Turning two everyday photos into one couple portrait is a solved step, and it is covered in the walkthrough for merging two solo photos. This article starts after that, at the part people ask about next: you have one good couple image, and you need a set — a full-length shot for the invitation, a close-up for the announcement, a wide landscape frame for the album cover — and then a save-the-date card with two names and a date on it, which is where the interesting trouble is.
Eleven stills, one generated clip and two editor exports were made to find out which parts of that job the model does and which parts it hands back to you. All of it ran on the AI image generator and the AI video generator at 16:9.
What came back across a set, and what did not
The first six frames were the same couple in three framings, twice: three described only by their framing, and three more with a locked style block appended.
Shared across all six. A photorealistic photograph of an engaged couple: a woman in a simple ivory silk wedding gown with a plain bodice and no beading, and a man in a charcoal gray two-piece suit with a white shirt and no tie. They are on a wide stone terrace overlooking a calm green lake at golden hour, warm low sun from camera left, no other people, no signage. The gown design and the suit are exactly the same garments in every shot.
The style block, added to the second three only. Visual spec held constant: warm golden-hour color temperature, matte silk and fine wool fabrics, the same stone terrace and green lake.
The three framings, and the bride from the first two of them.
The world held. All three frames have the same stone terrace, the same green lake, the same low warm sun from the left, an ivory satin gown and a charcoal suit. Put side by side they read as one location on one afternoon, which is the thing the brief asked for.
The people did not. In the full-length frame the bride has dark hair pinned up; in the close-up she has long loose auburn hair down her back. The neckline changes too — thin straps in one, a wider cut in the other. The second three frames, with the style block, behaved the same way.
The pattern is worth stating plainly because it explains both results at once. Everything that was written into the prompt came back. Everything left unwritten was invented again. Gown, suit, terrace, lake, light: all described, all reproduced. Hair, neckline, face, exact pose: none described, none carried over.
The same test in a different wardrobe
A reasonable objection is that the ivory gown is a vague thing to describe, and that a garment with real structure would pin itself down. Two more frames tested that with Chinese wedding attire, which is far more specific: crimson silk, a mandarin collar, frog closures, gold phoenix embroidery. Same terrace, same lake, same golden hour, one full-length and one close-up.
The described parts came back in both. The four crops are the parts nobody wrote down.
Everything named held. Both frames have crimson satin, a mandarin collar, frog closures, gold phoenix embroidery, the stone terrace and the green lake at golden hour. A stranger would take them for one outfit.
The details underneath moved. The bride's cuff carries a flat scrolling gold band in the full-length frame and a full phoenix in the close-up. The groom's closures are gold bars in one frame and self-colored red knots in the other, and his hair changes from a neat side part to a fuller swept-back cut. A more specific vocabulary raised the floor — the garment family is right in both — and it did not settle the detail, because the detail is the part nobody wrote down.
Why the style block did not help
The obvious fix for an inconsistent set is a style block: name the color temperature, the fabrics and the location, and paste it onto every shot. That is what the second three frames did, and it changed nothing about the problem.
It could not have. A style block is more description, and description was already the part that worked. Color temperature, fabric and location were being reproduced faithfully before the block was added, because they were already in the shared paragraph. Adding them a second time restates something that was not failing.
The failing part is the part that cannot be written down. You can describe a gown well enough to get the same gown. You cannot describe a face well enough to get the same face, and nobody can — that is what a face is. The same goes for the particular way hair sits, or where an embroiderer put the phoenix. So any amount of writing will reproduce the wardrobe and the weather and reinvent the person wearing them.
Carry the couple as an image, not as a sentence
Which points at the one move that does work, and it is not a prompt change. The couple has to travel between shots as a reference image rather than as a description — the same principle that keeps the same characters across a series. Generate or merge the couple once, then use that frame as the input for each subsequent shot, changing only the framing instruction.
The practical order for a set, then:
The shared paragraph still earns its place — it is what kept the terrace and the light steady across all eight frames here. It just cannot do the job people hope it will do, which is to hold a person together.
The poster, and the words the model insists on writing
The save-the-date itself was the next step: one of the full-length frames, sent back through image-to-image with a double-exposure brief. The prompt asked for a layout with an empty lower third, and it was explicit about lettering.
