This article turns one AI-generated selfie into an Independence Day poster in CapCut, at 16:9, 1:1 and 9:16, and measures what 25 posters did where the slogan, the portrait and the date meet. In nine requests that gave no size for the top band, the woman's hair covered the letters P, E and N of the slogan, including the request that told her head to stay inside the white middle band. Describing the band as slim, one fifth or one tenth of the poster height, kept the hair off the letters in seven of nine requests.
Both lines were spelled as asked on all 25 posters. The red band came back darker than my reference red, #D61F26, on 24 of them, and the white around the portrait differed from the white in the bottom band on all 25. A hex code typed into the request for the red was printed onto the poster in all five requests that carried one. The sections below follow these problems in the order you meet them, then take the outfit, a cut-out of the selfie, the three sizes and a phone-size check.
Before you start. The woman in these posters is AI-generated, and no real person appears in any of them. A poster puts a face in front of an audience, so use a photo of yourself, or of someone who agreed to be on it. If you post a poster like this, label it as AI-generated where your platform asks for that. The slogan is a generic greeting, and the posters carry no flag, national emblem or official slogan; this article does not cover those, print production or video. 17 AUGUST 2026 is the example date typed in these runs, so replace it with the date you need.
One request sentence sent with the 16:9, 1:1 and 9:16 chips and the selfie attached. The 16:9 band is 19.4% of the poster height, the 1:1 band 18.9% and the 9:16 band 31.6%.
One request drew all three bands of the Independence Day poster
The selfie is AI-generated. I made it in the Design Studio chat from this prompt, downloaded it as a PNG at 1,152 × 2,048 pixels and attached a JPEG copy of it to each request that follows:
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.
The posters came from Image mode of the AI image generator. Click + in the composer, choose Upload from device and pick the selfie; it 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 the line “Generating 1 item will consume 1 credit.” This is the plain request, the base for everything below:
Keep the same woman from the selfie: her face, hair and expression. Make a 16:9 Independence Day poster in three horizontal bands from top to bottom. Top band: a solid red band across the full width with the slogan "HAPPY INDEPENDENCE DAY" in bold white capital letters. Middle: a clean white background with her half-length portrait centered, wearing a red blazer over a white t-shirt. Bottom band: the date "17 AUGUST 2026" in bold red capital letters on white. Only red and white. No flags, no emblems, no other text.
What came back right is the structure. All 25 posters had a red band near the top with white capitals, a middle band with a portrait and a bottom band with the date in red. I read both lines back with a text recognizer: the date came back as written on all 25 posters and the slogan on 20. On the other five, her hair covers part of one letter, and I checked those five by eye; the slogan is spelled as asked on all of them. Her face sat at 49.9% to 50.9% of the width on 23 of the 25 posters, and at 62% and 63% on the sweater poster and one of the cut-out posters below.
The woman in the posters is a redrawn version of the woman in the selfie. A local-feature match between the face in each poster and the face in the selfie scored 35 to 69 on 24 of the posters, against 129 for a copy of the selfie that I scaled, shifted and degraded, 7 for the selfie mirrored, and 0 and 2 for two different generated women. One poster scored 3; it is the one-third band request below.
Each poster also carries a small “Ai” mark with a sparkle in its top-left corner, 44 × 48 pixels with its left edge at x 42 and its top at y 36, in the 2,560-pixel-wide files and the 1,440-pixel-wide files alike. The request said no other text, and the mark is there anyway. The Design Studio chat poster further down has none. I left the mark in place in each figure.
A slimmer band kept her hair off the slogan; a rule about her head did not
In the plain request the portrait rose into the top band, and her hair covered the lower part of the letters P, E and N in the middle of INDEPENDENCE. The same happened in all nine requests that gave no size for the band: the plain sentence twice, the plain sentence with a hex code for the red five times, the plain sentence with “one bright red, one pure white” once, and once with a sentence that told her head to stay out of the band. On those nine posters the top of her head rose 37% to 47% of the way up the band, measured as a share of the band's height. On the poster whose request told her head to stay inside the white middle band, it rose 47%.
