Pixel, voxel and clay are three different claims about what a face is made of, and a request that does not make the claim specific comes back as the photograph with a filter over it. This article converts one portrait into all three at 16:9 in CapCut, lays them out as a before-and-after strip, and settles two arguments with numbers taken off the downloaded files: whether the style actually arrived, and whether the person survived it.
Three surfaces, and one measurement that covers them
The three looks differ in one physical property. Pixel art is a lattice of flat squares, so it is defined by having large areas of exactly one color. Voxel is the same idea in three dimensions, solid cubes with lit faces, so it is flat in patches and it also has depth. Clay is the opposite of both: a modelled surface that carries thumb marks, tool grooves and slight unevenness, which means it should be less flat than the photograph it came from.
That gives one measurement covering all three. On the downloaded file, count the share of pixels whose nine-by-nine neighborhood varies by six levels or less, which is the share of the picture sitting inside a patch of one flat color. Pixel and voxel should push it up, clay should push it down, and a request that failed leaves it where the photograph had it.
One generated portrait and the three conversions made from it. Every conversion started from the picture in the top left, not from another conversion.
The source is a generated photograph rather than a real person's picture, made through the CapCut AI image generator flow in Video Studio's Image mode at Seedream 4.5, 16:9 and 2K. It has three features worth watching through the conversions: thick eyebrows, a specific hair length and part, and a small dark mole on her cheek below the eye.
The pixel request that did not pixelate anything
The first request described the grid in the terms most tutorials use, a block count and a palette size, and asked for the pixel-stretch glitch that the look is known for.
Rebuild this portrait as pixel art on a 16:9 canvas. Quantize the whole picture to a visible square pixel grid about 120 blocks across the frame width, so every block is a hard-edged square with no smoothing, and reduce the palette to about 24 flat colors with visible dithering in the shading. Keep the same face: the same eyebrow shape and thickness, the same eye spacing, the same hair length and part, the same dark mole on her cheek below the eye, and the same olive-green sweater. Across a horizontal band about one fifth of the frame high at the level of her mouth, smear the blocks sideways into long horizontal streaks, a pixel-stretch glitch. Flat light gray background. No text, no logos, no brand marks.
What came back has the glitch band at the right height and blocks in places, and it is not pixel art. The flat-color share in the lower half of the face is 14.6 percent against the photograph's 29.3, and on the sweater it is 4.3 percent against 7.1. By the measurement, the conversion made the picture less flat than the photograph it started from.
The second request kept the same subject clauses and changed only the scope of the quantization, naming the parts that were left alone and forbidding the escape route.
Rebuild this portrait as strict pixel art on a 16:9 canvas. Every part of the picture, including the eyes, the eyelashes and the mouth, is quantized to the same square grid of about 100 blocks across the frame width; each block is one flat color with a hard edge, and there are no smooth gradients, no soft edges and no photographic detail left anywhere in the image. Reduce the whole image to 24 flat colors. Keep the face readable at that resolution: the same eyebrow shape, the same eye spacing, the same hair length and part, one darker block for the mole on her cheek, and the same olive-green sweater. Flat light gray background. No text, no logos, no brand marks.
The same clauses about the face, two different scopes for the quantization. The lower row is the same part of the sweater from each file at full size.
The second version measures 55.9 percent flat in the lower face and 57.0 percent on the sweater. Naming the eyes and the eyelashes as parts that also have to be quantized is what moved it, and even so they came back smoother than everything around them: the eyes still hold soft gradients while the cheek beside them is in squares. In this run the request could raise the resolution of the effect everywhere else without ever fully reaching the eyes.
Voxel brings the lattice with it
The voxel request needed less insistence. Asking for a figure built from equal-sized solid cubes, roughly thirty across the width of the head, each cube one flat color with visible edges and simple shading, produced 43.6 percent flat in the lower face on the first attempt, with the mole rendered as a single darker cube and the hair stepped in blocks.
Flat-color share in two regions of each file. The sweater is the more honest of the two, because a face has soft skin in it either way.
