acceptodds
Under review as a conference paper at ICLR 2027

DDFR: Degradation-Disentangled Diffusion-Regularized Feature Correction for Blurred-Query Face Retrieval

Abstract

Face retrieval across varying degrees of Gaussian blur is challenging because clean-gallery and blurred-query representations are inherently asymmetric: conventional systems assume sufficiently clear probe and gallery images, whereas real deployments often match blurred probes against clean stored identities. Gaussian blur suppresses discriminative cues and displaces the blurred-query embedding from its clean-gallery identity anchor, degrading cosine ranking. Existing recognizers compare degraded queries directly with clean-gallery embeddings, while blind face restoration (BFR) preprocessing pipelines optimize visual quality rather than retrieval alignment. We introduce DDFR, a degradation-aware and diffusion-regularized framework that corrects blur-induced displacement in feature space. A shared encoder produces identity and degradation representations; the latter estimates an identity-space correction vector and is regularized by diffusion-style latent denoising. Directional separation discourages identity-aligned correction, and the estimated vector is subtracted from the blurred-query representation before retrieval. We further construct a dedicated identity-disjoint Gaussian-blur face retrieval dataset covering σ = 1, ..., 10. On this benchmark, DDFR achieves 88.05 ± 1.26% average Rank-1 accuracy, while ArcFace and AdaFace trained on the same constructed dataset achieve 59.52% and 65.58%, respectively. At σ = 10, DDFR retains 60.76 ± 3.74% Rank-1 accuracy, compared with 21.37% and 28.06% for the same references. Component analyses support degradation modeling, latent denoising, and directional separation; the extended sweep identifies the limit imposed by severe information loss. These results suggest a path for lower-resolution camera inputs to approach the retrieval utility of clearer, higher-quality captures when identity cues remain recoverable.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
face retrieval, low-quality face recognition, Gaussian blur, feature correction, degradation disentanglement, diffusion regularization, representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_mgKHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This work propose DDFR, a method to distangle degardation of low-quality faces to further accomplish the face retrival task. To evaluate the performance of DDFR, the authors curate a benchmark where input is low-quality (blurred) faces and the query target is clean faces. The DDFR achieves best performance on this benchmark.

Critical strengths
  • As claimed in this work, the authors try to tackle with face retrieval task, where the input face is blurred, as an asymmetric embeddingalignment problem.
  • The reprodulicity is available as the code and guidance are provided.
Critical weaknesses
  • The experiement result is not sufficient. As the authors have only evaluated on the self-curated benchmark, is there any similar benchmark for this setting? If so please also evaluate DDFR with them.
  • The discussion with other competing methods are not enough. What is the difference and novelty between DDFR and other blurred face retrieval methods?
  • Method presentation is not enough. How "distangle" is achieved, and how it could help to the face retrieval task?
Questions for authors

Please refer from the weakness part.

1 comment

Response to Reviewer mgKH

Thank you for recognizing DDFR's formulation of blurred-face retrieval as asymmetric embedding alignment and for noting the accompanying code and guidance.

1. Public benchmarks and the controlled blur protocol

We use a controlled protocol to measure how retrieval changes with the applied blur strength. Each Gaussian query is derived from its clean gallery photograph, so the source, pose, illumination, and crop stay fixed as the blur radius varies. This gives each evaluation point a known and reproducible degradation level. TinyFace provides a complementary measure of retrieval on native low-resolution images, whose standard annotations do not provide a calibrated blur-severity scale. TinyFace benchmark

We also evaluated the submitted model on two public datasets. For CFP, whose standard task is pair verification, we constructed a blurred-query, clean-gallery retrieval protocol. At σ=10, DDFR achieved 74.132% Rank-1 over 22,487 query evaluations in the same-source control, where the clean source photograph was the gallery positive. The study also evaluated cross-photograph retrieval with the query source excluded from the gallery. CFP benchmark · CFP same-source and cross-photograph results

