QualiTeacher: Quality-Conditioned Pseudo-Labeling for Real-World Image Restoration¶

Conference: ECCV2026
Paper: ECCV Paper
PDF: Full Paper
Code: https://github.com/fengyang1399-pixel/QualiTeacher.git
Area: Image Restoration
Keywords: quality conditioning, pseudo-labels, mean teacher, no-reference image quality assessment, preference optimization

TL;DR¶

QualiTeacher turns teacher pseudo-label quality into a control condition for the student, then constrains quality ordering and cropped consistency so that imperfect supervision supports controllable real-world restoration, reducing RTTS dehazing FADE from 0.824 for CORUN with Colabator to 0.590.

Background & Motivation¶

Real-world image restoration rarely provides a precisely aligned clean target: the same scene usually cannot be photographed again with identical viewpoint and lighting after removing snow, haze, or sensor noise. Synthetic paired data can establish basic restoration capability, but it does not cover the complex combinations and spatial variation of real degradations. The Mean Teacher framework generates pseudo-labels for real degraded images, trains a student against them, and updates the teacher through an exponential moving average of the student. This loop exploits unlabeled images, but residual snow, blurred textures, and color shifts in teacher outputs also become supervision targets. Methods such as Colabator improve target reliability through label banks and quality selection, yet a high score does not necessarily mean that the retained image is genuinely clean.

Unconditionally fitting every pseudo-label mixes restoration targets of different quality into one mapping; retaining only a small set of top-scoring targets sacrifices diversity in both image content and quality variation. No-reference image quality assessment (NR-IQA) cannot fully resolve this problem, because evaluators may favor oversmoothed outputs or mistake sharp but artificial repetitive textures for high quality. Rather than training a larger teacher, this paper changes how the student interprets supervision: targets for a degraded input can occupy different quality levels, and that level should be visible to the network. A low-quality target need not represent the ideal restoration; it can instead represent the output associated with a low-quality condition. This also explains why adding a score alone is insufficient: the network may ignore it, or create a quality gap simply by worsening its low-score output.

The authors combine quality conditioning with explicit ordering objectives, aiming to learn a restoration direction along the score axis and select a high-quality condition at inference time. Here, surpassing the teacher is a learning objective and an interpretation of the experiments, not a mathematical guarantee for every input. In particular, increasing an evaluator's score is not equivalent to recovering true structure when the evaluator is biased. Core Idea: turn pseudo-label quality from a binary retention criterion into a learnable conditioning coordinate, while requiring high-score outputs to improve and local quality evaluations to remain consistent instead of unconditionally imitating the teacher.

Method¶

Overall Architecture¶

Inputs include labeled image pairs and real degraded images without clean references; the output is an image restored by the student under a specified quality condition. The unlabeled branch first builds candidates and a historical bank through Quality-Graded Pseudo-Labels, then uses Quality Conditioning and Spatial Weighting to incorporate target quality and regional reliability into student training. Quality-Guided Optimization further constrains the ordering of output quality across conditions and suppresses evaluator-driven artifacts. The teacher is an exponential moving average (EMA) of the student and supplies training targets; inference only requires the conditional student, not the full candidate-selection pipeline for every test image.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    Input["Real degraded image"] --> Labels["Quality-Graded Pseudo-Labels<br/>Teacher augmentation, scoring, and bank"]
    Labels --> Condition["Quality Conditioning<br/>and Spatial Weighting"]
    Input --> Condition
    Paired["Paired images and clean targets"] -->|Supervised branch| Condition
    Condition --> Optimize["Quality-Guided Optimization<br/>Preference ordering and cropped consistency"]
    Optimize -.->|EMA of student parameters| Labels
    Optimize --> Student["Trained conditional student"]
    Test["Test image and condition score 7"] --> Student
    Student --> Output["Restored image"]

The label bank, quality evaluation, and loss branches belong to training; the conditional student is the deployed model. Quality scoring is not merely loss weighting: the global score enters network features, while block-level scores weight pixel supervision, addressing different problems. Figure 2 also distinguishes weak augmentation of labeled samples from strong augmentation of unlabeled student inputs; the teacher separately uses reversible geometric augmentation to produce aligned targets.

