Degradation-Consistent Paired Training for Robust AI-Generated Image Detection

Xiaokun Yang; Yinghan Hou; Zongyou Yang

arxiv: 2604.10102 · v2 · pith:WIIXMMCXnew · submitted 2026-04-11 · 💻 cs.CV · cs.AI

Degradation-Consistent Paired Training for Robust AI-Generated Image Detection

Zongyou Yang , Yinghan Hou , Xiaokun Yang This is my paper

Pith reviewed 2026-05-10 16:12 UTC · model grok-4.3

classification 💻 cs.CV cs.AI

keywords AI-generated image detectiondegradation robustnesspaired trainingconsistency lossJPEG compressionimage corruptionsSynthbuster benchmark

0 comments

The pith

Degradation-Consistent Paired Training raises AI-generated image detector accuracy on corrupted inputs by 9.1 percentage points with only a 0.9 percent drop on clean images.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

AI-generated image detectors lose accuracy when inputs undergo everyday corruptions such as JPEG compression, Gaussian blur, or resolution downsampling. The paper introduces Degradation-Consistent Paired Training as an explicit training objective that pairs each clean image with a degraded version and forces the model to keep their features and predictions aligned. Two losses achieve this: a feature consistency term that minimizes cosine distance between representations and a prediction consistency term that minimizes symmetric KL divergence between output distributions. The resulting detector shows large gains on degraded test cases across nine generators while adding no parameters and no inference cost. Ablations indicate that these consistency constraints outperform attempts to improve robustness by expanding model architecture.

Core claim

By creating a clean view and a degraded view for every training image and enforcing both feature-level consistency through minimized cosine distance and prediction-level consistency through minimized symmetric KL divergence, the training process produces detectors whose accuracy under the eight degradation conditions rises by 9.1 percentage points on average while clean accuracy falls by only 0.9 percentage points, with the largest lifts observed under JPEG compression.

What carries the argument

Degradation-Consistent Paired Training (DCPT), which constructs clean-degraded image pairs and applies a cosine-distance feature consistency loss together with a symmetric-KL prediction consistency loss.

If this is right

Accuracy under JPEG compression improves by 15.7 to 17.9 percentage points relative to the identical baseline.
The method requires no extra model parameters and adds zero computation at inference time.
Training-objective changes prove more effective for robustness than adding new architectural components, which lead to overfitting on limited data.
Gains hold across all nine generators and all eight degradation conditions in the Synthbuster benchmark.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The paired-consistency idea could be applied to other detection tasks that suffer from distribution shift, such as video deepfake detection.
Future experiments could check whether the same losses remain effective when multiple degradations are combined or when degradation strength varies continuously.
If the robustness generalizes, practitioners could rely less on collecting separate corrupted datasets for every new corruption type.

Load-bearing premise

The degradations and consistency losses chosen for training will produce robustness that transfers to real-world corruptions never seen during training.

What would settle it

Testing the trained detector on AI-generated images that have undergone a new degradation type such as additive Gaussian noise or gamma correction not present in the training degradations and finding no accuracy gain over the non-paired baseline would falsify the claim.

Figures

Figures reproduced from arXiv: 2604.10102 by Xiaokun Yang, Yinghan Hou, Zongyou Yang.

**Figure 1.** Figure 1: Comparison of (a) our baseline training and (b) DCPT training under the frozen-backbone setting. Both use the same frozen DINOv2 ViT-B/14 backbone [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 5.** Figure 5: Per-degradation accuracy improvement of DCPT over baseline. JPEG [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

**Figure 4.** Figure 4: Clean vs. degraded accuracy comparison. DCPT nearly halves the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

read the original abstract

AI-generated image detectors suffer significant performance degradation under real-world image corruptions such as JPEG compression, Gaussian blur, and resolution downsampling. We observe that state-of-the-art methods, including B-Free, treat degradation robustness as a byproduct of data augmentation rather than an explicit training objective. In this work, we propose Degradation-Consistent Paired Training (DCPT), a simple yet effective training strategy that explicitly enforces robustness through paired consistency constraints. For each training image, we construct a clean view and a degraded view, then impose two constraints: a feature consistency loss that minimizes the cosine distance between clean and degraded representations, and a prediction consistency loss based on symmetric KL divergence that aligns output distributions across views. DCPT adds zero additional parameters and zero inference overhead. Experiments on the Synthbuster benchmark (9 generators, 8 degradation conditions) demonstrate that DCPT improves the degraded-condition average accuracy by 9.1 percentage points compared to an identical baseline without paired training, while sacrificing only 0.9% clean accuracy. The improvement is most pronounced under JPEG compression (+15.7% to +17.9%). Ablation further reveals that adding architectural components leads to overfitting on limited training data, confirming that training objective improvement is more effective than architectural augmentation for degradation robustness.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

