Sources for these images

The fixed no-JavaScript collage uses the following images. Every image was center-cropped to 3:2, resized to 720 × 480, metadata stripped, and encoded as WebP for the blog.

Park Bench by KSDIGITAL.eu, CC BY 2.0.
Brighton station visit002 by Elsie esq., CC BY 2.0.
Doree on keyboard by stickywikit, CC BY 2.0.
Erie Lackawanna Terminal Clock Tower in Hoboken, NJ by gargudojr, CC BY 2.0.
Cat and bear by celesteh, CC BY 2.0.
Green light by jglsongs, CC BY 2.0.
apple & knife by Muffet, CC BY 2.0.
Giraffe in Zoo by Hanumann, CC BY 2.0.
Driving while talking on his cell phone, tsk. by k.steudel, CC BY 2.0.
Untitled by Dripps, CC BY 2.0.
good thing my fridge is empty by dennis, CC BY 2.0.
Brown Teal duck. New Zealand by Bernard Spragg. NZ, CC0 1.0.
Canadian Geese by Jonas Salcedo, CC0 1.0.
Dsc 0451 Jpeg by Christianvelandia, CC0 1.0.
Generated samples ul-hR6Zn, aUWBhzt2, MrV2JwtI, xRONmFsk, jTSOYOdJ, UVTVe9WO, _aQqTIgK, EZmSRSl_, AGiAAxBk, zsf7bND2, 9c7qKW20, iI7HKKKU, and u9Hq0LCS from AIGI-Detection-Quality-Paradox by Coxy7, CC BY 4.0.

Percept-Lens: A Deep Dive into AI-Generated Image Detection

Across two complementary studies, we introduce Percept-Lens, an out-of-distribution evaluation framework, and a Mah-NCM-based detector built on frozen visual features. The framework exposes failures in released AI-generated image detectors, while our Mah-NCM-based detector recovers a strong real-versus-generated signal from a general-purpose frozen visual encoder.

Shashank Kotyan Makoto Shing Yuki Imajuku Rujikorn Charakorn Tarin Clanuwat

Sakana AI

Why detection matters

When an image loses its context

On 20 March 2023, Eliot Higgins posted a fictional sequence of Donald Trump's arrest made with Midjourney.¹ The original thread identified the images as generated, but as copies travelled beyond it, that context did not always travel with them. AFP traced reshared versions and found comments treating some of them as genuine.²

This was not an isolated incident, as only four days later another Midjourney image of Pope Francis in a white puffer coat went viral and fooled many viewers.³ In May 2023, an AI-generated image of an explosion near the Pentagon spread as breaking news before officials confirmed that no such incident had occurred.⁴ More recently in January 2025, synthetic images showing the Hollywood sign burning circulated during the Los Angeles wildfires even though the landmark remained intact.⁵ The cases differ in intent and harm, but share a failure mode that fabricated pixels moved faster than their provenance.

AI-generated fictional image of Donald Trump on the ground amid police officers — Select any image to view it at full size.

AI-generated image of Pope Francis wearing a large white puffer coat — Select any image to view it at full size.

As images move between platforms, provenance information can be lost. Context still determines where an image came from, what it means, and how it is being used. When that context is missing, a detector can contribute one clue by estimating whether the image appears real or generated.

The hard test begins outside familiar data

Many existing detectors perform strongly when their training and evaluation images share familiar generators, prompts, subject matter, and processing.⁶⁷⁸ Real-world deployment is far less controlled, as a new generator may introduce unfamiliar styles, images may come from a different domain, and a platform may resize or compress them before they reach the detector. The central question is whether the real-versus-generated distinction still holds when these changes arrive together. Here, transfer means remaining effective as generators, prompts, content, and processing change. The benchmark measures how well detectors do so in unfamiliar settings.

This work makes two complementary contributions. First, the Percept-Lens evaluation framework measures how released detectors perform under broad distribution shift. Second, the representation study proposes a new detector that combines a frozen visual encoder with a Mah-NCM decision rule.

