The Real Fear Behind Every Restoration Project
Nobody worries about megapixels. The actual fear — the one that stops people from clicking “restore” — is: “Will this tool change the person I remember?”
It is a legitimate concern. AI restoration models can and sometimes do alter facial features in subtle ways: narrowing a jawline, smoothing out a distinctive nose shape, making eyes slightly more symmetrical than they were in life. These changes are small enough that the photo still looks “good” — but the person no longer looks quite like your grandmother. Researchers call this phenomenon identity drift.
This guide explains why it happens at a technical level, how to spot it, and what practical steps you can take to minimize it.
Reconstruction, Not Recovery: How AI Restoration Actually Works
When a photograph has severe blur, grain, or physical tears across a face, the original pixel data is destroyed. It does not exist anymore. There is no hidden layer of information waiting to be unlocked — the silver halide crystals that formed the image are physically gone.
What AI restoration models do instead is reconstruct — they predict what the missing pixels probably looked like based on two inputs:
- The surviving pixels in your photo. The intact parts of the face (an undamaged left eye, a clear jawline on one side) provide reference data that the model uses to infer the damaged areas.
- A generative prior — millions of other faces. Models like GFPGAN and CodeFormer are trained on large datasets of high-quality face photographs. When your photo is missing detail, the model fills the gap with patterns it learned from these other faces — not from your grandmother's actual face.
This is the core tension: the more damaged the input, the more the AI has to rely on its training data (other people's faces) rather than your actual photo. And the more it relies on training data, the higher the risk of identity drift.
The 5 Causes of Identity Drift
1. Extreme low input resolution
When a face occupies fewer than roughly 64×64 pixels in the source image, the AI has so little reference data that it must generate up to 90% of the facial structure from its training prior. At this point, the model is essentially painting a new face that is statistically plausible — not restoring the original one. This is the single biggest cause of identity drift, and it is why scanning at high DPI matters so much.
2. Physical damage directly over key facial landmarks
There are specific facial features that carry most of your identity signal: the inter-pupillary distance (space between your eyes), the nose bridge width, the lip contour, and the jawline angle. When a scratch, tear, or water stain cuts directly through one of these landmarks, the model has to hallucinate that geometry. Researchers at Tencent ARC documented this as a core limitation of GAN-based face restoration — the generator prioritizes producing a “plausible-looking” face over an accurate one.
3. Over-aggressive face enhancement (“beautification bias”)
Many AI restoration tools optimize for perceptual quality — how “good” the output looks to a human viewer — rather than pixel-level accuracy to the original. This creates a systematic bias toward smoothing skin, straightening noses, and making faces more symmetrical. The result looks impressive as a photograph, but it no longer looks like the specific person. Wrinkles, moles, scars, and unique asymmetries that made your grandfather's face his face get quietly erased.
4. Training data demographic bias
AI face models are trained predominantly on contemporary, well-lit photographs of younger adults. When the input is a 1940s portrait of an elderly person with deep wrinkles, strong light-and-shadow contrast, or non-Western facial structures, the model has fewer reference examples to draw from. This can cause it to subtly shift features toward the “average face” in its training set — a documented problem in NIH-published research on AI facial recognition bias.
5. Colorization-induced skin tone shift
When restoration and colorization are applied simultaneously, the AI is performing two generative tasks at once. The color layer can alter the perceived shape of facial features — warm skin tones can make cheekbones appear flatter, cool tones can make a jawline look sharper. This is why the restore-first, colorize-second workflow produces better identity preservation than a single combined pass.
The Uncanny Valley Effect in Restoration
The uncanny valley is a well-documented psychological phenomenon: when a face looks almost real but has subtle imperfections, our brains register discomfort rather than simply noticing “low quality.”
In AI restoration, this often shows up as:
- Skin that looks “waxy” or plastic — the AI removed natural texture (pores, fine lines) and replaced it with a smooth, synthetic surface.
- Eyes that feel “dead” or vacant — the AI reconstructed iris detail and specular highlights that are technically correct but lack the micro-asymmetry of real eyes.
- Teeth that are too perfect — generic AI face priors default to straight, white teeth because that is what appears most frequently in modern training data. A 1950s portrait subject likely did not have Hollywood-perfect teeth.
When a family member sees the restored version and says “something feels off” without being able to pinpoint what, they are likely experiencing the uncanny valley. The photo is too perfect to feel authentic.
How to Verify Likeness After Restoration
You do not need specialized software to check for identity drift. Here are practical techniques anyone can use:
The side-by-side test
Open the original scan and the restored version next to each other at 100% zoom. Focus specifically on these five invariant landmarks — features that should not change during restoration:
- Inter-pupillary distance — the space between the centers of both eyes
- Nose bridge width — measured at the narrowest point between the eyes
- Lip contour shape — the Cupid's bow and commissure angles
- Ear shape and attachment angle — often overlooked, but highly individual
- Jawline asymmetry — real faces are never perfectly symmetrical
If any of these five landmarks shifted noticeably, the AI introduced identity drift.
The family member test
Show the restored photo — without the original beside it — to a family member who knew the person. Do not prompt them. Simply ask: “Does this look like [name]?” If they hesitate, furrow their brow, or say “something's different,” the AI likely drifted. Human facial memory is remarkably specific, especially for loved ones.
The second reference photo test
If you have a second photograph of the same person from a different angle or era, compare it against the restored version. Consistent facial proportions across multiple source photos is a strong signal that the restoration preserved identity accurately. This is particularly useful for composite projects where you are working with reference photos of the same person.
Input Quality Makes the Difference
The single most effective way to prevent identity drift is to give the AI more data to work with. Here is a practical breakdown:
| Face Size in Scan | AI Reliance on Prior | Identity Drift Risk |
|---|---|---|
| 256px+ across face | Low — uses mostly your actual pixels | Minimal |
| 128–256px | Moderate — blends your pixels with prior | Moderate |
| 64–128px | High — prior dominates reconstruction | High |
| Under 64px | Near-total — essentially generating a new face | Very high |
To maximize the face pixel count, scan at 600 DPI or higher. A 4×6” print scanned at 600 DPI produces a 2400×3600px image — giving even a small face in a group photo enough resolution for faithful reconstruction. See our complete scanning guide for DPI recommendations by print size.
When AI Can't Help: Knowing the Limits
Honesty about limitations is more valuable than false promises. There are situations where no AI tool — including BringBack — can restore a face without significant identity drift:
Face is a tiny speck in a large group photo — if the face is under ~40 pixels across in the scan, even 1200 DPI rescanning won't provide enough data. The AI will produce something that looks like a face, but it will be a generated face, not the original person.
The entire face area is physically missing — a hole in the print, a burn, or a water stain that dissolved the emulsion across the full face. With zero surviving reference pixels, the AI has nothing to anchor reconstruction to.
Heavy motion blur on the only copy — motion blur is directional information loss. Unlike grain or noise (which is random), motion blur systematically smears facial geometry in one direction, and the model cannot reliably reverse it without introducing drift.
What to do instead: In these extreme cases, consider using the add person to photo tool with a separate, clearer reference photo of the same person. This lets you place their likeness from a better source into the scene, rather than asking AI to invent facial detail from almost nothing.
Ready to try restoration with identity-aware processing? BringBack's side-by-side comparison tool lets you inspect original vs. restored pixels at full zoom before downloading — so you can verify likeness before committing.
Try the Restoration Tool