How many pixels does the face have?
Measure the face, chin to hairline, in the source. Above roughly 200 px, a faithful upscaler has real information to work with, and the result will look like the person. Below about 60 px, every method is reconstructing from very little, and restoration models are effectively drawing a face from a description. That is when identity drift appears.
A safe workflow
- Faithful upscale of the whole image to the target size.
- Face restoration as a separate layer, if the faces are small or blurry.
- Blend the restored faces at 30–60% opacity over the faithful result.
- Check eyes, teeth, ears and hairline at 100%, the areas models change most.
- Compare against another photo of the same person taken at a similar age.
What to look for
| Artefact | Looks like | Fix |
|---|---|---|
| Waxy skin | Plastic, poreless cheeks | Lower restoration strength; add back fine grain |
| Identity drift | Different eye shape, nose or jaw | Raise fidelity / resemblance; blend less |
| Mismatched eyes | Different iris colour or direction | Restore each face at a higher input resolution, or mask manually |
| “Beauty filter” effect | Younger, smoother, symmetrical | Use faithful only; skip creative passes on faces |
Creative upscalers and faces
Creative upscalers can produce stunning skin texture, but at medium or high creativity they also change expressions, ages and features. For real people, keep creativity at its lowest, or mask faces out of the creative pass entirely and take them from the faithful version.
Know the limits
An upscaled face isn’t a better photograph of that person. It’s a plausible reconstruction. That’s fine for a family album, and inappropriate for identifying someone, as evidence, or in news photography without disclosure.
FAQ
Why does the upscaled face not look like the person?
Face restoration and creative upscalers generate facial detail from learned priors. When the source face is only a few dozen pixels wide, the model is mostly guessing, and its guesses lean towards an average face. Lower the strength, blend the result, or accept a softer but truer face.
What are GFPGAN and CodeFormer?
Two widely used open-source face restoration models. GFPGAN (2021) uses a generative face prior from a StyleGAN-type model; CodeFormer (2022) uses a learned codebook and exposes a fidelity weight that trades quality against faithfulness. Many commercial “face enhance” features are built on similar ideas.
Is it OK to upscale photos of other people?
Upscaling your own photos is routine. Be careful with images that will be used as evidence, in journalism or for identification. Upscaled facial detail is generated, and it can mislead. Label enhanced images where that matters.