Measuring transfer — Benchmark Study

A wider benchmark makes the gap visible

The Percept-Lens evaluation framework brings together more than 36 million images from 26 datasets in a common test of distribution shift. It covers changes in generators, prompts and styles, and image domains, allowing familiar-data performance to be compared with performance under much broader conditions.

To measure how existing detectors perform under this broad shift, we evaluate four released AI-generated image detectors. These include UnivFD⁹ with a CNNSpot¹⁰ training prior and three DRCT⁷ variants with GenImage⁸ or Stable Diffusion¹¹ priors. The transfer benchmark below compares each detector's familiar-data result with its broader out-of-distribution result.

Transfer benchmark

Accuracy degrades outside familiar distributions

Hover, focus, or tap an out-of-distribution point to see its change from the paired in-distribution accuracy.

Swipe horizontally to inspect all four detectors.

Static summary. All four released detectors show lower accuracy outside their familiar distributions, with drops ranging from 19.83 to 35.79 percentage points.

Paired transfer results. Each line connects one detector's reported in-distribution accuracy with its accuracy on the benchmark's out-of-distribution suite.

Exact values and accessible table

Transfer benchmark dataIn-distribution and out-of-distribution accuracy for four released AI-generated image detectors.
Detector · Training prior	ID accuracy	OOD accuracy	Accuracy change
UnivFD · CNNSpot	84.68%	48.89%	−35.79 pp
DRCT · GenImage	82.36%	62.53%	−19.83 pp
DRCT · SDv1	76.85%	57.00%	−19.85 pp
DRCT · SDv2	85.92%	59.85%	−26.07 pp

The benchmark study also points to a practical way to improve transfer. In controlled comparisons, detectors trained with diverse, open-ended web captions transferred better than those trained with object-centric captions, while simply adding generators yielded smaller gains. Prompt and image variety mattered alongside generator count. These comparisons show how training data shapes transfer, but not which detector components preserve a usable real-versus-generated distinction as distributions shift.

Diagnosing transfer — Representation Study

Follow the signal inside the detector

The representation study broadens the Percept-Lens evaluation framework to 39 public datasets and 7.1 million images. An AI-generated image detector has three components. The labeled reference set used to fit it is the support prior. The frozen encoder represents each image while its weights remain fixed, and the classification head turns those features into a decision. We hypothesize that a frozen visual encoder already captures the distinction between real and generated images, so we use a prior-conditioned Gaussian head to recover the decision boundary directly from its features. Rather than updating the encoder or training the head iteratively, our method fits the head from the class means and covariance of the labeled support features. This turns classical Gaussian rules into a closed-form detector that we test under broad out-of-distribution shift.

From images to a shared feature geometry

Our proposed detector first uses a frozen encoder φ to map an image x to a feature vector z = φ(x). It then fits a closed-form Gaussian head from labeled support images. The method illustrated here uses Mahalanobis-Nearest Centroid Matching (Mah-NCM), which estimates the real and generated class means μ_c and one pooled covariance Σ. The illustration projects this high-dimensional process into two dimensions for visualization. The study uses the full-dimensional feature vectors. In the interactive illustration below, we show how the method classifies an image. Choose and label one image as the query. The other five form the labeled support set used to fit the shared geometry.

Unlabeled visual sample 1 — **Fit, then classify.** Five labeled support images estimate two means and one shared pooled covariance, which define the ellipses and linear boundary. We then map the selected query into the same feature space, and the lower squared Mahalanobis score determines its class.

Unlabeled visual sample 2 — **Fit, then classify.** Five labeled support images estimate two means and one shared pooled covariance, which define the ellipses and linear boundary. We then map the selected query into the same feature space, and the lower squared Mahalanobis score determines its class.

Decision rule

Choose the nearest class in the shared feature geometry

prediction (z) = \underset{c \in {real, generated}}{\arg \min} {(z - μ_{c})}^{T} Σ^{†} (z - μ_{c})

Here, z is the query feature, μ_c is the estimated class mean, and Σ^† is the stable pseudoinverse of the regularized pooled covariance Σ. With uniform class priors, the query is assigned to the class with the lower squared Mahalanobis score. Unlike Euclidean nearest-centroid matching, this score accounts for feature scale and correlation, so displacement along a broad, noisy direction matters less than the same displacement across a narrow one. Sharing one covariance across classes preserves this geometry while keeping the decision boundary linear.

