Projecting Old Face Photos Forward in Time to Match Low-Resolution Probes

Many face-matching problems pair a sharp photo taken years ago with a blurry, tiny face captured today. A missing child's school portrait is compared against a 16-pixel-wide face from a street camera. A decades-old passport photo is compared against CCTV footage. Two separate research lines address the two halves of this problem. One line builds face embeddings that ignore age. The other trains a network that sees low-resolution faces to imitate a network that sees high-resolution ones. To our knowledge, no published method or benchmark handles both at once.

This note proposes treating the old high-resolution photo and the new low-resolution probe as two different modalities. They are matched in a shared space learned with a contrastive objective, meaning a loss that pulls matching pairs together and pushes non-matching pairs apart. The key difference from existing work is that the gallery side is not made age-blind. Instead, we feed it the elapsed time and the subject's age at enrollment, which forensic and identity-document settings almost always record. The gallery embedding is then pushed forward in time in feature space, not image space.

The underlying bet concerns what downsampling erases. Downsampling removes adult texture aging such as wrinkles and skin creases. It leaves intact the coarse change in skull and jaw proportions that happens during childhood. Low-resolution cross-age matching is therefore mostly a geometric-growth problem. An age-invariant model trained on sharp faces is solving a texture problem that has already disappeared from the probe.

Core Claim

We make three claims. First, at probe resolutions of 16×16 pixels and below, the age information left in face embeddings is mostly geometric. It shrinks sharply for adults but much less for subjects enrolled in childhood. Second, because of this, distilling an age-invariant high-resolution model into a low-resolution student will help adult-to-adult matching but stall for child-to-adult gaps. Third, a gallery encoder conditioned on enrollment age and elapsed years, trained contrastively against low-resolution probes, will close much of the remaining child-to-adult gap.

This should hold without the identity drift that image-space aging generators introduce. It should also hold without the accuracy loss on sharp images that cross-resolution fine-tuning is known to cause.

The work differs from its closest neighbours in what it keeps.

  • Age-invariant models. MTLFace uses attention to split high-level feature maps into an identity part and an age part. It then penalizes the identity part for carrying age, and reports 96.23% on AgeDB-30 and 95.62% on CALFW with 112×112 crops. Hidden Factor Analysis, Decorrelated Adversarial Learning and Orthogonal Deep Features Decomposition follow the same discard-the-age-factor logic. They motivate the problem and supply our teacher, but they all throw away age. We instead use it as a known input.
  • Image-space aging. AGR-GAN and latent-diffusion aging also condition on target age, but they do it by generating pixels. Both are trained on sharp inputs and are prone to changing identity-bearing structure. We move the prediction into the embedding.
  • Cross-resolution distillation. Cross-Resolution Learning (CRL) and CRRCD supply the method for the probe side, distilling a high-resolution teacher into a low-resolution student. Both evaluate only on pairs taken at roughly the same time.
  • Text-guided matching. TGFR shows that a second modality can stabilize recognition of degraded faces. Its cross-modal loss is the template for our training objective. We use elapsed-time metadata as the conditioning signal rather than captions, for reasons given below.

Research Program

Diagnostic stage. Before building anything, we test the premise directly. We take a frozen face recognition backbone and an MTLFace-style age-invariant backbone. We extract embeddings of AgeDB and FG-NET faces after downsampling to 8, 12, 16, 24 and 112 pixels. We train a small regression probe on each to predict age. We report the probe's error separately for subjects under 15, 15 to 40, and over 40.

The geometric-growth hypothesis predicts a specific pattern. Age error for adults should rise steeply as resolution falls, while error for children should rise much more slowly. If both groups degrade at the same rate, the asymmetry the method relies on is absent and the conditioning idea loses its main motivation. This stage is cheap and publishable on its own. It is also the first point at which the project can stop.

Model. The gallery tower $f_H$ is a frozen high-resolution face encoder. A small residual network $h$ shifts its output by an amount that depends on enrollment age $a$ and elapsed years $\Delta$. The resulting gallery embedding is

$$g = f_H(x) + h\big(f_H(x), a, \Delta\big), \qquad h(\cdot, a, 0) = 0,$$

where $x$ is the enrollment photo. The constraint at $\Delta = 0$ is enforced by multiplying $h$'s output by a learned gate that is zero at zero elapsed time.

Freezing $f_H$ and forcing a zero shift at zero elapsed time means that high-resolution-to-high-resolution matching is mathematically unchanged. This removes, by construction, the roughly 4% loss on sharp images that CRL reports after cross-resolution fine-tuning.

The probe tower $f_L$ is initialized from $f_H$ and sees low-resolution probes. Degradation is applied by a curriculum from 112 down to 8 pixels, following CRL, and mixes in blur, sensor noise and JPEG compression. This addresses the point that real surveillance degradation is never plain bicubic downsampling.

Training objective. Training combines two terms. The main term is a symmetric InfoNCE loss between $f_L(y)$, the embedding of a later low-resolution probe $y$, and the forward-projected gallery embedding $g$ of the same person. The second term is CRRCD's relational distillation on same-time pairs, which keeps the student's neighbourhood structure faithful to the teacher.

Negatives are drawn partly at random. The rest are matched to the anchor by estimated gender, ethnicity and age bracket. This copies the construction of CALFW, in which attribute-shortcut verification falls from 86.23% to 51.80% once negatives are demographically matched. Without such negatives, a low-resolution student can score well by recognizing demographics rather than identity.