On TinyFace, we evaluated naturally degraded probes against the official gallery of genuine matches and distractors. The protocol included 3,728 planned queries and 157,871 gallery entries. TinyFace results and denominators

2. Difference from prior blurred-face methods

Vageeswaran et al. model the possible blurred appearances of a gallery face as a convex set and match each probe against those sets. Blind restoration methods such as CodeFormer reconstruct a face image, which a retrieval system can then encode. DDFR learns to estimate a correction vector from the query's degradation features and applies it directly to the retrieval embedding. One deterministic latent reconstruction produces the correction, and cosine search uses cached clean-gallery embeddings. Set-theoretic blur-robust recognition · CodeFormer

3. How feature separation supports retrieval

A shared ResNet-18 supplies an identity feature aa and a degradation feature dd through separate branches. DDFR projects dd into a latent, applies degradation-conditioned latent denoising, then predicts a 512-dimensional correction γ\gamma. It ranks the normalized embedding r=(a−γ)/∥a−γ∥2r=(a-\gamma)/\lVert a-\gamma\rVert_2 against the gallery. Inference uses the same fixed condition at every blur radius and requires no query blur label.

During training, identity classification and supervised contrastive learning organize the corrected embeddings by identity, and paired consistency aligns their clean and degraded views. Severity regression supervises the degradation latent, and noise prediction trains the latent denoiser. The directional penalty ∣cos⁡(a,γ)∣|\cos(a,\gamma)| discourages corrections along the identity feature direction, while a clean-anchor loss limits drift of the clean embedding. This is a functional separation between the feature used for identity retrieval and the pathway that estimates its degradation correction.

We tested the complete design against a residual MLP that replaced all components after the raw ResNet-18. The non-backbone parameter counts were 18.690M for DDFR and 18.469M for the MLP. Across five seeds, we paired backbone initialization and matched successful encoder updates in the base and direct training stages. Each model encoded the same clean gallery and degraded queries with its own checkpoint.

Source-paired Rank-1, mean ± SDDDFRWhole-backend MLPDDFR gain
Average across σ=1–1091.462 ± 0.451%86.543 ± 3.993%+4.919 pp
σ=1069.029 ± 1.420%58.663 ± 8.023%+10.366 pp

DDFR led in four of five seeds at σ=10. The comparison evaluates the structured backend together with its training objectives and supports its effectiveness under severe blur in the source-paired setting. We will annotate the method figure to show where each training loss acts and how the corrected query embedding reaches the gallery search. Five-seed results and training protocol

Reviewer_e7iLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies asymmetric face retrieval between blurred query images and clean gallery images and proposes DDFR. The method separates identity and degradation representations. The degradation branch predicts the blur-induced displacement in identity space, which is then subtracted from the query feature. The model also uses diffusion-style latent denoising, blur-severity supervision, and a directional separation loss.

Critical strengths

1.The problem is clearly formulated as feature alignment between clean gallery images and blurred queries. 2.The method operates directly in feature space, avoiding image restoration and gallery re-encoding. 3.DDFR substantially outperforms ArcFace, AdaFace, and blind face restoration pipelines under the proposed protocol, particularly under severe blur.

Critical weaknesses

1.Each test query appears to be created by directly blurring its corresponding gallery image. The pair therefore shares the same pose, illumination, and crop, making the task easier than real cross-image retrieval and potentially overestimating practical performance. 2.The experiments mainly consider synthetic Gaussian blur and do not evaluate motion blur, downsampling, compression, noise, or detection errors. The pseudo-identity construction for WIDER FACE may also introduce labeling errors or duplicated identities. 3.DDFR uses additional branches and paired supervision compared with ArcFace and AdaFace, but it is not compared with parameter-matched MLPs, conventional denoisers, or non-diffusion feature-correction models. It is therefore unclear whether the gains specifically come from diffusion regularization. 4.Only DDFR reports means and standard deviations over multiple runs, while the main baselines appear to use single runs. Additional parameter counts, computational costs, and query latency are also not reported.

