at 7.5%; the second result had also come back at a smaller face than its scene picture.
The request asked to keep the light, and it was not kept. The six results came back yellower than their scene pictures, with b* higher by 2.5 to 16.1. The two night scene pictures came back lighter by 8.1 and 6.1 in mean L*, and the two daytime close-up scene pictures came back as night pictures with strings of lights, 21.0 and 18.4 lower in mean L*. Each composite was a request listed at 1 credit.
Four of the six pairs: the scene picture made without the selfie, and the result of attaching it with the selfie. The match with the selfie is under each picture.
Two requests, one for the face and one for the outfit
No single sentence gave both in these runs. The close-up sentence with the soft white light made a face of 12.7% to 14.5% of the width in three of four runs, with matches of 43 to 58 and the outfit visible from the shoulders to the chest; the fourth run came back at 8.0% with a match of 31. The named full-length sentence gave the whole outfit and the street, with a face of 6.4% to 7.1% on the daytime street and matches of 13 to 23. Cheek color after the white light sentence with the close-up wording was 5.1 to 11.3 from the selfie's, so the 1.6 to 8.5 above did not carry over to the close-up runs.
The template is therefore two requests for each selfie. The first is for the face:
Keep the same woman from the selfie: her face, hair and expression. Make a 16:9 photograph of her at a 14 August Independence Day celebration. It is a close-up portrait from the front: her face and shoulders fill the frame. She wears a long straight cotton tunic in deep green with matching trousers and a white scarf over one shoulder. Behind her: a street in daylight, with strings of green-and-white bunting and a few small Pakistani flags across the street, and blurred people celebrating in the distance. A soft white light from the front lights her face. Green and white decorations only. No text anywhere.
The second is for the outfit, with a half-length sentence and no light words, as tested on the daytime street:
Keep the same woman from the selfie: her face, hair and expression. Make a 16:9 photograph of her at a 14 August Independence Day celebration. She stands in the center of the frame, half-length, facing the camera, wearing a long straight cotton tunic in deep green with matching trousers and a white scarf over one shoulder. Behind her: a street in daylight, with strings of green-and-white bunting and a few small Pakistani flags across the street, and blurred people celebrating in the distance. Green and white decorations only. No text anywhere.
For a night version, swap the scene sentence for the one at the top of this article. Send the first request again if the face comes back at less than about one eighth of the picture's width, which happened in one of the four runs. Then upscale the pictures you keep to 2K, which cleared the specks around the eyes in ten of ten. Check any large flag against its band before posting.
Configuration and cost. Pictures: Video Studio composer, Image mode, Seedream 4.5 on the settings line, 2K, 16:9 (the side-lit selfie at 9:16). Of 82 pictures, 65 were scene pictures with the selfie attached, 10 had no selfie, 6 carried a scene picture and the selfie together, and 1 was the side-lit selfie. After each send the chat showed “Generating 1 item will consume 1 credit.”, except for one request that showed “Generating 4 items will consume 4 credits.” I had not set a count; I stopped that request and the credit details still list 4 credits for it. The chat replies logged for the 82 pictures named Seedream 4.5 in 80 and no model in 2. Tools: Design Studio canvas, Upscale (the credit details list 1 credit for one picture at 2K, 2 for one picture at 4K, 5 for five pictures at 2K and 4 for four pictures at 2K) and Recolor (16 credits for each of 4 runs). The account's credit details for October 8, 2026 list 165 credits for the work in this article between 14:21 and 18:16: 86 Image-mode requests for 89 credits (85 of 1 credit and one of 4), 12 for Upscale and 64 for Recolor. Four of the 86 requests (three of 1 credit and the one of 4) produced no picture I kept. The header balance is not a total of this work, because rows labeled Expired and Task credits also change it. These figures apply to those settings on that date; other models, durations, resolutions and modes are charged differently.
Files. The pictures are 2,560 × 1,440 PNGs. The Recolor results are 2,752 × 1,536, and the Upscale results are 3,840 × 2,160 at 2K and 7,680 × 4,320 at 4K. The selfie is a 1,152 × 2,048 PNG attached as a JPEG copy; the side-lit selfie and its flipped copy are 1,440 × 2,560 and appear only inside a figure. Each picture, the selfie included, carries a small “Ai” mark in its top-left corner, which I left in place. Prompt text is quoted as sent.
How the face was read. The face box comes from two frontal-face detectors run on the upper part of the picture; of the candidates, the one with the most skin-colored pixels is kept, because bunting and lights gave false faces. The match crops the box with a 12% margin, scales it to 240 pixels wide, evens out the contrast and counts local features that agree under one scale, rotation and shift within 5 pixels (ratio test 0.75). Planted checks: the selfie scaled, shifted and degraded scored 129, flipped left to right 7, and two other generated women 0 and 2; the same photo at face widths of 100, 150, 204, 300 and 400 pixels scored 94, 124, 124, 135 and 133. Face width is the box's width as a share of the picture's width.
How the light was read. Cheek lightness is the median L* in a window on each cheek, 24% of the face box's width and 22% of its height, at 10% to 34% and 66% to 90% across and 52% to 74% down. The background measure is the mean L* of the part of the picture left of and right of a band 1.9 face widths either side of the face, over the top 75% of the height; a planted picture bright on the left read +65. The planted checks for the cheek measure are in the text.
How color and Recolor were read. Colors are medians in CIE L*a*b*, and differences are CIE76 ΔE. The Recolor results are 2,752 × 1,536 and the originals 2,560 × 1,440, so the original was registered to each result with an affine fit to 1,088 to 1,142 matched points (median residual 0.6 pixels) and the windows were mapped through it. A planted transform of 1.075 and 1.0762 with a shift of 0 and −6 pixels read back as 1.0750, 1.0762, −0.03 and −6.10, and a planted color patch read 60.2 in the mapped window against 60.2 worked out directly.
How the flags and specks were read. The green field's left and right ends come from the lightness steps along a row band above her head, the band's outer edge from the largest drop in b* between the wall and the band, and the field's height from the dark green mask near each end; the planted flags are in the text. The specks are pixels whose a*b* color differs from the 7 × 7 median around them by more than 9 while their L* differs by less than 6, counted in a band from 20% to 55% of the face box's height and 5% beyond its width on both sides. Edge steepness is the mean Sobel gradient of the luminance.