"Enhance" covers several different operations with very different reliability. The distinction that matters is whether the information still exists in the file.

Genuinely Recoverable

Noise — random variation layered over real signal. Because it's random and the signal isn't, denoising can separate them statistically. Modern AI denoising is excellent, and RAW files denoise better than JPEGs. The limit: aggressive denoising removes fine texture along with noise, producing plastic-looking skin and smeared foliage.

Underexposure (in RAW) — a RAW file holds more range than the preview shows, so shadow detail genuinely exists and can be lifted. In a JPEG, shadow information was already discarded, and lifting reveals banding and noise instead of detail.

Colour casts and white balance — fully recoverable in RAW, largely recoverable in JPEG.

Compression artifacts — models trained on compressed/original pairs can plausibly reverse JPEG blocking. Results are usually good because the damage pattern is well understood.

Mild softness — sharpening enhances edge contrast that is genuinely present.

Not Recoverable (the Tool Invents)

Motion blur and defocus — the information about what was behind the blur was never recorded. Deblurring models produce a *plausible* sharp version, not the real one. Fine detail, especially text and faces, is fabricated.

Blown highlights — pure white pixels contain no data. Any "recovery" is synthesis.

Resolution beyond the source — as with upscaling, detail is invented.

Occluded content — anything hidden behind an object.

The practical test: if you can't tell what was there by looking, neither can the model. It's producing a statistically likely answer.

The Face Problem

Face enhancement deserves special attention. Models trained on portraits will:

  • Smooth skin and remove texture, sometimes drastically.
  • Reconstruct eyes, teeth and features from learned averages, which can subtly alter identity.
  • Apply an aesthetic — symmetry, brightened whites, defined jawlines — that people didn't ask for.

For personal photos this may be welcome. For identification, journalism, historical archives or anything documentary, it's a serious problem: the enhanced face may not be that person's actual face. Several documented cases involve upscalers producing faces of a different apparent ethnicity or age from low-resolution inputs.

Practical Guidance

  • Always keep the original. Enhancement is not reversible.
  • Compare at 100% before and after, specifically at faces, text and fine texture.
  • Use the least aggressive setting that solves the problem. Most tools default to too much.
  • Denoise before sharpening, and do both at the end of your workflow.
  • Shoot RAW if you expect to rescue anything; it moves several items from the "invented" list to the "recoverable" list.
  • Disclose enhancement in journalistic, evidentiary or archival contexts, and never enhance images used for identification.

The honest framing: enhancement recovers information that survived, and fabricates information that didn't. Knowing which is which for your specific image is the whole skill.