Matcha portraits keep surfacing in social feeds: a single green palette, a background of slow marbled paint, and a person who still looks like a photograph in the middle of it. The usual way to fail at one is to watch your own face melt into the swirls. This walks through the version that produced the result below, built in CapCut Design Studio from one kitchen selfie and one prompt written as two lists.
The source selfie and the finished portrait. The kitchen dissolved, the face did not.
What makes the matcha look: one palette, moving paint, an intact face
The effect has three properties, and naming them is what makes the prompt writable.
One green palette covers everything. There are no competing colors anywhere in the frame. The range runs from deep forest green to pale cream, and even the person's skin picks up a soft cast of it.
The paint behaves like liquid. The background is not blurred or tinted. It moves in thick marbled swirls with visible brush texture, the way oil paint looks mid-stir.
The face stays photographic. This is the property that separates a portrait from a green puddle, and it is the one the model will sacrifice first unless the prompt defends it.
Every stronger version of the first two properties costs you some of the third. The whole craft of this effect is deciding where that trade stops.
Start from a chest-up photo
The protection you will write into the prompt needs something to protect. In a chest-up selfie the face occupies a large share of the frame, which gives the model enough pixels to keep the features sharp while it repaints everything around them. The smaller the face, the less there is to defend.
A busy background helps rather than hurts. This source was shot in a kitchen with shelves, jars, plants, and a window, and all of it became raw material for the swirls. The transformation reads clearly because there was something recognizable to transform.
The source: chest-up framing, face large in frame, detailed background. The photo was generated in Design Studio rather than photographed, so no real person's likeness is involved.
Write the prompt as a keep list and a transform list
A single instruction like "make this matcha fluid art" leaves the model to decide what survives, and the face is usually what it spends. Splitting the prompt into two lists removes that decision.
The keep list comes first. Name the person and every attribute that must survive: face, features, smile, pose, framing, realistic skin. Leading with protection matters, because it frames everything that follows as an exception to it.
The transform list names each conversion. The background dissolves completely. The outer edges of hair and shirt melt into the currents. The palette locks to one green range. The face receives only a cast of color, nothing more.
The prompt used for the portrait in this article:
Keep this exact woman clearly recognizable: same face, same features, same smile, same pose and framing, realistic skin on her face. Repaint everything else as matcha-toned fluid art: the kitchen background dissolves completely into slow swirls of thick matcha green oil paint with visible marbling and brush texture, and the outer edges of her hair and t-shirt melt into the same green paint currents. One single green palette across the whole image, from deep forest green to pale cream. Her face keeps only a soft green cast from the surrounding color. Soft even light, high detail.
The load-bearing phrase is the opening one. Keep this exact woman clearly recognizable sets the rule that the rest of the prompt operates under.
Run it in Design Studio
Open Design Studio from the left rail of the CapCut workspace, or start from the CapCut AI image generator page. The composer in the middle of the screen takes both the photo and the prompt, and the plus button on its left edge is where the photo goes in.
The plus button attaches your selfie. The small counter on the right shows the credit cost of a generation, 16 at the time of writing.
The work then happens on a canvas: your photo sits on the left, and a chat panel on the right holds the conversation with the model. The prompt goes into the chat box, and each generation returns a new image beside the original instead of overwriting it, so the source stays available for comparison.
Source and result side by side on the canvas. The toolbar above the image covers further edits, and the chat on the right offers one-click refinements.
Reading the result
The repaint for this article came back in under half a minute. Before treating a result as finished, check the three places where this effect either holds or breaks.
Eyes and teeth. These are the smallest details the keep list protects. Here they stayed sharp, and the smile is the same one from the source.
The hair boundary. A good result feathers the tips of the hair into the surrounding swirls while the mass of it stays solid. A bad one either leaves a hard cutout edge or lets the paint climb toward the scalp.
The background. Nothing from the original room should survive. In this result the shelves, jars, and window are gone without a trace of their geometry.
The finished portrait. The face carries only a green cast, the shirt keeps its folds, and the room is entirely paint.
If one aspect is off, the chat offers single-change refinements under the reply rather than requiring a rewrite. For this result the offered chips were Explore other color palettes, Adjust the fluid art style, and Refine green cast on face, and the last one is the lever to pull when the skin has taken too much green.
The same photo without the keep list
To show what the first half of the prompt is doing, here is the identical source run again with the protection removed. The instruction asked for heavy marbling across the whole frame, face included.
The same selfie with and without the keep list. On the right, the paint took the face; the gold earrings are nearly the only detail that survived.
The full-melt is not a failure of the effect, and in the video form of this trend it is the dominant look: the most popular clips run the marbling across the entire frame, person included, until the footage reads as flowing paint. What it stops being is a portrait. The marbling erased the smile, redrew the features as paint ridges, and left the earrings as the last recognizable objects in the frame, which is the right outcome for an abstract clip and the wrong one for a profile picture.
Between these two poles you can set the strength anywhere, and the control is list membership rather than a slider. Moving hair and shirt fully onto the keep list gives the subtlest version, where only the background liquefies. Letting their edges melt, as the featured prompt does, is the middle setting. Moving the face onto the transform list is the far end.
What this effect will not do
The palette is total, skin included. If you need true skin tones with a green background, this single-palette effect is the wrong tool, because the cast on the face is part of the look.
The protection depends on face size. On a full-body shot where the face is a small fraction of the frame, there is little for the keep list to hold onto, and results drift toward the melted end regardless of wording.
This article covers the still-image version. The same palette also appears in a video form of the trend, where footage is graded toward matcha green, and that is a different workflow from repainting a photograph.
Written 27 August 2026. The source selfie is AI-generated in CapCut Design Studio and does not depict a real person. Interface labels, timings, and credit figures reflect a single session on that date and may change.