A vintage save-the-date poster: the couple as a dark silhouette double-exposed with the lake and hills, warm faded paper texture, a thin white rule inset from the edge, and a clean empty pale panel across the lower left. No text, no letters, no numbers, no logos, no printed words.
The negative instruction was ignored, and what came back was misspelled.
The layout arrived as asked. So did three lines of type nobody ordered: WEDDING, Save-Date, and underneath, SAVE-OFE DATE. The last one is not a phrase in any language, and the middle one is missing a word. This is the standard failure mode of generated lettering, and it is why the negative instruction is worth including even when it is ignored: the model is drawing shapes that look like a wedding poster, and wedding posters have words on them.
Removing the type took one more image-to-image pass at 1 credit, asking for the same picture with the lettering gone and the panel left clear. That worked: the silhouettes, the lake, the hills, the rule and the faded grade all survived, and the lower left came back empty.
There is no masked edit here, so plan around that
Which raises the obvious question: if image-to-image can take the words off, can it be pointed at one small area and told to leave the rest alone? A separate test says no, and the answer is worth the credit it cost because it changes how you sequence the work.
The instruction was as narrow as it could be made: redraw the four hands, and <em>do not alter their faces, their clothing, the terrace, the lake or the lighting in any way</em>. The reference was the full-length frame from the set above.
The region the instruction named moved least. The regions it protected moved most.
Measured against the reference in CIELAB, over four regions checked on both frames before the numbers were taken:
The hands came back in the same pose with the same fingers. The hillside came back as different vegetation with buildings in it. The balusters changed shape and spacing, the paving changed its joint pattern, and the plaque on the wall changed color. The picture is a new render of a similar scene, and the scope in the sentence did nothing.
The practical reading is a sequencing rule. Treat every image-to-image pass as a re-render of the whole frame, which means: fix things in descending order of how much of the picture they touch, accept that anything you were happy with may come back different, and stop as soon as the frame is good rather than making one more small correction. If a detail has to be exact, it does not belong in the generated layer at all.
Setting the names and the date
Two names and a date are the definition of a detail that has to be exact, and the poster above has already shown what the model does with words. So they go in the editor, on top of the picture, where they stay as typed.
Text, Add heading, then the size and fill controls. The checkbox at the bottom is the one to watch.
The route is Text in the left rail, then <em>Add heading</em> for the names and <em>Add body text</em> for the date. Both land in the middle of the frame at a default size and in white, and both need the same three changes: drop the size until the line fits the clear panel, set the fill to something dark enough to read on a pale ground, and drag the block into the panel.
One control deserves attention. <em>Apply to all</em>, at the bottom of the Style section, was switched on by default and it does what it says — setting the date line to a small size pulled the names down to match. Turning it off first, then styling each line separately, avoids undoing the work twice.
The finished card, exported at 3840 by 2160.
Export sits behind the Export button, then Download. The resolution list runs 360p, 480p, 720p, 1080p, 2k and 4k, and the last two are below the fold in the dropdown — worth scrolling for, since the default is 720p. Selecting 4k and exporting gave a file that measured 3840 by 2160 at 30 fps, and the first frame of it is the card above. Two things are worth knowing before you plan around this. The format list offers MP4 and MOV and no still-image option, so a poster comes out as a one-frame grab from a video file. And the export carries a small Ai mark in the top-left corner, about 40 by 50 pixels in a 3840-wide frame, which is part of the deliverable whether or not your layout expected it.
What happens if you animate the finished poster
A save-the-date that moves is the obvious next ask, so the cleaned poster went to image-to-video with a deliberately conservative brief: hold the framing, keep the double exposure, let the water ripple and the haze drift, very slow push in.
The poster went in. A photograph came out, within about a quarter of a second.
The subject survived and the mood survived. The poster did not. The thin white rule, the paper grain, the double exposure and the pale panel are all gone inside the first second, and what plays is a good clip of two people at a lake. Measured in one fixed screen box, the pale panel reads mean 218.1 with a standard deviation of 10.4 at the first frame — a flat plate — and mean 145.8 with a deviation of 21.9 a quarter of a second later, which is open water.
