Improving Low-Resolution Face Recognition under Limited Data: How Synthetic Data Generation Can Close the Domain Gap

Luis S. Luevano, Ünsal Öztürk, Hatef Otroshi Shahreza, Anjith George, Sébastien Marcel

Idiap Research Institute, Switzerland

Accepted at IJCB 2026 — Focus Session on Generative AI for Fair and Secure Biometrics under Limited Data

Abstract

Face Recognition (FR) systems in surveillance settings often encounter Low Resolution (LR) faces, those whose face region falls below the standard 112×112 input size. While labelled High Resolution (HR) training data is abundant, labelled native-LR data, and above all paired native-LR/HR data, is scarce. One workaround is to synthesize LR data from the available HR faces, but how much synthesis effort is repaid in recognition accuracy remains unclear. We present a study of simple synthetic generation strategies for a compact, edge device-oriented face recognition system, spanning interpolation-based degradation, knowledge distillation, a Prepended Domain Transformer (PDT), Real-ESRGAN-style degradation, and a learned Super Resolution (SR) front-end with an identity-aware loss. We evaluate these strategies on synthetic cross-resolution face benchmarks (LFW, CFP-FP, AgeDB-30) and on TinyFace, a real-world native LR dataset, and expose a synthetic–real gap: the degradation setting that is optimal on synthetic benchmarks is not the one that is optimal on real LR. We find that more synthesis effort does not help monotonically: the learned SR front-end does not surpass a direct feed of the aligned LR image into a strong backbone, while simple interpolation augmentation of a compact backbone is the only synthesis that improves over its own baseline. We conclude that generative methods for LR face recognition must be validated on real LR and against a direct-feed baseline, and release our pipeline.

56×56 px

28×28 px

14×14 px

7×7 px

The same pair of faces captured at smaller and smaller sizes, each scaled back up to the size a face recognition system expects. The cues that identify a person disappear along the way. Surveillance cameras routinely produce faces this small, yet there is very little labelled low-resolution data to train on.

Key Findings

Using synthetically-generated degraded images is misleading. The degradation that gives the best results on artificially downsampled benchmarks is the one that gives the worst results on genuinely captured low-resolution faces, worse than not adapting the model at all. A gentler degradation is what actually helps on real data.
More elaborate generation is not better generation. For a small model that can be retrained, plain downsampling used as data augmentation beats both a realistic degradation simulator and a learned super-resolution front-end.
Always compare against the simplest option. Passing the aligned low-resolution image straight into a strong recognition model is a hard baseline to beat, and our learned super-resolution front-end never did. Any restoration or translation step should be measured against it before it is claimed to help.
Better average accuracy did not mean a more fair system. On RFW, recognition errors remain unevenly distributed across demographic groups no matter which degradation the model was trained on.

Approach

Adapting a face recognition system to low resolution involves two choices: how much effort to put into generating the low-resolution training images, and how to adapt the recognition model to them. We compare three levels of generation effort:

Low — simply downscaling high-resolution faces and scaling them back up, used as data augmentation while retraining a small model.
Medium — a Real-ESRGAN-style degradation simulator that adds blur, resizing, noise and compression, so that the generated faces look like genuinely captured ones.
High — a learned super-resolution network that reconstructs a high-resolution face and is trained to keep the person's identity, placed in front of a recognition model that stays frozen.

These choices are independent of how the recognition model itself is adapted: retraining it on the generated data, distilling it from a high-resolution teacher, or leaving it untouched and only transforming the incoming image. Every option is compared against the simplest one, which we call the direct feed: align the low-resolution face, resize it, and pass it to a model trained on high-resolution images, with nothing in between.

Learned super-resolution network and image translator in front of a frozen recognition model

The highest-effort option we tested: a learned super-resolution network and an image translator turn a low-resolution face into something a frozen recognition model can handle.

The Gap Between Generated and Real Low Resolution

Models that look good on artificially downsampled benchmarks (LFW, CFP-FP, AgeDB-30) do not keep their ranking on genuinely captured low-resolution faces (TinyFace). The degradation that comes first on the artificial benchmarks comes last on the real ones, and a much gentler degradation performs better instead. Training on a gentle degradation keeps the model closer to what real low-resolution capture actually looks like.

The same holds for methods that transform the incoming image instead of retraining the model. An image translator fitted to artificially degraded faces does not carry over to real ones and mostly just smooths them out. And knowledge distillation, which gives by far the largest gains on artificially degraded data, also ends up behind a plainly retrained model on real data.

Translator outputs on real low-resolution TinyFace images. Having been trained on artificially degraded faces, the translator mainly smooths the input.

Super Resolution and Prepended Domain Transformers (PDT)

The learned super-resolution network does extrapolate enough detail to translate into an increased recognition accuracy. Used on its own, the network does not perform as well as passing the aligned low-resolution image directly to the recognition model. In the second stage of training with PDT, the super resolved output changes in favor of input useful for the domain transformer and increases recognition performance, but it still does not catch up with the direct feed baseline.

Sub-pixel reconstructions: input, reconstruction, high-resolution target

Real-ESRGAN-style reconstructions: input, reconstruction, high-resolution target

Reconstructions from our two super-resolution variants (input / reconstruction / high-resolution target) in two stages. We trained the Super Resolution module with an identity-aware reconstruction loss and then retrained with Contrastive Loss with the Prepended Domain Transformer (PDT) module.

Bias and Fairness

We also measured how evenly the models perform across demographic groups, using RFW, which covers African, Asian, Caucasian and Indian subjects. As in earlier reports on this dataset, the Caucasian group is recognized most reliably and the African group least reliably, in every setting we tested.

Training on low-resolution data does not reduce this difference. The settings that give the best average accuracy do not spread the errors more evenly across the four groups; the imbalance moves in both directions depending on the test resolution, with no consistent improvement. Better average accuracy therefore does not make the system fairer.

Source Code

Our training, data generation and evaluation pipeline will be released at:

[Idiap Gitlab] Code for the IJCB 2026 paper (Coming soon)

BibTeX


@INPROCEEDINGS{luevano2026improving,
  author={Luevano, Luis S. and Öztürk, Ünsal and Otroshi Shahreza, Hatef and George, Anjith and Marcel, Sébastien},
  booktitle={2026 IEEE International Joint Conference on Biometrics (IJCB)},
  title={Improving Low-Resolution Face Recognition under Limited Data: How Synthetic Data Generation Can Close the Domain Gap},
  year={2026},
  volume={},
  number={},
  pages={},
  note={Accepted at IEEE International Joint Conference on Biometrics (IJCB) 2026, Focus Session on Generative AI for Fair and Secure Biometrics under Limited Data}}