A ladder of simple decision rules

The Gaussian ladder compares closed-form rules on the same frozen features and labeled support data. Euclidean and cosine nearest-centroid rules use class means. Gaussian naive Bayes (GNB) adds per-dimension variances. Mah-NCM uses one full covariance shared by both classes, while quadratic discriminant analysis (QDA) fits a separate full covariance to each class. The study stabilizes these covariance estimates in high-dimensional feature spaces with diagonal loading and shrinkage.

Compare every ladder assumption in the table below.

Decision-rule comparisonGaussian discriminants used to study a frozen encoder feature space with fixed training prior.
Rule	What it tests	Covariance assumption	Boundary
Euc-/Cos-NCM	Whether class centroids alone separate the features	Shared isotropic	Linear
GNB	Whether per-dimension scales suffice without correlations	Class-specific diagonal	Non-linear
Mah-NCM	Whether one correlation-aware geometry fits both classes	Shared full	Linear
QDA	Whether the classes need different shapes and orientations	Class-specific full	Quadratic

Mah-NCM is strongest in 9 of 11 matched comparisons, while QDA leads in the two DRCT cases. The result distinguishes settings well described by one shared geometry from those that benefit from class-specific covariance.

Keep the features, change the decision

To compare our Mah-NCM-based detector with released methods, we replace each released classification head with Mah-NCM while keeping the corresponding public training data and frozen features fixed. We compute mean class accuracy within each evaluation dataset and then average all 39 datasets equally. In this like-for-like comparison, our Mah-NCM decision rule achieved higher mean class accuracy in 10 of 11 cases. The six detector families are UnivFD⁹, AIDE¹², Effort¹³, DRCT⁷, CoDE⁶, and Community Forensics¹⁴.

CF-384 was the only exception. Its released decision rule reached 87.54% mean class accuracy, compared with 84.55% for Mah-NCM under the same encoder and support prior. For this fixed representation and evaluation, the original decision rule performed better. Mah-NCM was stronger in the other 10 comparisons, while CF-384 shows that it is not universally optimal.

Matched comparison

Released head versus matched Mah-NCM

Hover or focus a Mah-NCM point to see its difference from the released head.

Headline result. Mah-NCM is higher in 10 of 11 matched comparisons. CF-384 is the one exception.

Swipe horizontally to inspect every matched pair.

Static summary. Mah-NCM has the higher mean class accuracy in 10 of 11 matched rows. CF-384 is the exception, with 87.54% for the released head versus 84.55% for Mah-NCM.

Matched head replacement. Each connector links the released head and Mah-NCM fitted on the same support prior and frozen encoder feature space. Horizontal position is mean class accuracy, averaged equally across the 39 Percept-Lens evaluation datasets.

View exact values

Matched comparison dataReleased detector head and matched Mah-NCM mean class accuracy.
Released detector	Support prior	Released head	Mah-NCM
UnivFD	CNNSpot	54.63%	57.62%
AIDE	CNNSpot	56.41%	62.76%
AIDE · GenImage-SDv1	GenImage-SDv1	48.99%	57.35%
DRCT	Full GenImage	65.77%	73.01%
AIDE · full GenImage	Full GenImage	54.98%	64.64%
DRCT	SDv1	63.36%	65.21%
DRCT	SDv2	62.37%	67.91%
Effort · GenImage-SDv1	GenImage-SDv1	72.58%	76.82%
CoDE-kNN · ELSA	ELSA	64.01%	66.15%
CF-224	CommunityForensics	81.98%	82.61%
CF-384	CommunityForensics	87.54%	84.55%

Representation and support data shape transfer

Having isolated the classification head, we next examine the two ingredients it depends on. The frozen encoder defines the feature space, while the labeled support prior is used to estimate the class distributions within it. The encoders in this sweep are general-purpose visual encoders that were not trained or fine-tuned specifically for real-versus-generated detection. Community Forensics¹⁴ and Mah-NCM stay fixed while only the encoder changes. PE-Core-bigG-14-448¹⁵ achieves the highest mean class accuracy among the tested encoders, 94.46%. We therefore use it as the fixed encoder for the complementary support-prior sweep. There, PE-Core and Mah-NCM stay fixed while only the labeled support data changes. The encoder candidates differ in architecture, training objective, data, and resolution. The sweep therefore compares complete encoder models rather than isolating any one design choice.

