When a picture made from a poor source photo comes out looking wrong, almost everyone checks the same place. They zoom into the face, look for mush around the eyes, and conclude the photo was not sharp enough. That is the wrong place to look, and this article has the measurements to say so.
One portrait was generated as a clean starting photo, then degraded four ways — full size, then 1280, 640 and 320 pixels wide with progressively harder compression, ending at a 23 KB file about the size of a screenshot pulled off a compressed video. Each version was handed to the same image-to-image prompt through the AI image generator, at 16:9 and 2K, 1 credit a frame. Then four fixed regions of each result were measured.
Every result exports at the same size, which is the first trap
All four came back 2560 by 1440. The source being a 2 MB photograph or a 23 KB thumbnail changed nothing about the pixel dimensions of the file. Anyone checking their export settings to work out whether the source was good enough is reading a number that cannot answer the question.
The PNG file sizes do fall steadily, which is the first hint that something inside the frame is thinning out even though the frame stays the same size. A PNG of a flat, smooth picture compresses smaller than a PNG of a detailed one.
Top row: the four results. Bottom row: the source each one was given, shown at the same display size so the degradation is visible.
The person survives, the room does not
Measuring the whole frame shows detail falling as the source degrades, which would seem to settle it. Measuring four fixed regions separately shows something more specific and more useful.
The person's row barely moves. The surfaces around the person lose between half and two thirds of their detail.
The subject is rendered about as well from a 23 KB source as from a 2 MB one — 1.06× between the best and worst, with the middle two measurements actually higher than the best-source measurement, so this row does not show a consistent trend across the four conditions. The best-to-worst ratios range from 1.88× for the floor to 3.01× for the upper-right ceiling region.
This makes sense once stated. The model is given the person to work from and a text description for everything else. A degraded photo still carries enough to say who is in the frame, what they are wearing and how they are lit — that information is coarse and survives compression. What a degraded photo stops carrying is fine surface texture, and with no texture to take its cue from, the model renders the rest of the scene smooth.
Which is why the result reads as wrong without any obviously broken part. Plaster goes plastic, floorboards lose grain, a wall becomes a flat gradient. Nothing is mangled, so zooming into the face finds nothing, and the picture still looks synthetic — because the emptiness is spread across every surface that is not the subject.
What to do about it
Three things follow, and none of them is "find a better photo of yourself", which is usually the one thing not available.
What the test cost
Attaching a reference image costs the same as generating without one: the panel reads Generating 1 item will consume 1 credit either way. The whole comparison is therefore a five-credit test, and it is worth running on your own material rather than trusting these numbers, because the answer depends on how much texture your scene was going to contain. These figures are what the interface showed for this configuration on 16 September 2026; the panel states the cost before you send.
Choose a scene with nothing to lose
The instinct when handed a bad source photo is to try to restore it — upscale it, sharpen it, regenerate until something sticks. The measurements say that effort is aimed at the part of the picture that was not in danger. The person comes through.
The better move is to change what you ask for. A poor source and a richly textured scene are a bad pairing, because the scene is exactly what goes flat. A poor source and a deliberately plain scene are a fine pairing, because a smooth wall rendered smooth is correct rather than degraded. Decide which one you are making before you write the prompt, and a 23 KB screengrab stops being a compromise and becomes simply the input you had.
Written 16 September 2026. Every image was generated in CapCut for this article; no photograph of a real person was used. The starting portrait is a 2560 by 1440 PNG from Seedream 4.5 at 16:9 and 2K. The four sources were produced from it by resampling to 2560, 1280, 640 and 320 pixels wide and re-encoding as JPEG at quality 82, 75 and 62 respectively, then restoring a PNG container so the container format was constant across the test. Detail is the mean gradient magnitude of CIELAB lightness within a fixed pixel rectangle, computed on the native 2560 by 1440 result in each case; the four rectangles were placed once and reused unchanged across all four results. The whole-frame figures quoted earlier use the same gradient measure plus a tile-based sharp-area fraction on a 1280 by 720 copy. Each condition was run once, so the regional ratios describe these four results rather than an average over repeats; the subject row is reported as flat precisely because its variation has no order to it, while the three surface rows fall in source order. An earlier run of the same four conditions was discarded in full: it recorded no binding between each result and the source that produced it, and tab ordering turned out not to match generation order, so the outputs could not be attributed. The generation script now writes the canvas identifier at the moment of sending, and every downloaded file is compared pixel by pixel against the starting portrait before use, because one recovered file in the discarded run was a byte-identical copy of it.