ID-PreFeR

The problem

Sharpen a ruined photo and the identity drifts away.

A degraded face is consistent with many plausible sharp faces. A restorer that does not know who it is looking at will produce a convincing one, and it will not be the right person. The obvious fix is to show it other photos of the person — but the methods that do this either need a sharp reference to work from, or hours of training and hundreds of megabytes per person.

Measured against the degraded input itself

ArcFace1 similarity to the real person, on CelebRef2. The dashed line is not a baseline we chose — it is the score the degraded input itself gets. Bars to its right are restorers that made the identity more recognisable; bars to its left made it less. Five of the ten prior methods never get past the line.

What it asks for

Five photos of the person, not all of them sharp.

From Fig. 4 of the paper. Ground truth for this identity: view the reference photograph.

The approach

Learn who they are, learn how badly they were photographed, then throw the second part away.

Three components, each inexpensive. A rank-16 adapter on the decoder's cross-attention is the only part that is trained — 6.41 M parameters, 0.25% of SDXL3, about 13 MB in FP16. A second, frozen copy of the same backbone provides the regularization target, so no extra images are collected. At sampling time, the predicted noise is nudged toward the face in the reference, once per step over the last few steps.

Qualitative

Drag the handle. The restored face is still theirs.

Eight cases, six restorers, and the same degraded input on the left of every handle. The first four are CelebRef portraits degraded by the shared evaluation pipeline, so ground truth exists. The last four are in-the-wild photographs — two from the internet, two captured on a phone — where the only thing to check against is the person’s other photos.

Input ID-PreFeR

The album it read

Five references. The framed one is the photo chosen for identity-preserving sampling.

Ground truth

Panels are taken unaltered from Figs. 5 and 6 of the paper. Every method sees the same input and, where it accepts them, the same references.

Benchmarks

Everyone else trades identity against quality. The corner where both hold was empty.

Two axes, both from Table 1: how well a restoration matches the real person, and how good the image looks on its own. Sharpen aggressively and identity suffers; protect identity and the image stays soft. No prior method reached 0.70 identity and 0.60 quality at the same time.

On photographs nobody degraded on purpose

The synthetic benchmark controls the degradation; real phone photographs do not. Blur, noise and compression arrive together, and there is no clean original to score against. We therefore measure identity twice — once with a face recognizer, and once with a vision–language judge asked how much the restored face still looks like the person — because ArcFace embeddings themselves become unreliable under heavy mixed degradation.

Take one away

Each component fails in its own way, and the failure is visible.

The three components together account for the same 6.41 M parameters, so the ablations are not a size argument. They are a behaviour argument: remove a component, and the failure it was holding back returns — each with its own signature.

Without the latent prior loss

Colour drifts — usually toward magenta — and small facial artefacts appear. This is the adapter forgetting what faces look like in general while it learns what this face looks like in particular. The drift gets worse as sharp references grow scarcer.

Without ID–quality disentanglement

Without the quality tokens and the Min-SNR-γ cap, the model has no way to tell “this is what she looks like” from “this photo was taken badly,” and it learns both. The results become visibly blurrier, most strongly when almost every reference is degraded. The identity score barely moves — this component buys detail, not recognisability.

Without ID-preserving sampling

The face stays sharp and plausible, but the details that make it this person go soft — iris colour is the easiest to spot. This is the single largest identity component: 0.767 drops to 0.734. It is also the most expensive one, which is worth stating plainly.

Fig. 4 Rows one to three of the paper’s ablation figure, split so each variant can be compared against the complete model at the same reference budget.

What it costs

Four minutes per person, thirteen megabytes to keep.

Per-person tuning is the part one expects to be expensive, so it is worth being precise about. The adapter is small enough to store next to the album, and quick enough to train while the phone is on the charger.

Where it breaks

It learns what the references share. Anything they disagree about, it averages away.

This is the mechanism working as designed, but it has an edge case users will hit. When the references differ in makeup, jewellery, or glasses, those features cancel out during training — the adapter keeps only the shared identity. ID-PreFeR may then vary in exactly those attributes rather than follow any single reference.

All three are partial mitigations. A thorough study is left to future work.

Cite

ID-PreFeR

ID-Preserving Face Restoration with Mixed Data Quality. European Conference on Computer Vision (ECCV), 2026.

@inproceedings{zhu2026idprefer,
  title     = {{ID-PreFeR}: {ID}-Preserving Face Restoration with Mixed Data Quality},
  author    = {Zhu, Chengxuan and Hong, Yuchen and Zhang, Qi and Fu, Bingtao and
               Liang, Jinxiu and Chen, Jinwei and Zhang, Huaqi and Xu, Chao and
               Shi, Boxin and Fan, Qingnan},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}