There is no strength slider in a written request. When a stylized look comes back too strong, the reflex is to ask for less of it, and "about half" is not a quantity the request can hand back. So the useful question is not how to turn the effect down but what the words actually buy, which is something that can be checked.
This article runs one portrait four times at 16:9 in CapCut, at a fifth of full strength, at half, at full, and once with the strength unmentioned and the region named instead, then measures two things on each file: how much structure the effect put behind her, and how far her skin color moved.
What "one fifth of full strength" bought
The starting photograph is a plain one, generated through the CapCut AI image generator flow in Video Studio's Image mode at Seedream 4.5, 16:9 and 2K: a woman in a white t-shirt against a smooth warm gray wall, even light, natural skin. The wall is deliberately empty, because the wall is where a stylized look has room to happen.
The same photograph and the same look at three requested strengths. The changes are not evenly spaced, and they are not the same kind of change.
The light version tinted the wall and left everything else alone, and on the file that is a color change with no structure in it. The spread of brightness across the background, which is 6.3 in the original photograph, went to 9.4, and a standard edge detector finds no edges at all in the background of either. Nearly the whole sampled background is inside the green hue band, so the wall is fully green and it is still a wall.
That is worth knowing before asking for a fifth of anything: the light request did not produce a fifth of the effect, it produced one of the effect's two ingredients. The color arrived and the paint did not.
Where the effect goes once you let it in
At half strength the background stops being a wall. Its brightness spread goes from 9.4 to 40.1 and 6.06 percent of it now registers as edges, which is the marbling: the effect has become a surface with structure rather than a wash.
Two numbers per version. The blue bar is how much structure the effect added behind her; the orange bar is how far her skin moved toward green.
The half-strength request also spread onto her, and the order it spread in is the useful part. Her t-shirt picked up a green cast and a few streaks, a green rim appeared along her shoulder and hair, and her face did not move: her skin sits 4.8 degrees of hue from where it started, against 4.3 for the light version. The clause that held it there was written as a place rather than as an amount.
Her face and neck keep their own natural skin colour with no green in them at all, and her features stay sharp and unchanged.
Full strength puts paint on the face
Asked for the whole look with nothing held back, the effect takes the skin. Her hue moves 14.7 degrees, roughly three times the shift in any other version, and at full size the paint is running down her cheek and across her jaw.
The two strongest versions and their faces at full size. Both have a fully painted background; only one of them still has her in it.
This is the version most people mean by the effect being too strong, and describing it as too strong is what makes it hard to fix. What actually happened is that the paint was given no boundary, so it went everywhere, including the one surface a portrait cannot spare.
Naming the region beats naming the strength
The fourth request leaves strength out of it. It asks for the background to be replaced at full strength, with a clean edge along her hair and shoulders, and for her to be left as she is.
Replace only the background of this photograph with full-strength matcha fluid art: thick swirling matcha-green and cream paint with visible marbling, filling the whole area behind her right up to a clean edge along her hair, shoulders and arms. The woman herself is left exactly as she is in the original photograph, with her skin, her hair and her white t-shirt completely unchanged and no green cast anywhere on her.
That version has the most background detail of the four, 7.42 percent of the sampled area registering as edges against 7.07 for full strength, and it moved her skin 4.5 degrees, which is where the light version left it. In other words the request that named a region beat the request that named a strength on both counts at once. Its one leak is at the bottom of the frame, where her forearms picked up a faint green cast that the request had excluded.
What the five requests cost
The cost line before one of the conversions, read on 10 September 2026.
Five credits for the whole ladder, which is the practical argument for running one: at this price the way to find the dose is to generate the light and the strong version and look at both, rather than to reason about a number in the request. 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, and video is priced separately.
What goes first as you turn it up
Across the four versions the effect arrived in a consistent order, and knowing the order is what lets you stop at the right place. The wall goes first, and it goes as color before it goes as texture. The clothing goes next, and a white t-shirt is the most willing surface in any portrait, which is why it picks up a cast a full step before anything else does. Then comes the rim: a line of the effect's color along the shoulder and the hair, which is the last change that still reads as lighting. After that it is the skin, and the skin is where the picture stops being of a person and starts being of a paint effect.
So the place to stop is the rim, and the way to stop there is to name the face rather than to name a fraction. If a stronger background is wanted after that, ask for the background by name at full strength instead of turning the whole thing up, since that is the version that gave the most paint and the least damage here.
One boundary belongs with any effect that recolors a person. The portrait in this article is generated, and with a real photograph the version that puts paint on somebody's face is a picture of them that they did not sit for; use your own, or ask. What this article does not cover: masking or repainting by hand, background removal as a separate step, applying the same treatment to video, upscaling and 4K export.
Written 10 September 2026. The portrait and all four conversions are generated images made in CapCut for this article; 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. Background contrast, edge share and skin hue were computed on the downloaded files, with the face located in each file by a standard frontal-face detector.
内部审校参考（发布前删除此节）
锚定玩法：Matcha Filter（ID 37，M4 滤镜与色调，P0，互动热度 204）；其他锚定玩法 抹茶滤镜效果(h27)、Matcha color(h105) 同源。标题与 slug 与规划行一致。
与第 02 篇的区分：02 篇（matcha-fluid-art，v1 期 9:16）讲的是"保留清单 vs 转换清单"的二选一写法；本篇讲的是**浓度分档与作用范围**，
四版对照 + 分区实测（背景对比度/边缘占比、肤色色相偏移），结论是"点名区域优于点名强度"，与 02 篇不重合。
规划行"CapCut 功能落地"中的编辑器调色、局部重绘、背景移除、画质增强、4K 导出本次未做，已在文末范围说明列出。
核对口径（2026-09-10 生产前核验 TikTok）：
"matcha filter photo trend" → "super-viral Matcha Filter effect sweeping across TikTok"（830，3 天前）、"matcha trend filter #Matchatrend"（9014）、"My kind of Matcha Trend"（1342）；
"ai filter only background not face" → "Background filter effect tutorial" 同一创作者两条（9334 / 2573），正是"只让背景吃满效果"的做法。两条主线均在更新。
matcha filter photo trend；
ai filter only background not face