Training data. We train on LCAF/SCAF from the MTLFace paper if they are obtainable, and otherwise on CACD with estimated ages. The label noise in web-crawled ages is a known risk. We therefore report results with ground-truth ages on AgeDB, and also with noise injected into $a$ and $\Delta$, to measure how sensitive the method is to bad metadata.

Baselines

Naive reference points. The do-nothing reference is the frozen high-resolution teacher applied directly to bicubically upsampled probes. CRL reports that the unadapted state-of-the-art model reaches only 4.8% True Acceptance Rate (TAR) on IJB-B at 8×8 pixels. TAR is the fraction of genuine pairs accepted at a fixed false-accept rate. We also report MTLFace applied the same way, to show how much an age-invariant model loses when its texture cues are gone.

Strongest competitor. The strongest competitor is an age-invariant teacher distilled into a low-resolution student. Concretely, MTLFace's identity branch serves as teacher, and the student is trained with CRL's curriculum MSE distillation plus CRRCD's relational contrastive loss. This is the most natural composition of existing work. It is exactly what a reviewer would propose, and it isolates whether our age conditioning adds anything over age invariance.

For fairness, it receives the same backbone, the same training images, the same degradation pipeline, the same demographically matched negatives and the same hyperparameter search budget. The reported numbers closest to this setting are:

  • CRL: 42.0% TAR on IJB-B at 8×8.
  • CRRCD: 95.25% accuracy on 16×16 LFW.
  • MTLFace: 96.23% on AgeDB-30 at full resolution.

No reported result was found in the research file for any method on cross-age benchmarks with downsampled probes. We will therefore establish this baseline ourselves.

What counts as winning. Beating this baseline means a statistically reliable improvement in TAR at a false-accept rate of 1e-3 on child-enrollment pairs at 16×16 and 12×12, measured across three seeds. It also means matching the baseline within noise on adult pairs. A gain that appears only on adult pairs would contradict our own hypothesis.

Generative baseline. The image-space alternative ages the high-resolution gallery photo to the probe's age with AGR-GAN or a latent-diffusion aging model, then matches with the distilled student.

Ablations. Three ablations isolate the key ingredient:

  • Replace $h$ with an age-blind residual whose age and elapsed-time inputs are set to constants. This tests whether gains come from extra parameters rather than temporal information.
  • Feed $h$ shuffled values of elapsed time.
  • Remove the demographically matched negatives.

Why not just…?

Why not just distill an age-invariant model into a low-resolution student? This is the real threat, and it may well suffice for adults. Our diagnostic stage predicts that adult aging cues are already destroyed at low resolution. In that case, age invariance and age conditioning should converge for adults. Where they should differ is childhood growth. That change is geometric and identity-specific, and it survives downsampling. An age-blind embedding can only suppress it, whereas a conditioned one can predict it. The comparison stratified by enrollment age decides this question. If the distilled baseline matches us on FG-NET-style child pairs, the conditioning idea is unnecessary and we will report that.

Why not just age the gallery photo with a GAN or diffusion model, then match? These generators condition on target age just as we do, but they render the result as pixels. MTLFace notes that face age synthesis often fails to preserve identity. CRL found that pixel-space hallucination by super-resolution damages biometric fidelity compared with feature-level distillation. The generative baseline is run under identical conditions. If it wins, the claim that feature-space projection avoids identity drift is wrong.

Why not just super-resolve the probe and use MTLFace? Super-resolution invents high-frequency texture from dataset priors, and MTLFace's age branch is built to read exactly that texture as age. CRL reports that super-resolution preprocessing underperforms direct distillation. We include the pipeline as a baseline mainly to check whether invented wrinkles mislead the age branch. To do so, we measure the age that MTLFace predicts on super-resolved faces against the subjects' true ages.

Why not just use text descriptions of stable facial traits as the second modality, as TGFR does? This was our first design. TGFR reports an equal error rate of 17.16% on very-poor-quality images. However, no age-separated benchmark in the file has captions. Captions generated by an attribute model run on the gallery image would be another teacher signal passed through language. DiFace's observation that text describes many people at once also caps how much identity information captions can carry. Elapsed-time metadata, by contrast, is exact, available in real case files, and directly targets the factor that changes. We therefore drop text from the core method. We keep one side experiment that adds stable-trait attribute descriptors to $h$'s input, to see whether they add anything beyond age.

Proposed Experiments

Benchmarks. Evaluation uses AgeDB-30 and CALFW for adult gaps and FG-NET for child-to-adult gaps. In each, the probe side is downsampled to 8, 12, 16 and 24 pixels, with and without compound noise, while the gallery stays at full resolution. We report verification accuracy and TAR at false-accept rates of 1e-3 and 1e-4, stratified by age gap (AgeDB's 5- to 30-year protocols) and by enrollment age. For FG-NET, no reported result was found in the research file even at full resolution. We will report MTLFace's full-resolution number ourselves as the ceiling.

Sanity checks. TinyFace and IJB-B at 8×8 serve as checks on the same-time setting. There our model should match the distilled baseline, because the shift is zero when elapsed time is zero. Full-resolution AgeDB-30 confirms that the 96.23% regime is untouched.

Decisive results. Three results decide the project:

  • The diagnostic probe must show that adult age information collapses faster with falling resolution than child age information does.
  • The conditioned model must beat the distilled age-invariant baseline on child-enrollment pairs at 12–16 pixels, while tying it on adults.
  • The shuffled-elapsed-time ablation must erase that gain.

Any one of these failing tells us which part of the story is wrong.