Key Designs¶

1. Quality-Graded Pseudo-Labels: obtain quality variation through augmentation and reject inconsistent candidates

For each degraded input, the teacher restores horizontally flipped, vertically flipped, and 90-degree-rotated versions, then maps the three outputs back through their inverse transformations. All pseudo-labels are therefore spatially aligned with the same input, although they may retain different amounts of snow, texture errors, or color distortion. These transformations are theoretically lossless, but restoration networks are not strictly geometrically equivariant, so the authors exploit response differences to broaden the supervision distribution rather than simply averaging the outputs. The scoring ensemble combines MUSIQ-KonIQ, BRISQUE, and CLIP-IQA: the first two assess image statistics or perceptual distortion, while the last introduces vision-language priors to reduce dependence on one evaluator. Because lower BRISQUE is better, the paper reverses its direction using \(100-S_{\mathrm{BRISQUE}}\) before score normalization.

Dual-Drop checks two axes: excessive disagreement between evaluators on the same pseudo-label, and excessive variation across the three geometric outputs for the same evaluator. Both checks use variances of normalized scores and thresholds; the text states that an augmented set is discarded when a candidate fails the consistency check. Passing candidates are merged with that input's historical bank, ranked by ensemble quality, and reduced to the historical Top-3. This is therefore not complete removal of filtering or permanent retention of every low-quality label: consistency rejection and Top-3 gating remain, while retained labels now carry explicit quality conditions into learning. Equation (2) is damaged in the extracted text, preventing exact recovery of its two-threshold inequalities; the prose does not provide threshold values, and this note does not invent them.

2. Quality Conditioning and Spatial Weighting: specify the target level and reliable regions separately

The student receives a degraded image and the quality score of its selected pseudo-label; training pairs the condition with its corresponding target instead of assigning the maximum score to every target. A sinusoidal encoding followed by an MLP maps the score to a vector matching the channel dimension at the injection point; the vector is broadcast spatially and added to intermediate features. The paper uses a U-Net encoder-decoder bottleneck as an example because its compact features allow a small amount of conditioning to influence the restoration process. This does not require a probabilistic restoration backbone or replace it with a diffusion model, allowing integration with existing models such as NAFNet, HistoFormer, and CORUN. Under this objective, a low-score pseudo-label is still reconstructed, but under its associated low-score mapping; the global score should not be mistaken for an automatic reduction in the image's loss weight.

Quality can also vary within one pseudo-label, for example when the sky contains little residual snow but building details retain stripes. The authors divide the label into \(2\times2\) non-overlapping blocks, score each block with the IQA ensemble, upsample the resulting weight map, and apply it to the pixel-level reconstruction loss. Higher-quality regions consequently provide stronger supervision, reducing pressure to copy local artifacts, while the global score continues to specify the output quality level. The readable portion of Equation (4) lists combination coefficients of 0.4, 0.4, and 1 for MUSIQ, reversed BRISQUE, and CLIP-IQA, and describes normalization to \([0,1]\). However, the implementation section uses a maximum condition of 7 without explaining the complete conversion between these scales; this note therefore neither equates raw IQA scores with injected conditions nor invents a conversion formula.

3. Quality-Guided Optimization: improve the high-score output rather than merely degrade the low-score one

Conditional reconstruction alone may let the model ignore the score or produce nearly identical outputs for different conditions. The authors inject high and low quality conditions for the same input, evaluate the two student outputs with NR-IQA, and derive an ordering signal without human preference annotations. This is inspired by Direct Preference Optimization (DPO), but is not standard DPO: a deterministic image regressor has no output log-probabilities to compare, so the objective operates directly in quality-score space. The first term, \(\mathcal{L}_1\), requires sufficient separation between high-condition and low-condition outputs, with margin \(\delta\) and preference sharpness controlled by temperature \(\beta\). The second term, \(\mathcal{L}_2\), anchors the high-condition output to the quality of the teacher's best pseudo-label, discouraging satisfaction of the margin solely by degrading the low-condition output. The third term, \(\mathcal{L}_{\mathrm{reg}}\), keeps score-injection weights near their initialization to limit disruption of the original restoration capability.

