How to Upscale Photos with AI Without Making Them Look Plastic

AI upscaling can turn a 600-pixel-wide grainy photo into something usable for print. It can also give your subject the now-famous "AI sheen" — skin like polished porcelain, eyes that look painted, a faint plastic glow on every surface. The difference between a clean upscale and a plastic one is mostly about which model you use and how hard you push it.

What AI upscaling actually does

Traditional upscaling (bicubic, Lanczos) just interpolates between existing pixels. Result: a bigger but blurrier image. AI upscaling, often called super-resolution, uses a neural network trained on millions of high-low resolution pairs. Instead of guessing pixel values mathematically, it invents plausible detail based on patterns it learned in training.

That last bit is the key. AI upscaling does not recover detail that was never in the original. It invents detail that could plausibly belong there. Sometimes that lands. Sometimes the model invents detail that looks fake.

When AI upscaling works well

Mildly low-resolution photos. A 1024-wide image upscaled 2x usually looks great. The model has enough information to work with.
Photos with clear subjects. Faces, animals, objects with recognizable shapes. Models have lots of training data here.
Sharp originals, just small. If the source is in focus and properly exposed, upscaling is mostly safe.

When it produces the plastic look

Aggressive 4x or 8x upscales on tiny inputs. The model has to invent too much. Skin gets airbrushed, pores disappear, hair looks like brush strokes instead of strands.
Photos with motion blur or noise. The model often misreads noise as texture and tries to "clean it up," producing the smoothed-over plastic effect.
JPEG-compressed inputs. Compression artifacts confuse the model. It sees the blocky 8x8 patterns and either preserves them as fake texture or smooths them into goo.
The wrong model for the content. A model trained on photos applied to anime, or vice versa, produces uncanny outputs. Many apps use a single general-purpose upscaler that is never the best fit.

How to use AI upscaling without the plastic look

Start with the best original you have. Original camera file beats screenshot beats re-saved JPEG. Each generation of compression compounds.
Stay at 2x when possible. 2x almost always looks natural. 4x is a coin flip on difficult photos. Anything beyond is asking for trouble.
Match the model to the content. If the app offers a "face" or "portrait" mode, use it on people. A photo-trained model will do less weird stuff to a face than a generic one.
Watch for the over-sharpened look. If the upscaled image looks sharper than the original could plausibly be, the model is hallucinating. Back off the strength or pick a less aggressive option.
Compare at 100% zoom. An upscale that looks great fit-to-screen often looks fake at full resolution. Always inspect at 100%.

Upscale vs. enhance: not the same thing

Plenty of apps blur the line. Upscale increases resolution. Enhance adjusts color, contrast, and sharpness. Restore repairs damage. Most modern apps bundle them. That is fine, except a single button hides three different operations, each with its own failure modes. If your photo looks weird after, try unbundling: upscale only, no enhance, no restore. See what each layer is doing.

What upscaling cannot do

Recover detail that was never there. A 200x200 thumbnail of a stranger's face cannot be upscaled into a recognizable portrait. The model will invent a face, but it will not be the right one.
Fix motion blur. Some motion-deblur models exist, but they are separate from upscalers. A blurry photo upscaled is still blurry, just bigger.
Remove watermarks or text. An upscaler will faithfully reproduce them at higher resolution.

If you are restoring an old family photo, upscaling is one step in a longer chain — see our practical guide to restoring old family photos with AI. If you are upscaling because the source is a phone screenshot or a small social export, also check the scan-and-enhance workflow for old prints — the rules carry over. Piko includes a tuned upscaler and a separate enhance pass so you can apply each one deliberately.