What changed the result was a different description of the band, or of the portrait. The table lists each 16:9 request by what I added to the plain sentence, with the 21 posters they produced.
In the slim-band requests the top-band sentence became this one, with the rest of the request unchanged:
Top band: a slim solid red band, one fifth of the poster height, across the full width with the slogan "HAPPY INDEPENDENCE DAY" in bold white capital letters.
Where the band came back at 19.4% to 19.6%, on five posters, the hair stayed off the letters on all five. Where it came back at 27.9% to 30.3% with a portrait of ordinary size, on 13 posters, the hair touched them on 11. The band did not follow the numbers closely: the one-tenth request came back at 19.4%, the same band as one fifth, and one third came back at 50.0%. On that poster the top of her head rose 71% of the way up the band, and her face scored 3 in the match, so I do not count it as her.
The small portrait kept the hair clear in both requests that asked for one, but not in a form I would use. In both her face took 8% of the poster's width, against 11% to 15% on the other 16:9 posters. In the first of the two she also sat in a rectangle of whiter wall (L* 99.2 against 96.1 around it) with a visible edge. In the second the slim band was not followed, and the band came back at 31.7%.
Six requests, one poster each: the plain sentence, the sentence that tells her head to stay in the white band, a small chest-up portrait, a slim band, both of those, and a one-third band.
The hex code I typed was printed on the poster
To get a particular red I typed its hex code into the band sentence, as “a solid red band (#D61F26)”. All five requests that carried the code printed it in the band. In four it sat near the top middle, above the slogan: dark red in run 1, where it showed only after I stretched the contrast, and white in runs 3 to 5. In run 2 it was white in the top right corner. I read the code with the recognizer and by eye: the # was read as @ in run 1, and in run 5 it was drawn as a broken glyph.
The 20 requests without a code printed nothing in the band that my detector could find. The detector found a planted white code at 40 and at 24 pixels tall and a planted code 12 levels lighter than the band, and it missed one 8 levels lighter, so a fainter mark is not ruled out.
The Eraser removed the printed code. In the Design Studio canvas I selected the poster from run 2, clicked Eraser in the image toolbar, brushed over the code and clicked the check mark. The credit details list that run as 1 credit (an Edit image row, October 8, 2026). Afterwards the recognizer found no text in the corner, and outside the brushed corner 98.2% of the pixels stayed within 10 levels of the original. The shorter fix is to leave the code out, because the same sentence without it printed nothing.
The typed code in two runs (panels 1 to 3) and the same crop from a request without it (panel 4). Panels 2 and 4 are copies of the band with the green channel stretched.
The red came back darker than my reference red, and the poster has two whites
My reference red is #D61F26, the code I typed into five of the requests, and its lightness is L* 46.3. I took the median color of the red pixels in each band. On 24 of the 25 posters it was darker, at L* 36.1 to 40.0, and its distance from #D61F26 (color difference, CIE76 ΔE) was 9.7 to 15.5, with a median of 11.6. A typed code, a color word or a slim band did not move the numbers out of that range. The 25th poster was the one whose request said “one bright red, one pure white”: its band matched the lightness (L* 46.3) but came back more saturated, #DC0D01, which put it 15.5 from #D61F26, the farthest of the 25.
The darker red has one advantage. White letters on the band measured a contrast of 6.4 to 7.5 on 24 posters and 5.1 on the bright-red one, by the relative-luminance formula, against 5.14 for white on #D61F26 itself; the red date on the white measured 6.4 to 7.6 and 5.1. If a deeper red suits your poster, the contrast is not the problem. If you need a particular red, the next two sections show what two routes did about it.
The colors of six posters against the reference red and white. Each swatch is the median of the pixels in its area.
The poster also has two whites. The white around the portrait measured L* 93.7 to 96.9 on the 25 posters, and the white in the bottom band measured L* 100.0 on 24 of them and 98.0 on the 1:1 poster. The color difference between the two whites was 1.5 to 6.4, with a median of about 4, and it was 3 or more on 24 posters. The selfie's own wall measures L* 70.2, so the light wall in the posters is a redrawn wall, not the selfie's. At full size the step is visible where the wall meets the bottom band, and the contrast-stretched copies in the figure below show it as a gray panel.
