Two images come out of the same camera on the same day. One is a ceramic mug on a seamless backdrop. The other is a portrait lit with a softbox. Both are 1800×1200, both get upscaled 2× with the same preset, and both come back slightly wrong — in opposite directions. The mug looks like plastic. The face looks like a mannequin. Turning the strength slider up fixes neither, because the two failures are not the same defect at different intensities.
Product work and portrait work should not share a pipeline, and the reason is not "one needs more sharpening." The two source classes have different statistical structure, and a learned upscaler is a machine that exploits statistical structure.
The two sources are statistically opposite
A material surface — brushed aluminium, woven cotton, stone, cardboard — is stationary in the texture sense. The statistics of a 32×32 patch of the left side of the fabric are close to those of a 32×32 patch on the right. A convolutional network exploits exactly this: it estimates the local texture distribution from anywhere in the image and applies it everywhere, because there is a large consistent sample to learn from within one input.
Skin is not stationary. Pore density and pore size vary systematically across the face — finer and denser on the nose and medial cheek, coarser toward the jaw and temple — over a low-frequency colour field (subsurface scattering, capillary variation) with no counterpart in the pore layer. The upscaler has no concept of "this region is skin"; it has only learned which local patterns co-occur, and skin and ceramic both present as smooth low-variance regions with fine high-frequency structure on top. So a model treating skin as stationary applies nose-scale detail to the forehead and forehead-scale detail to the nose. At 2× that reads as unevenness; at 4× it reads as a texture map pasted onto a face.
Skin: the training prior is aimed at the wrong target
Here is the uncomfortable part. The portraits that dominate public upscaling datasets are, overwhelmingly, retouched. Commercial portraits, stock photography, magazine covers — all have been through frequency separation and skin smoothing before publication. A generative model trained on that corpus learns that plausible high-resolution skin is smooth skin with a few preserved pores. Hand it a raw portrait and its maximum-likelihood answer is the retouched look. It is not failing; it is succeeding at the wrong objective.
The second half is perceptual and not fixable with better models. Human cortex has dedicated machinery for faces, and face detection thresholds are far lower than for comparable contrast on non-face content. An artifact a viewer would never notice on a wall is obvious on a cheek, so a single global preset strength cannot be correct for both.
Material: regular structure the model can lock onto, and the ways that goes wrong
Periodic texture aliasing. A woven fabric at 1800 px wide still has a weave period of a few pixels. If that period sits near the model's effective sampling limit, the reconstructed weave can drift in phase across the image — threads bend, or a false moiré appears. The stricter the regularity, the more visible the phase error, because the eye is excellent at continuing a periodic pattern and spotting where it breaks.
Hallucinated micro-texture. Wood grain, leather pores, concrete aggregate: the model has a strong prior for these and inserts grain where the source had none, especially in smooth or defocused areas. If the detail is selling a finish, invented finish is a defect.
Edge bleed at the cutout. Studio product shots are usually cut against white. That boundary is a step edge at very high contrast, and step edges are where reconstruction ringing lives. A halo of a few pixels around a silhouette is the common giveaway that a catalogue image was upscaled.
Specular highlights contain no information, and that is where models invent
A clipped specular highlight — the blown-out stripe on a glass bottle, the hot spot on polished metal — has pixel values pinned at 255 across its core. No gradient, no texture, no recoverable structure. Whatever a model puts there is invented from the prior.
For a portrait this is mostly harmless: skin speculars are broad and soft because of subsurface scattering, rarely clip hard, and sit in a low-frequency region anyway. For a product it is the difference between glass and plastic. The reference for a blown highlight is the shape of the light source and the curvature of the object — information that lives in the highlight's boundary gradient, not its core. An upscaler that paints texture into the core destroys the read of the material. The fix is not a stronger preset; it is protecting the highlight and letting it stay smooth.
Two different colour tolerances
On a product, colour is often contractual. A brand's red has a specified value, and a customer comparing screen to object notices a two-point hue shift. Generative upscalers weight luminance far more heavily than chrominance — the same bias as every practical codec — so chroma is reconstructed more loosely and saturated brand colours drift.
On skin the same looseness is worse. The visual system calibrates skin tone against memory with remarkable precision, so a face that shifts slightly orange reads as a lighting error, and because skin chroma lives in the low-frequency subsurface layer the shift is spatially broad: not noise, but the wrong white balance. If your pipeline has a separate chroma pass, this is where to spend it.
What the presets should actually differ in
The distinction that matters is not sharpness strength but how much the model may synthesise versus interpolate. For material, respect structure — reconstruct periodic detail, stop at the edge of the highlight. For skin, the opposite bias: interpolate more, synthesise less, accept slight softness, because the alternative is a face silently retouched that every viewer reads as uncanny without being able to say why.
Decision table
| Source class | Dominant risk | Preset bias | What to protect | Avoid |
|---|---|---|---|---|
| Product on seamless backdrop | Halo at cutout edge | Structure-respecting, light synthesis | Silhouette edge, specular core | High sharpening near background |
| Textile, leather, wood | Hallucinated micro-texture, weave phase drift | Moderate, with grain control | Periodicity of the weave | Presets that add film grain |
| Glass, polished metal, gloss | Invented texture inside blown highlights | Low synthesis | Highlight boundary gradient | Anything that fills clipped areas |
| Portrait, beauty, headshot | Over-smoothing, silent retouch | Interpolation-biased, low synthesis | Pore variation across zones | Default "enhance" modes |
| Portrait in a wide scene | Wrong: skin is not the subject | Treat as scene, not portrait | Overall consistency | Per-face tuning at low pixel area |
| Print or packaging artwork | Text and logo hallucination | Near-lossless, minimal synthesis | Letterform geometry | Generative detail models |
| Cosmetics applied to skin | Both risks at once | Half strength, verify by zone | Product colour, skin chroma | Single global strength |
When you should not split the presets this way
When the product is skin. A skincare campaign, a dermatology reference, a tattoo portfolio — here the "material" is skin, and product-accurate colour lands on top of face-detection sensitivity. Neither preset is right; you need the conservative one plus a colour check against a physical reference.
When the texture is the product. A fabric swatch catalogue, a stone sample set, a laminate selector. Any invented weave misrepresents goods; the honest answer is a capture-side fix — more resolution, better light.
When the face is small. A portrait preset applied to a face occupying 40 pixels is wasted effort and usually a net loss. Tune for the frame, not the subjects you know are in it.
When the source is already retouched. A synthesis-heavy preset on commercial retouched photography amplifies the retouching, not the detail — you are sharpening someone else's frequency separation. And if review happens as a chat thumbnail, the preset choice is not being evaluated at all; both failure classes are invisible below 100%.
A two-minute check that separates the two cases
Crop a 200×200 region of the most important texture — a patch of weave, or the cheek beside the nose. Upscale the crop alone at 100% and compare against the same crop upscaled with plain interpolation. Then ask: did the model add anything not traceable to a gradient in the source?
On material, a small amount of added structure is acceptable and usually correct. On skin, any added structure is a hallucination, and the correct preset is the one that adds least. That test — traceability of every high-frequency detail to the source — settles most preset arguments faster than any slider comparison, and it is the criterion I use when choosing between the conservative and aggressive ends of a preset set, as documented at upgolab.com.