The model links below identify the exact checkpoints. The encoders come from the CLIP¹⁶, MAE¹⁷, BEiT¹⁸, SigLIP¹⁹, BLIP²⁰, BLIP-2²¹, DINOv2²², and DINOv3²³ families.

Frozen encoder sweepCommunity Forensics + Mah-NCM remain fixed, with no detector fine-tuning.
Frozen Encoder	Mean CA
ResNet-50	72.00%
ResNet-101	70.27%
ResNet-50x4	72.34%
ResNet-50x16	74.98%
ResNet-50x64	74.57%
ViT-B/16	74.18%
ViT-B/32	70.61%
ViT-L/14	76.96%
MAE-Huge	69.37%
BEiT-Large	61.78%
CLIP-XLM-RoBERTa-Large	77.67%
SigLIP-Large	64.14%
BLIP-Large	70.17%
BLIP2	77.73%
DINOv2-giant	73.30%
DINOv3-ViT-7b	88.91%
PE-Core-bigG-14-448	94.46%

Support prior sweepPE-Core-bigG-14-448 + Mah-NCM remain fixed.
Support Prior	Mean CA
CNNSpot	77.93%
GenImage	92.43%
DRCT	83.25%
ELSA	94.45%
Community Forensics	94.46%

How much labeled support is enough?

Our strongest tested detector combines PE-Core, Mah-NCM, and the Community Forensics support prior. The next question is how many labeled images are needed to estimate its class geometry. We use progressively smaller subsets of the Community Forensics prior while keeping the encoder and Mah-NCM fixed. With 219 labeled images (0.005% of the full labeled set), PE-Core + Mah-NCM reaches 90.89% mean class accuracy, compared with 94.46% with the full set.

Support sweep

Approaching the full-support result with fewer labels

Number of labeled images 219 images · 90.89% CA · reported min–max 89.55–91.62%

Headline result. 219 labeled images reach 90.89% mean class accuracy, 3.35 points above CF-384's 87.54% result. The full labeled set reaches 94.46%.

Swipe horizontally to inspect the full accuracy curve.

Static summary. PE-Core + Mah-NCM rises from 85.56% mean class accuracy with 44 labeled images to 90.89% with 219 and reaches 94.46% with the full labeled set. A grey dashed line marks CF-384's 87.54% result.

Strong performance with limited labeled data. Red circles show Community Forensics subsets, and the red dashed line marks 94.46% with all 4,386,820 labeled images. The grey dashed line marks CF-384's 87.54% result, the highest released-detector result among the 11 matched comparisons. CF-384 was fine-tuned end to end for generated-versus-real classification, whereas PE-Core is a general-purpose encoder kept frozen. Since the systems use different encoders and decision rules, the line provides context rather than an encoder-only comparison. Whiskers are reported minima and maxima, not confidence intervals. The x-axis is logarithmic.

View exact values

Support sweep dataMah-NCM mean class accuracy as the number of Community Forensics labeled images is reduced.
Support	Images	Mean CA	Reported min–max
0.001%	44	85.56%	83.81–87.40%
0.005%	219	90.89%	89.55–91.62%
0.01%	439	91.55%	90.86–92.22%
0.1%	4,387	93.18%	92.64–93.60%
1%	43,868	93.52%	93.35–93.80%
100%	4,386,820	94.46%	Not reported

Transfer in practice

Where released detectors lost the signal

The benchmark recorded the examples below as errors from six released detectors. Our Mah-NCM-based detector classifies the same images correctly, showing where its frozen representation retains a useful real-versus-generated distinction as generators, prompts, styles, and image domains change.