The gap between voxel's face and its sweater, 43.6 against 18.8 percent, comes from what the two are: the head is built from cubes and the knitted sweater keeps a woven texture on the cube faces. That is worth knowing before you write the request, because a cube style that stops at the collar reads as a head sitting on a photograph.
Clay is the style that gets less flat
Clay went the other way, as the surface argument predicts. The named clay conversion measures 12.8 percent flat in the lower face and 2.2 percent on the sweater, both below the photograph, because the request asked for thumb marks, tool grooves and slightly uneven edges and got them. The wording that did it is a list of surface facts rather than a style label.
Remodel this portrait as a hand-made clay figure on a 16:9 canvas. The head and shoulders are shaped from matte modelling clay with soft rounded volumes, visible thumb marks and tool grooves in the surface, slightly uneven edges, and a completely matte finish with no gloss. Keep the same face: the same eyebrow shape and thickness, the same eye spacing, the same hair length and part shaped as clay curls, the same dark mole on her cheek below the eye, and the same olive-green sweater. Even soft studio light, flat light gray background. No text, no logos, no brand marks.
Read against the other two, clay is the style that cannot be checked by flatness alone, since a photograph would also score low. It is checked by what appears rather than by what flattens: tool marks across the cheek, a seam where the hair meets the forehead, a matte finish with no bright highlight anywhere on the skin.
What a one-line request did to her face
The same photograph was converted a fourth time with the request most people would actually type: "Turn this photo into a cute 3D clay style portrait." It is a clay figure, it is competent, and it is not her.
The same photograph, the same style, two requests. The mole survived both; the proportions did not.
The visible change is in the eyes, and it can be counted. Measuring the share of near-white, low-saturation pixels across the eye band of the detected face box, the photograph reads 0.28 percent, the named clay conversion 1.29 percent, and the one-line version 7.59 percent, which is twenty-seven times the photograph. The one-line request also raised the eyebrows, shortened the nose, added a smile and lengthened the neck. It kept the hair, the sweater and the mole, which is what makes the result so easy to accept at a glance and so wrong beside the original.
The clause that prevents it is the one carried through every request above: the same eyebrow shape and thickness, the same eye spacing, the same hair length and part, the same mole, the same sweater. Those five items are cheap to write and they are the difference between a portrait of someone and a doll wearing their haircut.
What the six requests cost
Every image here came from the Video Studio home composer, which prints a cost line before it starts.
The cost line and the settings it belongs to, read on 9 September 2026.
Six credits produced one photograph and five conversions, of which four are in the article and one, the first pixel attempt, is here as the version that did not work. 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.
Laying out the before-and-after frame
The comparison is the deliverable, and it has one rule: the crop has to be identical across the panels, or the reader is comparing framings rather than styles. Each of the four files here is 2560 by 1440 with the head in the same place, so a single center crop taken at the same width from all four lines the faces up without any adjustment.
The four panels at the same crop. Cover the labels and the styles are still told apart by their surfaces, which is the test the strip has to pass.
Two things to do before exporting a strip like this. Put the photograph first and leave it unlabeled if the styles are obvious, because the value of the frame is in the comparison rather than in the captions. And check the strip at the size it will be seen: at thumbnail width the voxel and pixel panels stay legible while the clay panel reads much like the photograph, so if only one style is going out, clay is the one that needs the original beside it.
The photograph in this article was generated for it, and that is the setting where a strip like this is uncomplicated. A real photo of a real person makes the same strip a picture of them, in three forms they did not choose; use your own face, or ask first. What this article does not cover: background removal, inpainting individual features, upscaling, printing, and animating any of the three styles.
Written 9 September 2026. The portrait and all conversions are generated images made in CapCut for this article; no real person, product or brand appears, and the styles here are described by their surfaces rather than by reference to any existing product or title. Interface labels and the cost line reflect a single session on that date and the configuration named above, and may change. Flat-color share and eye-white share were computed on the downloaded files; face regions were located with a standard frontal-face detector.