Direct IQA optimization can still produce high-scoring grid artifacts, motivating cropped quality consistency. Two paths use the same quarter-image location: restore the full image and then crop and score it, or crop the degraded input first and then restore and score it; both use the same quality condition. The objective reduces the difference between quality scores, rather than requiring identical pixels, because full-image and cropped restoration have different context. Following the textual definition of Equation (9) in Section 3.4, the constraint can be written as:

\[ \mathcal{L}_{\mathrm{cropped}}=\lVert S_1-S_2\rVert_1. \]

Here, \(S_1\) and \(S_2\) are ensemble quality scores from those two paths; this expression represents the difference explicitly described in the prose, not a reconstruction of the damaged preference-loss Equation (8). The authors argue that local statistical shortcuts and evaluator scale dependence can make the two processing orders disagree, so this constraint reduces reward hacking. It does not prove that all artifacts disappear; the failure analysis still reports occasional grids.

A Worked Example¶

Consider a real image containing snow and building textures; the following illustrates the stated pipeline rather than an additional experiment. The teacher restores three geometric augmentations, inversely aligns the outputs, and applies multi-evaluator scoring and consistency checks to the three candidates. Passing candidates are merged with historical labels, retaining at most three; the student uses the condition associated with its selected target and distributes pixel supervision according to the four block scores. For that same input, the student also produces high- and low-condition outputs: the high-condition branch must not only outperform the low one but also receive upward anchoring from the teacher's best target quality. The two crop-and-restore processing orders are then checked for quality consistency to discourage grid textures from exploiting an evaluator at one scale. At deployment, the test image is directly supplied with condition 7; its clean ground truth is not required, and the restored output need not first return to the label bank.

Loss & Training¶

Training combines spatially weighted pixel \(\ell_1\) reconstruction, a VGG-feature perceptual loss, the preference objective, and cropped quality consistency. Reconstruction compares the student output with the corresponding bank pseudo-label, while perceptual loss preserves higher-level structure and appearance; together they constrain image content and cannot be replaced by IQA ordering alone. The EMA coefficient is 0.998; the paper reports two NVIDIA A6000 GPUs, AdamW with momentum parameters \((0.9,0.999)\), a learning rate of \(5\times10^{-5}\), and 10K iterations. Because pseudo-labels scoring above 7 are scarce, the authors merge them into one high-quality space and set the maximum inference guidance score to 7. Continuous conditioning therefore does not imply experimentally established extrapolation to arbitrarily large scores; the practical control range is limited by training support and this merging rule. Equations (8) and (11) lose several operators in the cached extraction, so the complete preference expression and term weights cannot be transcribed reliably and require readable equations or implementation for confirmation. The paper provides a code URL but describes a future release; this note does not infer that the implementation is already available.

Key Experimental Results¶

Main Results¶

The following excerpts from Table 1, page 11, compare identical backbones against Colabator; they are not a universal ranking across all methods and metrics. Real test images lack clean references, so results primarily measure no-reference perceptual quality; lower FADE is better, while the other listed metrics are higher-is-better.

Task / Test Dataset	Backbone	Metric	Colabator	QualiTeacher	Difference
Desnowing / RealSnow 10K-Test	NAFNet	HyperIQA	0.470	0.559	+0.089
Desnowing / RealSnow 10K-Test	NAFNet	Q-Align	3.936	4.030	+0.094
Dehazing / RTTS	CORUN	FADE	0.824	0.590	-0.234
Dehazing / RTTS	CORUN	LIQE	2.946	3.317	+0.371
Low-light enhancement / DICM	SemiLL	MFRQA	58.138	59.372	+1.234
Underwater enhancement / Seathru	AIM-Net	URANKER	2.374	2.467	+0.093
Deraining / Real 3000	HistoFormer	TOPIQ	0.520	0.508	-0.012
Deraining / Real 3000	HistoFormer	LIQE	2.680	2.539	-0.141

Desnowing training uses paired Snow100K and unpaired RealSnow 10K-Train data; the low-light experiment follows SemiLL, training on VE-LOL and testing on DICM. The training evaluators are MUSIQ, BRISQUE, and CLIP-IQA, whereas the main table uses a different evaluator set and task-specific measures, helping avoid reporting only the scores directly optimized. The deraining prose claims superiority on every metric, but Table 1 shows lower TOPIQ and LIQE than same-backbone Colabator and identical MANIQA of 0.308; this note follows the tabulated values.

Ablation Study¶

Table 2, page 12, uses NAFNet for desnowing on RealSnow 10K-Test; components are added sequentially, and all listed metrics are higher-is-better. These are cumulative ablations rather than independent removals, so differences do not establish component contributions independent of interactions.