Asking for one white did not close the gap. The “one pure white” request came back with a wall of L* 96.1, 4.2 from the bottom band. I also asked for the selfie itself to be edited so that its gray wall became a plain white one (Image mode, 9:16, one sentence), and the wall came back at L* 92.9. I did not find a request that made the two whites match.
Three posters as delivered (top) and with only the light end of the lightness scale stretched (bottom). The numbers under each are the wall's L*, the bottom band's L* and the color difference between them.
Recolor moved the red closer once in three runs
Recolor is in the toolbar of an image selected on the Design Studio canvas. Its panel has two tabs, Entire image and Select area. I used Select area, clicked the blazer and then the empty part of the band to place two numbered pins, and typed D61F26 into the hex field of each pin. I did not try Entire image.
Three runs gave three results. On the slim-band poster the band moved from #BE0300 (11.6 from #D61F26) to #CC1215 (6.0), and the blazer from #AC0A13 (13.3) to #BF181E (8.1). The same two pins on the poster I had cleaned with the Eraser returned a white band with a red slogan. A single pin on the wall at the left, set to FFFFFF, whitened the blazer, left the band at #C00201 (11.2 from #D61F26) and moved the wall from L* 96.5 to 96.9, which took its difference from the bottom band from 3.53 to 3.14. All three results came back at 2,752 × 1,536 pixels, and the posters I sent in were 2,560 × 1,440. In the Design Studio canvas on October 8, 2026, the credit details list three Recolor rows of 16 credits.
For comparison, the 25 Image-mode requests (Seedream 4.5, 2K, a cost line of 1 credit on October 8, 2026) landed 9.7 to 15.5 from #D61F26, so sending the request again did not close the gap either.
The printed code before and after the Eraser, the slim-band poster before Recolor, and the three Recolor results.
The design chat kept the red and boxed the portrait
The other route is the Design Studio chat. I selected the selfie on the canvas, sent the slim-band request (the one for 16:9, with no hex code) and downloaded the design card that came back. Its band was #D71F27, 0.4 from #D61F26. An earlier message in the same chat had carried that code, and the reply to my follow-up said it kept #D61F26, so the 0.4 is what the chat did with a code it had already been given. I did not test a chat with no code anywhere in it.
The type was the problem. In the first design the slogan's capitals were 4.3% of the poster height and the date digits 2.9%, against 12.0% and 15.0% on the Image-mode slim-band poster. I followed up with this message:
Make the slogan and the date much larger: the slogan's capital letters should be about one seventh of the poster height. Leave a gap of white space between the red band and her head. Keep the same red and the same white.
The next design had slogan capitals at 9.4% of the height, against the 14.3% that one seventh would be, and date digits at 5.9%. In both designs she is a figure cut off at the thighs, in a rectangle of wall (L* 96.1) on a white page, 4.1 and 3.9 from the white around it. In the first the top of her head sits against the band's lower edge, so the follow-up also asked for a gap. Each download was a one-page PDF, 960 × 540 points, holding a single 1,920 × 1,080 JPEG, with no text and no fonts in it, so the type is part of the picture. The chat's replies ended with “18 credits consumed” and “15 credits consumed”, and the credit details list the same two amounts. The first reply also described scale and position sliders for the portrait; I did not use them.
The chat's second design (left) and the Image-mode slim-band poster (right), with the type sizes measured as a share of the poster height.
A different outfit also changed the band and where she stands
The outfit is one clause of the middle sentence. I changed only that clause in three requests that otherwise repeat the slim-band sentence: her white t-shirt alone, a red sweater, and a white shirt with a red scarf. Each outfit is one request, so the figure shows four posters, not a result about outfits.
With the blazer the band came back at 19.4% and the hair stayed clear. With the white t-shirt it came back at 19.6% with the hair clear, and nothing below the band was red. With the sweater and with the scarf the band came back at 27.9% and 28.2%, which is not slim. On the sweater poster her face sat at 62% of the width instead of 50%, and on the scarf poster her hair covered P, E and N again. The blazer's red measured 13.4 from the band's red, the sweater's 7.1 and the scarf's 6.6. After you change the outfit, check the band's height and her head again, as you would after any other change.