Questions for authors

1.Is each test query obtained by directly blurring its corresponding gallery image? Could the authors evaluate retrieval between different original images of the same identity? 2.It's better the method was evaluated on real low-quality face data or other degradation types. 3.How much does the diffusion objective improve over parameter-matched MLPs, denoising autoencoders, or non-diffusion correction modules? 4.Do ArcFace, AdaFace, and DDFR use identical training budgets? Could multi-seed results be reported for all principal methods? 5.How are the pseudo-identities generated and verified?

1 comment

Response to Reviewer e7iL

Q1. Query source and cross-photograph performance

Yes. Each original Gaussian query is a blurred version of a clean gallery photograph. We separately evaluated CFP using a different gallery photograph of the same labeled person, with the query photograph excluded. DDFR attained 2.446% cross-frontal Rank-1 and 0.630% cross-pose Rank-1 at σ=10. These results restrict the advantage reported in Q3 to the source-paired protocol.

Q2. Low-quality images and detection

The constructed benchmark lets us measure Rank-1 at known applied blur strengths on the same source photographs. TinyFace supports objective identity retrieval, while its standard annotations provide no calibrated blur scale and natural blur may co-occur with resolution, pose, illumination, and crop changes. Fixing the source and gallery lets us vary the GaussianBlur radius reproducibly. TinyFace complements this measurement by testing naturally mixed degradations.

Across the tested source-paired perturbation settings, DDFR Rank-1 was 99.66–100.00% for motion blur, 99.94–100.00% for downsampling, 81.43–100.00% for JPEG compression, and 98.91–100.00% for additive noise. Simulated misalignment was more disruptive, with Rank-1 spanning 3.03–94.51%. These ranges describe the tested settings and common-source comparison.

The CFP blur-plus-noise experiment and real low-quality TinyFace subset also test detection and five-point alignment on observed face crops. We report detection coverage and Rank-1 on aligned pairs separately to show which stage limits retrieval. We also report Rank-1 over all planned queries, counting detection failures as misses. On the same CFP queries, adding noise reduced the fraction of pairs that passed detection from 93.1% to 53.9%. Detector results and denominators

Q3. Whole-backend MLP comparison

The two systems tied at σ=1–2. At σ=7–10, the structured DDFR backend led by 8.38–10.37 points in mean Rank-1. DDFR estimates a degradation-conditioned correction in feature space and uses one deterministic query pass against the cached gallery. The residual MLP control keeps the raw ResNet-18 and replaces the full downstream backend. Non-ResNet parameter counts differ by 1.18%, and each model re-encodes both sides with its own weights.

Rank-1 over five seedsFull DDFRWhole-backend MLPDifference
σ=1–10 average91.462 ± 0.451%86.543 ± 3.993%+4.919 pp
σ=1069.029 ± 1.420%58.663 ± 8.023%+10.366 pp

At σ=10, DDFR led on four of five seeds. The familywise paired bootstrap 95% interval was [1.821, 17.063] pp and the exact two-sided seed sign-flip p was 0.125. This comparison measures the effect of replacing the complete backend and changing its applicable training objectives. The observed gain therefore supports the combined DDFR design on this protocol. The isolated effect of the diffusion objective remains outside the scope of this comparison. Whole-backend training contract

Q4. Scope of the five-seed budget comparison

The original table used historical reference checkpoints with different training lineages. We will withdraw the wording that their training data and budgets were identical. The five-seed structural comparison pairs the raw backbone initialization and matches successful encoder-update budgets in the base and direct stages. Selected checkpoints and later training-stage exposure differ.