AI-generated zebra made with ADM, misclassified as real by Effort and correctly classified as AI-generated by our Mah-NCM-based detector — **These examples show where released detectors lose the signal while our Mah-NCM-based detector retains it.** The comparison spans changes in generator, prompt and style, and image domain. Select any image to view it at full size.

AI-generated zebra made with Midjourney, misclassified as real by DRCT trained on GenImage and correctly classified as AI-generated by our Mah-NCM-based detector — **These examples show where released detectors lose the signal while our Mah-NCM-based detector retains it.** The comparison spans changes in generator, prompt and style, and image domain. Select any image to view it at full size.

Application to current image generators

We also applied our Mah-NCM-based detector to newly generated photorealistic images from ChatGPT Images 2.0 and Nano Banana 2. The examples span varied subjects, styles, and image domains. The wider benchmark tests whether detector improvements continue to transfer as generators evolve. If detectors that perform well on it fail on future generators, the benchmark will need to expand.

AI-generated image of a botanist inspecting a leaf in a greenhouse, created with ChatGPT Images 2.0 — **Our Mah-NCM-based detector classified all six displayed images as AI-generated, while CF-384 classified them as real.** Three were created with ChatGPT Images 2.0 and three with Nano Banana 2. Select any image to inspect the full-size version.

AI-generated macro photograph of a mechanical wristwatch, created with ChatGPT Images 2.0 — **Our Mah-NCM-based detector classified all six displayed images as AI-generated, while CF-384 classified them as real.** Three were created with ChatGPT Images 2.0 and three with Nano Banana 2. Select any image to inspect the full-size version.

Takeaway

What to improve when transfer fails

Detection can contribute one signal to media verification alongside context and provenance. The detector estimates whether an image is real or generated, while context and provenance explain where it came from and how it is being used.

Together, the studies move from evaluating transfer to improving it. Percept-Lens reveals where released detectors fail, while the representation study introduces our Mah-NCM-based detector and shows whether the next improvement should target the data, visual encoder, or decision rule. In the tested settings, our strongest detector reached 94.46% mean class accuracy, showing that frozen visual features already separate real and generated images.

Resources and citations

Benchmark Study

@inproceedings{kotyan2025perceptlens,
  title     = {Percept-Lens: A Large-Scale Benchmark for Generalization in AI-Generated Image Detection with Diverse Images and Generative Models},
  author    = {Kotyan, Shashank and Imajuku, Yuki and Shing, Makoto and Clanuwat, Tarin},
  booktitle = {CVPR Workshop on Responsible Generative AI},
  year      = {2025}
}

Representation Study

@inproceedings{kotyan2026prior,
  title     = {Prior-Conditioned Gaussian Discriminants for Generalizable {AI}-generated Image Detection},
  author    = {Kotyan, Shashank and Shing, Makoto and Imajuku, Yuki and Charakorn, Rujikorn and Clanuwat, Tarin},
  booktitle = {European Conference on Computer Vision},
  year      = {2026}
}

Percept-Lens: A Deep Dive into AI-Generated Image Detection

Sources for these images

A simple decision rule recovers the signal in frozen visual features

When an image loses its context

Viral synthetic images with their context restored

Fictional Trump arrest

Pope in a puffer coat

Hollywood sign wildfire hoax

The hard test begins outside familiar data

A wider benchmark makes the gap visible

Accuracy degrades outside familiar distributions

Follow the signal inside the detector

From images to a shared feature geometry

How the Mah-NCM detector classifies an image

Choose the nearest class in the shared feature geometry

A ladder of simple decision rules

Keep the features, change the decision

Released head versus matched Mah-NCM

Representation and support data shape transfer

How much labeled support is enough?

Approaching the full-support result with fewer labels

Where released detectors lost the signal

Three sources of shift, seen through detector errors

Generator

Application to current image generators

Recent images from ChatGPT Images 2.0 and Nano Banana 2

Prompts and usage details

Usage details

What to improve when transfer fails

Resources and citations

Benchmark Study

Representation Study

Further Reading