Four outfit clauses in the same slim-band request. dE is the color difference between the outfit's red and the band's red.
Cutting her out first did not help
One way to start this kind of poster is to cut the person out of the photo. In the Design Studio canvas I selected the selfie, clicked Remove BG in the image toolbar and chose Auto. The credit details list 1 credit. The result is a PNG with a transparent background at 1,152 × 2,048 pixels, and on a red fill its hair edge shows no gray fringe in the crop in the figure.
I then attached the cut-out instead of the selfie to two requests with the slim-band sentence. Neither poster did better than the three that used the selfie. The bands came back at 28.1% and 28.2%, against 19.4% to 19.6%. In the first run her hair covered P, E and N, and in the second the top of her head rose 24% of the way up the band, with her face at 63% of the width. The whites differed by 5.2 and 6.43, against 3.53 to 4.05, and the walls measured L* 94.9 and 93.7, against 96.1 to 96.5. In this run the cut-out was not a repair for the collision or for the two whites.
The selfie and its cut-out on red, then the slim-band request without and with the cut-out attached. dE is the difference between the wall and the white of the bottom band.
Ask for each size; Expand left the poster as it was
The same slim-band sentence went out three times, with the ratio chip set to 16:9, 1:1 and 9:16 and the sentence naming the same ratio. The downloads were 2,560 × 1,440, 2,048 × 2,048 and 1,440 × 2,560 pixels, and they were three different layouts. The 16:9 band was 19.4% of the height with the slogan on one line. The 1:1 band was 18.9% with the slogan on one line and a strip of white above it. The 9:16 band was 31.6% with the slogan on three lines and the date on two. The plain sentence gave other layouts at the two ratios: a 1:1 band of 28% with the slogan on two lines, and a 9:16 band of 39.4% with the slogan on three. Her face took 19% to 20% of the width on the square posters and 25% to 33% on the vertical ones.
Expand is the other way to get a second size. In the Design Studio canvas I selected the 16:9 poster, clicked Expand, opened the ratio menu in the Expand bar, chose 1:1 and clicked the check mark. The credit details list 1 credit. The result is a 2,560 × 2,560 canvas with the poster at rows 560 to 1,999 and 560 rows of white above and below it. The poster was not laid out again: her face took 12% of the width, against 19% to 20% on the native square posters, and the first 11 rows of the band came back white. The added white measured L* 99.6 above and 100.0 below. In this run a ratio chip gave a poster composed for its shape, and Expand gave the 16:9 poster in a taller canvas.
The 9:16 and 1:1 posters from their own chips, and the 16:9 poster after Expand to 1:1.
The request, the toolbar and panels used, and four credit-details rows.
Read each export at 135 pixels wide before you post it
A poster is judged at the size it is shown. I shrank the 16:9 poster from the slim-band request, the 1:1 poster and the 9:16 poster to widths from 360 pixels down to 90, with area averaging, and read the slogan and the date with the text recognizer at each width, requiring a match with the text as written. In a planted test with a stand-in font, capitals planted 8 pixels tall (measured at 9) were read and capitals 6 pixels tall were not.
At 360 pixels wide the slogan's capitals were 24 pixels tall on the 16:9 poster, 29 on the 1:1 poster and about 47 on the 9:16 poster. The 16:9 and 1:1 posters were read down to 112 pixels wide, where their slogan capitals were 7.6 and 9.1 pixels tall, and not at 90. The 9:16 slogan was read at each width, 90 included, but its date was read as T7 AUGUST 2026 at 112 pixels and as 7 AUGUST 2026 at 90. At 135 pixels all three posters were read in full.
The three posters at 360 pixels wide and at the narrower widths I tested.
The recognizer is not a viewer, so look at the 135-pixel copy yourself as well. On each of your three exports, shrink a copy to 135 pixels wide, read the slogan and the date, and look at the face. Where a line fails, do not post that size. Describe the band or the type differently, send the request again and read the new poster at the same width.