That is a reasonable thing for the model to do. It was given a picture and asked for motion, and a picture of a poster is, to it, a picture of a couple with some marks on it. The graphic layer is not protected by being graphic.
So the working order inverts. Animate the photograph, not the poster, and rebuild the poster on top of the clip in the editor:
The clip on the timeline with the two text layers over it, each on its own background plate. 1920 by 1080, 30 frames per second, 149 frames, 4.97 seconds. The type holds still while the picture moves, which is what a title card is for.
The plate under each line is a text background, set from the same Style section as the fill: <em>Background</em>, then a cream from the recommended swatches. It is standing in for the panel the video destroyed, and it does the job better, because it travels with the text block rather than sitting at a fixed place in a moving picture.
One thing the wide shot cannot tell you
A practical note that came out of trying to measure the first set. In the wide environmental frame the couple occupies a small part of the picture — in the frame above, the gown is a few dozen pixels across. At that size there is nothing to check: the dress cannot be compared with the dress in the close-up, because there is not enough of it to compare.
That matters when you are reviewing a set before sending it out. The wide shot will pass any inspection you give it, because it is too small to fail one. Whatever continuity you need has to be established between the frames where the couple is large, and the wide shot has to be accepted on the strength of the rest or re-rendered on suspicion.
What it cost
Thirty-six credits, and the shape of that number is the useful part. Eleven stills at a credit each answered the questions about what a set holds and what a prompt can and cannot pin down. The one video cost more than the eleven stills together and returned a single finding, which is that the video does not keep a layout. The figures are what the interface showed for these configurations on 16 September 2026; the panel states the cost before you send, and the balance moved by 25 across the video, which matches.
Decide which lines have to be correct, and keep the model away from them
The through-line in all of this is a split between two kinds of content that look similar on a page and behave nothing alike.
One kind is allowed to be interpreted. A gown, a terrace, a golden-hour sky, a mood: describe it, take what comes back, re-render if you dislike it. The model is good at this and the variation is mostly free.
The other kind has one correct form. Two names spelled the way the people spell them. A date that is the date. A venue. A time. That kind of content is not worth handing to the model, and the evidence is in this article twice: it printed SAVE-OFE DATE unprompted, and it dropped the panel those words were meant to sit on as soon as the picture moved.
So write the exact lines down first, in a text file, before any generation. Everything on that list is editor work, set last, over the top. Everything else can be a prompt. Sorting that before the first credit is spent tells you where the picture ends and the document begins, which is the part of this job that is not guesswork.
Written 16 September 2026. Every still and clip here was made in CapCut for this article; no photograph of a real person was used, and the garments are generic descriptions not modelled on any designer's work. The eleven stills are 2560 by 1440 PNGs from Seedream 4.5 at 16:9 and 2K. The generated clip is 1280 by 720 at 24 fps from Seedance 2.0 Mini. The card at the top of the typesetting section is the first frame of a 3840 by 2160 editor export; the finished announcement is a second editor export at 1920 by 1080 and 30 fps, trimmed by two frames so the type track and the picture end together. The region figures in the scope test are mean CIELAB difference and the share of pixels past a difference of 10, computed between the two full-size frames; all four region boxes were drawn on both frames and checked by eye before the numbers were taken. The pale-panel figures in the animation test are mean and standard deviation of pixel lightness in one fixed screen box, with the box likewise checked on the frames first. Set consistency is described from the frames rather than from a measurement: four quantitative approaches were built and discarded first, all defeated by the same thing — that the three framings are deliberately different, so any whole-frame or largest-region statistic moves with the framing rather than with style drift. Locating the gown as the brightest large low-chroma region caught the sunlit lake instead, giving mask areas from 42,900 to 261,105 pixels across shots that should have shown the gown getting smaller. Whole-frame palette spread was likewise dominated by how much sky and water each framing included. Hand-placed color patches on the bride's hair were checked against the frames before use and two of the four had landed on the lake and on skin, so those figures were dropped as well. A further 16 credits were spent on an approach that produced nothing usable, and it is not described here because its failure could not be separated from the way it was driven.