DCPT's paired consistency training delivers a clear 9-point robustness lift on degraded images with little clean cost, but the baseline comparison leaves open whether the gains come from the losses or just seeing the degradations.

read the letter

DCPT's paired consistency approach improves AI-generated image detector robustness to common degradations like JPEG and blur, with a reported 9.1 point average gain on degraded test conditions and only a 0.9 point drop on clean accuracy. The paper introduces Degradation-Consistent Paired Training as an explicit objective using cosine distance on features and symmetric KL on predictions for each clean-degraded pair. This stands out from prior work that gets robustness indirectly through augmentation. It keeps the model unchanged at inference and shows results across nine generators and eight degradation types on the Synthbuster benchmark. The ablation on adding architecture components is helpful, as it indicates overfitting on the training data and reinforces that the training objective matters more. The empirical gains are the strongest part, especially the pronounced improvements under JPEG compression. The method is simple and directly targets the observed failure mode of detectors on corrupted images. One soft spot is the baseline setup. The comparison is to an identical baseline without paired training, but it is not clear whether that baseline also processes degraded images or trains exclusively on clean ones. If the latter, the 9.1 point lift includes the benefit of data augmentation, which weakens the claim that the consistency losses are the main driver. The abstract lacks details on statistical significance, exact training protocols, and hyperparameter choices, so the results need more scrutiny to confirm they are not sensitive to those factors. The assumption that the chosen degradations represent real-world conditions is reasonable but untested beyond the benchmark. This paper is for computer vision researchers working on detection of synthetic media who need practical robustness improvements. A reader looking for training strategies that avoid extra model complexity will get value from the experiments and the focus on the objective rather than architecture. It has enough substance and clear empirical support to merit peer review, even with the need for tighter controls on the baseline. I recommend sending it for review.

Referee Report

3 major / 3 minor

Summary. The manuscript proposes Degradation-Consistent Paired Training (DCPT) as a training strategy for AI-generated image detectors. For each training image, it creates a clean view and a degraded view (JPEG compression, Gaussian blur, resolution downsampling) and applies two consistency losses: feature-level cosine distance between representations and prediction-level symmetric KL divergence between output distributions. The method adds no parameters or inference cost. On the Synthbuster benchmark (9 generators, 8 degradation conditions), DCPT improves average degraded-condition accuracy by 9.1 percentage points over an identical baseline without paired training while dropping clean accuracy by 0.9 percentage points, with larger gains under JPEG (+15.7 to +17.9 pp). An ablation indicates that architectural additions cause overfitting on limited data, favoring objective-level improvements.

Significance. If the central comparison holds, DCPT offers a simple, zero-overhead way to explicitly optimize for degradation robustness rather than obtaining it incidentally from augmentation. The reported gains on a multi-generator, multi-degradation benchmark and the observation that architectural changes overfit while objective changes do not would be useful for practitioners deploying detectors in real-world conditions. The approach is parameter-free and directly falsifiable on the stated benchmark.

major comments (3)

[Abstract and §3] Abstract and §3 (method): The central claim attributes the 9.1 pp degraded-condition gain to the paired consistency losses (cosine distance + symmetric KL). However, the description of the 'identical baseline without paired training' does not specify whether that baseline is trained exclusively on clean images or also receives the same degraded views (without the consistency terms). If the baseline uses only clean images, the reported delta is confounded by data augmentation and does not isolate the effect of the proposed losses.
[§4] §4 (experiments): No details are provided on the number of independent training runs, random seeds, or statistical significance tests for the 9.1 pp, 0.9 pp, and per-degradation deltas. Without these, it is impossible to assess whether the improvements exceed run-to-run variance on the Synthbuster splits.
[§4] §4 (ablation): The statement that 'adding architectural components leads to overfitting on limited training data' is presented as supporting the superiority of objective-level changes, but the manuscript does not report the specific architectures tested, the size of the training set, or quantitative overfitting metrics (e.g., train vs. validation gaps). This weakens the ablation's ability to support the broader conclusion.