Configuration and cost. Posters: Video Studio composer, Image mode, Seedream 4.5 on the settings line, 2K; 21 at 16:9, 2 at 1:1 and 2 at 9:16, each with the selfie attached. After each send the chat showed “Generating 1 item will consume 1 credit.” I logged the chat replies of 23 requests (22 posters and the white-wall edit of the selfie): 13 named Seedream 5.0 Pro and 10 named Seedream 4.5. I report the settings line and do not know which model drew which poster. Tools: Design Studio canvas, with the credit details listing 1 credit for each Remove BG, Eraser and Expand run and 16 credits for each Recolor run. Design Studio chat: two turns, whose replies ended with “18 credits consumed” and “15 credits consumed”. The selfie was a Design Studio chat generation, listed at 25 credits. The account's credit details for October 8, 2026 list 228 credits in charges for everything I ran for this article, which includes 86 credits on three Design Studio chat turns that produced nothing I used (two during a connection problem, one from a stray click on a suggestion chip), one Image-mode request whose poster I did not keep (1 credit) and one that produced four images by mistake (4 credits). The header balance is not a total: it fell by 253 at 12:00 when six rows labeled Expired (bonus) posted, and rows labeled Task credits added credits at several points without any action from me. These figures apply to those settings on that date; other models, durations, resolutions and modes are charged differently.
Files. The 16:9 posters are 2,560 × 1,440 PNGs, the 1:1 poster is 2,048 × 2,048 and the 9:16 posters are 1,440 × 2,560; the square and vertical posters appear only inside the composite figures. The Recolor results are 2,752 × 1,536 and the Expand result is 2,560 × 2,560. The Design Studio chat's download was a PDF; the PNG in the assets is that PDF converted at 2,560 × 1,440. The selfie is a 1,152 × 2,048 PNG, attached as a JPEG copy. Prompt text is quoted as sent. The woman is AI-generated, and 17 AUGUST 2026 is an example date.
How the layout was read. The top band is the first run of rows whose strips near the left and right edges are mostly red. Slogan letters are the white blobs inside the band that are at least 45% as tall as the tallest, and a slogan counts as whole at 20 letters. How far her head rose is the largest dark, non-red blob inside the band, from its top to the band's lower edge, as a share of the band's height. A letter counts as touched when dark, non-red pixels cover more than 3% of its box. On a planted poster with a red band and a dark ellipse over it, the clean version read 20 letters and no head, and the version with the ellipse read 19 letters and 171 pixels of head inside the band. Type sizes are the median height of the glyphs as a share of the poster height.
How the printed code was looked for. Inside the band, the slogan and the head are blanked, the green channel is stretched over the band's own range, lines of at least four glyph-sized blobs are read, and white blobs smaller than the slogan's letters are read separately. On planted bands the detector found a white code at 40 and at 24 pixels tall and a code 12 levels lighter than the band, and it found nothing on a band with no code or with a code 8 levels lighter. The share of pixels within 10 levels of the original after the Eraser is the share whose largest per-channel difference is 10 or less, outside the brushed corner.
How color was read. Each color is the median RGB of the pixels in its area (red pixels in the band; the white margins at the sides of the middle band and of the bottom band; the red pixels of the date), converted to CIE L*a*b*, and the color difference is the straight-line distance between two such colors (CIE76 ΔE). Planted checks: white against black read ΔE 100.0 and white read L* 100.0, and the contrast formula, (L + 0.05) / (Lb + 0.05) on relative luminance, read 21.0 for white on black and 4.54 for #767676 on white. The stretched copies map L* 90 and below to black and L* 100 to white.
How the faces were matched. A face box was found on the selfie and on each poster, the crop with a 12% margin was scaled to 240 pixels wide and evened out in contrast, and local features were matched (ratio test 0.75); the score is the number of matches that agree under one scale, rotation and shift within 5 pixels. The planted checks are in the text above.
How the small sizes were read. The slogan and date crops come from boxes measured on the full-size poster. Each crop was shrunk by area averaging to the width tested and read with a text recognizer on its green channel (inverted for the white slogan), with a letters-and-digits whitelist, and compared with the text as written after spaces were removed.