People make these grids for two reasons. One is a decision: whether to grow the beard, and which one. The other is a post: nine panels of the same face through a set of looks, which works because the viewer's eye is doing a comparison rather than looking at a picture. Both fail the same way. Somewhere around the sixth panel the man in it stops being the man in the first one, and by then the grid is showing a set of strangers with different haircuts.
This article builds one nine-panel grid at 16:9 in CapCut, five grooming panels and three era panels from a single generated portrait, and then measures the drift: how much of the frame moved, and how far each panel's face got from the original.
The finished grid. The top six panels change the lower face only; the bottom three change the clothing, the hair and the color of the photograph as well.
Fix everything except one thing
The source is a plain frontal portrait generated through the CapCut AI image generator flow in Video Studio's Image mode at Seedream 4.5, 16:9 and 2K: a clean-shaven man against a flat gray wall, even frontal light, a plain charcoal t-shirt. Everything about it is deliberately uninteresting, because everything in it has to survive eight more generations.
Each panel is that photograph plus one change sentence, and then the same block of constants, which is the part that does the work.
Keep everything else in the photograph exactly as it is: the same face, the same eyes, eyebrows and nose, the same short dark hair, the same head position and size in the frame, the same charcoal t-shirt, the same gray wall and the same flat frontal light. Change nothing above the upper lip.
Measured on the files, the constants held where it matters most. Across all nine panels the detected face box is between 198 and 209 pixels wide against 202 in the source, its center sits within three pixels horizontally and four pixels vertically of where it started, so nothing needs aligning after the fact and the panels can be laid straight into a grid. That is the payoff of naming the head position and size rather than assuming they will stay.
Below the upper lip
The five grooming panels are stubble, a moustache, a goatee, a short beard and a full beard, each one sentence. To ask whether the face survived, the top half of it can be compared directly: the region from the brow to just under the eyes is the part no beard touches, so if it changes, the change was not the one that was asked for.
The part of the face no beard reaches, cropped from each panel at the same size. The number is how closely it matches the source.
The short beard and the stubble panels score 0.983 and 0.973, which is the same face with something added under it. The moustache and full beard sit at 0.933 and 0.925, and the goatee at 0.904 is the loosest of the five, which is visible in the crop as a slightly heavier brow and a softer eye. Nothing was asked for above the upper lip in any of them, so those differences are drift rather than instruction.
Changing the year changes more than the year
The three era panels ask for a period look: an 1890s high-collared jacket in sepia, a 1950s shirt and tie in black and white, an 1980s patterned shirt in faded color. Each one changes three things at once, the clothing, the hair and the color of the whole photograph, and each one also changes the face more than any grooming panel did.
The eight panels ranked by how close the untouched part of the face stayed. The grooming panels average 0.944 and the era panels 0.901.
The 1980s panel at 0.881 and the 1950s at 0.889 are the two loosest in the grid, and the reason is in the request rather than in the model: a period look is a grading instruction as well as a wardrobe one, and once the whole frame is being re-rendered in sepia or in black and white, the face is inside the part being re-rendered. The 1890s panel does better at 0.933 because its sepia is warm and gentle, closer to the original's own tones.
The practical version of that is to treat era panels as a separate row rather than as more of the same. They cost more identity per panel, so a grid that mixes them should keep them together, where a viewer reads the color shift as the point rather than as an inconsistency.
What the eleven requests cost
The cost line before one of the panels, read on 10 September 2026.
Eleven requests at one credit each, nine of which are in the grid. Two panels did not come back on the first attempt and were sent again, which is worth budgeting for when a grid needs a specific number of panels to be complete. These figures are for this configuration on this date and come from the composer's own cost indicator; other models, resolutions and modes are charged differently.
What the grid is evidence for
If the grid was made to decide something, it is worth being precise about what it can settle. It is good evidence about shape: whether a full beard suits the jaw, whether a moustache reads as too heavy on that upper lip, whether the face is better framed by hair or by a clean neckline. Those are geometry questions and the panels answer them, because the geometry is what held steady across the set.
It is weak evidence about texture and color. The density of a real beard, how gray comes in, how patchy the cheeks are at three weeks, none of that is in the picture, and the panels here differ slightly in skin rendering even where nothing was asked for. So use the grid to narrow five options to two, and expect the real thing to differ in the details the grid cannot know about.
The portrait here was generated for this article. With a real photograph, this kind of grid is a set of pictures of a specific person in looks they did not choose, so keep it to your own face or to someone who has agreed to it. The era panels have one more line: they are styling exercises, not history, and they should not be presented as old photographs of anyone, put into a family album, or made to resemble any real person, living or dead. What this article does not cover: masking or repainting a panel by hand, background removal to unify the panels, the animated version that cuts between them, saving the grid as a reusable template, and 4K export.
Written 10 September 2026. The portrait and all eight panels are generated images made in CapCut for this article; the man in them is invented and no real person, product or brand appears. Interface labels and the cost line reflect a single session on that date and the configuration named above, and may change. Face-box sizes and positions were found with a standard frontal-face detector, and the match figures are correlations between the upper half of each face and the same region of the source, taken at the same size.
内部审校参考（发布前删除此节）
锚定玩法：男子AI胡子进化史（ID 81，M11 时尚与写真，P0，互动热度 171）；其他锚定玩法 AI穿越百年写真(h3) 已一并做出（三格年代版）。标题与 slug 与规划行一致。
规划行"内容结构建议"中的局部重绘蒙版、背景移除统一底色、字幕模板加标签、逐格切换短视频与卡点、存为可复用模板、画质增强与 4K 导出本次未做，已在文末范围说明列出；
一致性校验以"上半张脸相关度 + 人脸框尺寸位置"两项实测替代。
风险与备注"生成对象仅限本人或已获同意 + 跨年代不得对照任何真实人物、不得用于伪造历史或家族影像"：末节已逐条写入。
核对口径（2026-09-10 生产前核验 TikTok）：
"ai photo through the decades trend" → "This trend got me romanticizing a life I never lived"（2257）、"#capcut #capcutpioneer 如何做怀旧 90 年代照片"（9285）、"how to create the 80s photo trend"（1175，1 天前）、"I asked ChatGPT what would I have looked like in the 80s"（97，1 天前）；
"ai beard styles on my photo" → "viral no beard filter … capcut tutorial clean shave effect"（765）、多条 AI 胡子滤镜教程（262 / 138 / 66）；
"before after grid ai portrait" → "how to make four grid realistic portrait photos"（96）等。年代路线更旺，胡型路线为稳定的工具型需求。
ai photo through the decades trend；
ai beard styles on my photo