minor comments (3)

[Abstract] Abstract: The exact list of the 8 degradation conditions and the training dataset size/source should be stated explicitly rather than summarized.
[§2] §2 (related work): The positioning against B-Free and other methods would be clearer if the manuscript briefly restated how those methods incorporate (or fail to incorporate) explicit consistency objectives.
[§3] Notation: The precise formulation of the symmetric KL term and the weighting between the two consistency losses should be given as equations with hyperparameter symbols.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below and will revise the manuscript to improve clarity, reproducibility, and the strength of the ablation analysis.

read point-by-point responses

Referee: [Abstract and §3] Abstract and §3 (method): The central claim attributes the 9.1 pp degraded-condition gain to the paired consistency losses (cosine distance + symmetric KL). However, the description of the 'identical baseline without paired training' does not specify whether that baseline is trained exclusively on clean images or also receives the same degraded views (without the consistency terms). If the baseline uses only clean images, the reported delta is confounded by data augmentation and does not isolate the effect of the proposed losses.

Authors: We agree that the current description is ambiguous and could lead to misinterpretation. In the experiments, the baseline model is trained on the identical set of clean and degraded image pairs (i.e., the same data augmentation), but without the feature-level cosine consistency loss or the prediction-level symmetric KL loss. This design isolates the contribution of the paired consistency terms. We will revise the abstract and Section 3 to explicitly state the baseline training procedure, including that both models receive the same degraded views. revision: yes
Referee: [§4] §4 (experiments): No details are provided on the number of independent training runs, random seeds, or statistical significance tests for the 9.1 pp, 0.9 pp, and per-degradation deltas. Without these, it is impossible to assess whether the improvements exceed run-to-run variance on the Synthbuster splits.

Authors: We acknowledge the importance of reporting run-to-run variability and statistical significance for assessing the reliability of the reported gains. We will add details on the number of independent training runs (conducted with different random seeds), the specific seeds used, and the results of statistical significance tests (e.g., paired t-tests across runs) for the key deltas. These will be incorporated into Section 4 and the supplementary material. revision: yes
Referee: [§4] §4 (ablation): The statement that 'adding architectural components leads to overfitting on limited training data' is presented as supporting the superiority of objective-level changes, but the manuscript does not report the specific architectures tested, the size of the training set, or quantitative overfitting metrics (e.g., train vs. validation gaps). This weakens the ablation's ability to support the broader conclusion.

Authors: We agree that the ablation lacks the necessary specifics to robustly support the conclusion. We will expand this section to describe the exact architectural modifications tested (e.g., added convolutional layers or attention modules), report the training set size used in the experiments, and include quantitative overfitting indicators such as train-validation accuracy gaps. This will provide stronger evidence for the preference of objective-level improvements over architectural changes. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical validation of explicit consistency losses against a defined baseline

full rationale

The paper introduces DCPT as a training strategy with two explicitly defined consistency losses (feature cosine distance and symmetric KL on predictions) applied to paired clean/degraded views. The central result is an empirical accuracy delta on the Synthbuster benchmark (9.1 pp degraded average, 0.9 pp clean drop) versus an 'identical baseline without paired training.' No derivation chain, uniqueness theorem, ansatz, or fitted parameter is invoked that reduces the reported gains to the inputs by construction. The losses are new quantities, not renamings or self-citations; the comparison is presented as a controlled ablation rather than a prediction forced by data fitting. The method is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard loss functions and the domain assumption that consistency across degradations yields robustness; no free parameters or new entities are introduced.

axioms (2)

domain assumption Minimizing cosine distance between clean and degraded feature representations produces degradation-robust features
Invoked as the feature consistency loss in the proposed training strategy.
domain assumption Minimizing symmetric KL divergence between clean and degraded output distributions produces degradation-robust predictions
Invoked as the prediction consistency loss in the proposed training strategy.

pith-pipeline@v0.9.0 · 5526 in / 1440 out tokens · 56309 ms · 2026-05-10T16:12:16.091897+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

DCPT adds zero additional parameters and zero inference overhead... feature consistency loss that minimizes the cosine distance... prediction consistency loss based on symmetric KL divergence
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Experiments on the Synthbuster benchmark (9 generators, 8 degradation conditions)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.