Inference measureFull DDFRWhole-backend MLP
Conv/Linear MACs per query489.17M492.19M
Batch-1 p50 latency across seeds3.684–4.081 ms1.893–2.785 ms

The timing measures synchronized FP32 model execution with a cached gallery on an RTX PRO 6000 Blackwell. Image decoding, transfers, and similarity search are outside this timing scope. Training and measurement protocol

Q5. Pseudolabel generation and validation

WIDER keys come from upstream crop folders. The processed dataset applies the same key to a source crop and its synthetic variants, allowing us to specify clean and blurred pairs and positive gallery entries. We use the term pseudo-identity key for this dataset grouping. Image-level screening and registration found one photograph in both training and validation, with a different crop in each. Duplicate audit

Reviewer_ADZXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies retrieval of Gaussian-blurred face queries against a clean gallery. DDFR uses a shared encoder with identity and degradation branches, predicts a correction vector from the degradation representation, and subtracts it from the identity feature before normalization and retrieval. Training combines identity and paired-view alignment objectives with blur-severity regression, directional separation, and diffusion-style latent denoising. Inference uses a single deterministic latent reconstruction with a fixed condition across blur radii, rather than iterative diffusion sampling. On the constructed paired-image benchmark, DDFR reports 88.05 ± 1.26% average Rank-1 accuracy, compared with 59.52% for ArcFace and 65.58% for AdaFace trained on the same constructed dataset.

Critical strengths
  1. A coherent approach to a meaningful retrieval problem. Matching degraded queries to a clean gallery is practically relevant, and directly correcting retrieval features is well aligned with the downstream objective. Predicting and subtracting a degradation-dependent correction gives the method a clear organizing idea. The inference procedure is also concrete: gallery features are cached, and each query requires a single deterministic correction pass without image generation or iterative diffusion sampling.

  2. Substantial improvements under the stated controlled protocol. Table 1 reports a 22.47-percentage-point improvement in average Rank-1 over same-data AdaFace, with a substantial advantage also at the strongest reported in-range blur level. The paper explicitly trains ArcFace and AdaFace on the same constructed Gaussian-blur dataset, addressing the simple explanation that only DDFR receives blur-augmented training data. The component ablations provide additional evidence that the proposed training and correction pipeline is useful in this setting.

Critical weaknesses
  1. The evaluation does not adequately distinguish identity retrieval from matching a blurred image to its own source image. Section 4 states that queries are generated by blurring paired gallery images, and that the test set contains 1,715 manifest identities and 1,750 gallery entries, with only 35 identities contributing a second clean view. Consequently, the protocol largely permits matching a degraded image to the same photograph rather than recognizing the person in a different photograph. The gains may therefore reflect strong instance-level blur invariance without establishing comparable improvements in identity generalization. Furthermore, some manifest identities are pseudo-identities, and the paper acknowledges that disjoint keys do not guarantee disjoint real people. These limitations do not invalidate the explicitly defined paired-image benchmark or establish data leakage, but they materially restrict the identity-level interpretation and practical significance of the results. A cross-image evaluation with verified identity labels and the query's source photograph excluded from the gallery would directly address this concern.

  2. The evidence does not sufficiently isolate the proposed correction mechanism from paired-view supervision and additional modeling capacity. DDFR combines consistency and supervised contrastive objectives with a degradation branch, blur regression, directional regularization, and latent denoising. The same-data ArcFace/AdaFace comparisons establish an overall system advantage, but do not determine whether a simpler feature-alignment or residual-correction model with comparable supervision could obtain similar gains. Figure 4 shows substantial ablation drops, but the supplied manuscript does not fully specify whether each variant is retrained and how the remaining paths and objectives are maintained. In particular, poor retrieval from the uncorrected intermediate feature is not by itself evidence of recovered identity information, since the retrieval objectives supervise the corrected representation. The paper appropriately notes that directional separation does not prove statistical independence; correspondingly, the stronger degradation-disentanglement interpretation remains in

Questions for authors

Q1. Does the improvement persist when the blurred query and its correct gallery match come from different photographs of the same verified person, with the query's source photograph excluded from the gallery? This would distinguish identity-level generalization from same-image matching while retaining the Gaussian-blur setting.