Config	Score Condition	Preference Optimization	Cropped Consistency	Dual-Drop	MANIQA	ARNIQA	HyperIQA	Q-Align
(a)	No	No	No	No	0.329	0.574	0.452	3.850
(b)	Yes	No	No	No	0.341	0.601	0.510	4.001
(c)	Yes	Yes	No	No	0.352	0.613	0.535	4.021
(d)	Yes	Yes	Yes	No	0.358	0.620	0.543	4.027
(e)	Yes	Yes	Yes	Yes	0.362	0.628	0.559	4.030

Key Findings¶

Score conditioning corresponds to (a) to (b), increasing HyperIQA from 0.452 to 0.510, a gain of 0.058 and the largest single sequential increase for this metric in the table.
Preference optimization raises MANIQA from 0.341 to 0.352, but the authors note accompanying grid artifacts in Figure 5; adding cropped consistency increases it to 0.358 and reportedly reduces visible artifacts.
The user study in Table 3, page 12, involves 15 participants providing blinded ratings on a 0–5 scale; NAFNet desnowing scores are 3.23 for Colabator and 3.58 for QualiTeacher, without confidence intervals in the text.
Table 5, page 14, feeds enhanced ExDark images into YOLOv13; Mean AP is 76.7 for CIDNet with Colabator and 79.1 with QualiTeacher, supporting benefits beyond image-quality scores alone.

Highlights & Insights¶

Quality conditioning distinguishes the quality of an available label from the desired quality of the final output. It makes an attribute of imperfect supervision explicit instead of treating every pseudo-label as ground truth.
Global conditioning and local weighting are complementary: one controls the target quality level, while the other handles reliability variation within a target, so they cannot be reduced to a single sample weight.
A preference margin plus upward anchoring is more targeted than enlarging the gap between outputs alone. A gap can be achieved by damaging the low-condition branch; upward anchoring also places improvement responsibility on the high-condition branch.
Cropped consistency suggests checking evaluator shortcuts across processing orders. It regularizes vulnerabilities in proxy scores rather than proving the absolute correctness of no-reference assessment.

Limitations & Future Work¶

Section 5 acknowledges occasional grid artifacts and attributes them to MUSIQ-KonIQ's photographic texture preferences; replacing it with TOPIQ removes grids in Figure 8, but this is evidence from the displayed cases.
Evaluator substitution involves trade-offs: the TOPIQ variant in Table 6 achieves ARNIQA 0.614, below the default 0.628 and Colabator's 0.620, so it is not superior on every metric.
From a reader's perspective, NR-IQA, human ratings, and detection gains support perceptual utility but do not establish true pixel fidelity; text, medical details, and similarly sensitive content require dedicated validation.
The conversion between normalized scores and condition 7, the two threshold values, and the damaged preference formula still require readable equations or implementation, preventing unambiguous reproduction from this text alone.
Lightweight feature injection is presented as plug-and-play, but training still requires multiple teacher passes, multiple IQA evaluations, and label-bank storage; without a complete cost comparison, lightweight inference should not be confused with inexpensive training.
Further work could independently test extrapolation beyond the trained score range and audit reward hacking with structural-fidelity measures excluded from training; these are reader suggestions, not completed experiments in the paper.

Compared with Mean Teacher: the EMA teacher loop remains; the changes concern student conditioning and supervision organization, not replacement by a separate giant teacher.
Compared with Colabator: both improve real-world restoration through label selection; QualiTeacher retains a bank while adding quality conditions and ordering constraints, so the distinction is not simply filtering versus no filtering.
Compared with standard DPO: no policy log-probability ratio is used; IQA score differences constrain deterministic outputs. The accurate description is DPO-inspired quality preference optimization, not direct application of the original DPO derivation.
Research implication: a measurable defect in supervision can become a condition instead of an immediate reason to delete a sample; transfer requires checking whether the defect score separates from semantic correctness, lest wrong answers become merely another controllable style.

Rating¶

Novelty: 4/5. Quality-conditioned pseudo-labeling and upward-anchored preferences for deterministic restoration form a clear contribution.
Experimental Thoroughness: 4/5. Five restoration tasks, multiple backbones, cumulative ablations, human ratings, and detection are covered, although gains are not universal across metrics.
Writing Quality: 3/5. The motivation is clear, but some universal-superiority statements conflict with tables, and formula extraction limits reproducibility from the available text.
Value: 4/5. Useful guidance for restoration training with real unlabeled data, subject to evaluator bias and calibration.