Q2. How are the WIDER FACE manifest identity keys assigned and checked? Can you distinguish results on verified person identities from results on pseudo-identities, and clarify what real-person overlap checks are possible across the training and test sources?

Q3. How does DDFR compare with a model using the same backbone and paired consistency/contrastive supervision, but either direct feature alignment or a simple residual correction head without diffusion-style latent denoising? This comparison would help establish the incremental value of the proposed mechanism rather than merely the overall training recipe.

Q4. Are all Figure 4 variants trained separately, or are any evaluated by disabling components of the full model? For each variant, what replaces the removed module, which losses remain active, and are both gallery and query embeddings recomputed using that variant? Please also clarify whether the “identity feature only” result comes from a separately trained retrieval model or an intermediate feature of full DDFR.

1 comment

Response to Reviewer ADZX

Q1. Different photographs of the same person

The main benchmark blurs the clean gallery photograph at specified radii so that pose, illumination, and crop stay fixed. In the separate CFP test, we excluded the query photograph from the gallery and used a different photo of the same labeled person. At σ=10, the submitted DDFR achieved 2.446% cross-frontal Rank-1 and 0.630% cross-pose Rank-1. The gain over the MLP reported in Q3 was observed on the source-paired benchmark. Its persistence across different photographs remains open.

Q2. Why the benchmark uses WIDER pseudolabels

The dataset builder inherits each key from an upstream data_wider_distance folder and assigns that key to the stored source crop and its synthetic derivatives. We use these stable keys to generate the clean and blurred pairs and to specify the positive gallery entries. We call them pseudo-identity keys because they encode the dataset's grouping of stored crops. A key match defines a hit on this benchmark, whose purpose is to measure retrieval under controlled blur applied to those sources.

The resulting splits have 13,344 training, 1,716 validation, and 1,715 test keys. During the image-level audit, perceptual screening and registration confirmed that train_006593 and val_001769 contain different crops of one photograph. Duplicate audit

Q3. Value of the structured retrieval backend

The complete DDFR backend achieved 69.029% Rank-1 at σ=10, compared with 58.663% for the whole-backend MLP. DDFR predicts a degradation-conditioned correction in feature space and uses one deterministic query pass against a cached clean gallery.

The independently trained control keeps the raw ResNet-18 and replaces its entire downstream backend with a residual MLP retrieval head. Initial raw backbone weights are paired by seed. The two backends differ by 1.18% in non-ResNet parameters (18.690M versus 18.469M). Each system uses its own checkpoint to encode both the gallery and the queries.

The two systems tied at σ=1–2, and DDFR's mean advantage ranged from 8.38 to 10.37 points at σ=7–10. Over σ=1–10, mean Rank-1 was 91.462% versus 86.543%. At σ=10, DDFR led on four of five seeds. The 10.366-point difference had a familywise paired bootstrap 95% interval of [1.821, 17.063] pp and an exact two-sided seed sign-flip p of 0.125. The applicable training objectives and later-stage exposure differ, so the estimate concerns the complete architecture and training design together. Whole-backend training contract

Q4. Figure 4 and identity-only evaluation

The 27.09% row is an independently trained CE-only model (27.154% in replay). Full DDFR's internal uncorrected identity representation is a separate measurement at approximately 88.04%. All historical variants re-encode both query and gallery with their own checkpoints.

VariantRecorded change
Full DDFRStaged training with a final purifier fine-tune and direct-context retrieval.
CE-onlyIndependently trained with classification loss.
No degradation branchDisables degradation-driven correction and retains applicable identity and paired-view losses. The saved setup can retain identity-space purification.
No latent denoiserBypasses latent noising and the denoiser, with Ldiff off. Projections and correction remain.
No directional lossSets the directional penalty weight to zero.
No conditionFixes the condition index.

The revised caption will identify the two identity measurements separately. Checkpoint-specific objectives, stop rules and forward paths are recorded in the variant audit. Variant audit

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.