This February, JPEG AI International Standards was published after years of research aimed at generating smaller, transmittable, and preserved image codecs using machine learning techniques without compromising perceptual quality.
A comparison of peak signal-to-noise ratio (PSNR) and JPEG AI’s ML advanced approach from the official JPEG AI publication stream. Source: https://jpeg.org/jpeggai/documentation.html
One of the reasons why this Advent made most of the headlines is that the core PDF of this announcement was (ironically) not available on free access portals such as Arxiv. Nevertheless, Arxiv had already proposed many studies examining the importance of JPEG AI across several aspects, including the rare compression artifacts of the method and its importance to forensic medicine.
One study compared compression artifacts containing drafts of previous drafts of JPEG AI and found that new methods tend to blur text. This is not a minor issue when codecs contribute to the evidence chain. Source: https://arxiv.org/pdf/2411.06810
As JPEG AI modifies images in a way that mimics the artifacts of composite image generators, existing forensic tools make it difficult to distinguish real images from fake images.
After JPEG AI compression, according to a recent paper (March 2025), cutting-edge algorithms are no longer able to reliably separate authentic content from the manipulated areas of localization maps. The example of the source seen on the left is an operation/fake image, with the tampered area clearly depicted in standard forensic techniques (center image). However, JPEG AI compression provides a layer of reliability for fake images (the image on the far right). Source: https://arxiv.org/pdf/2412.03261
One reason is that JPEG AI is trained using a model architecture similar to that used in the generator system that forensic tools aim to detect.
The new paper shows the similarity between AI-driven image compression methodology and actual AI-generated images. Source: https://arxiv.org/pdf/2504.03191
Therefore, both models may generate several similar underlying visual properties from a forensic perspective.
Quantization
This crossover occurs because Quantizationwhich is common to both architectures and is used in machine learning as a way to convert continuous data into discrete data points and as an optimization technique that can significantly slim down the file size of trained models (being familiar with the wait between the release of cumbersome official models and the community-style quantized versions that can be run on local hardware).
In this context, quantization refers to the process of converting a continuous value of a latent representation of an image into a fixed discrete step. JPEG AI uses this process Reduce the amount of data you need Simplify internal numerical representations to save or send images.
Quantization makes encoding more efficient, but also imposes structural regularities that can resemble artifacts left by generative models.
According to this, the author of the “New Work of the Title” Three forensic queues for JPEG AI images We propose an interpretable non-neurotic technique for detecting JPEG AI compression. Determines whether the image has been recompressed. It is distinguished from the actual compressed image and the ones that were completely generated by AI.
method
Hue correlation
This paper proposes three “forensic cues” tailored to JPEG AI images. Color Channel Correlationwas introduced during the preprocessing procedure of JPEG AI. Measurable distortion of image quality The whole repetitive compression reveals recompression events. and Quantization patterns of latent space This helps to distinguish between images compressed by JPEG AI and images generated by AI models.
Regarding the hue correlation-based approach, JPEG AI’s preprocessing pipeline introduces statistical dependencies between image color channels and creates signatures that can act as forensic queues.
JPEG AI converts RGB images to YUV color space and performs 4:2:0 chroma subsampling. This includes downsampling of the chrominance channel before compression. This process leads to subtle correlations between the high frequency residuals of the red, green and blue channels. This is a correlation that is not present in uncompressed images and is of different intensity than that produced by traditional JPEG compression or composite image generators.
A comparison of how JPEG AI compression changes hue correlation in images.
Above, you can see a comparison of papers showing how JPEG AI compression changes the hue correlation of images and uses red channels as an example.
Panel A compares uncompressed images with JPEG AI compressed images and shows that compression significantly increases inter-channel correlation. Panel B separates the effects of JPEG AI preprocessing (color conversion and subsampling only). Panel C shows that traditional JPEG compression slightly increases the correlation, but not to the same extent. In Panel D, Midjourney-V5 and Adobe Firefly examine composite images that show moderate correlations, while others remain close to the compression level.
Rate distortion
The speed extension queue identifies recompression of JPEG AI by tracking image quality measured by peak signal-to-noise ratio (PSNR).
This study argues that by repeatedly compressing images with JPEG AI, loss of image quality gradually decreases and still measurable losses, as quantified by PSNR, and this gradual decomposition forms the basis of forensic cues to detect whether images have been recompressed.
Unlike traditional JPEG, where previous methods track changes in specific image blocks, JPEG AI requires a different approach due to its neural compression architecture. Therefore, the authors propose to monitor how both Bitrate and PSNR evolve in successive compression. Each round of compression changes fewer images than the previous one, and this reduced change (when plotted against Bitrate) can reveal whether the image has undergone a hypercompression stage.
An illustration of how repeated compression affects image quality of various codecs features the results of JPEG AI and neural codecs developed at https://arxiv.org/pdf/1802.01436. Both will steadily reduce the PSNR with every additional compression, even at low bitrates. In contrast, traditional JPEG compression maintains relatively stable quality across multiple compressions, except for high bitrates.
The image above shows the charted rate distortion curve for JPEG AI. The second AI-based codec. Traditional JPEGs find that JPEG AI and neural codecs show consistent PSNR reductions at all bitrates, while traditional JPEGs only show significant decomposition at much higher bitrates. This behavior provides a quantifiable signal that can be used to flag recompressed JPEG AI images.
The author similarly constructs signatures that help to flag whether or not an image is recompressed by extracting how bitrate and image quality evolve in multiple compression rounds, providing a potentially practical forensic queue in the context of JPEG AI.
Quantization
As we saw before, one of the more challenging forensic problems raised by JPEG AI is its visual similarity to the composite images generated by the diffusion model. Both systems use encoder-decoder architectures that process images in compressed latent space, leaving subtle upsampling artifacts.
These shared properties can confuse the detector – even those played on JPEG AI images. However, the key structural differences are as follows: JPEGAI applies quantization. This is a step of rounding the latent values to discrete levels for efficient compression, but generative models are not usually the case.
In a new paper, this distinction is used to design forensic cues that indirectly test the existence of quantization. This method analyzes how a potential representation of an image reacts to rounding, assuming that if an image is already quantized, its latent structure exhibits a measurable alignment pattern with a round value.
These patterns, although invisible, produce statistical differences that help separate the actual compressed image from the fully composite image.
Examples of mean Fourier spectra reveal that both JPEG AI compressed images and images generated by diffusion models such as Midjourney-V5 and stable diffusion XL exhibit normal grid-like patterns in the frequency domain. In contrast, the actual images do not have these patterns. This overlap of spectral structures helps explain why forensic tools often confuse actual compressed images with composite images.
Importantly, the authors show that this queue works in different generative models and remains effective even when it is strong enough to zero the entire section of latent space. In contrast, composite images show a very weak response to this rounding test, providing a practical way to distinguish between the two.
The results are intended as a lightweight, interpretable tool aimed at differences between compression and generation cores rather than relying on brittle surface artifacts.
Data and Testing
compression
To assess whether the hue correlation queues can reliably detect JPEG AI compression (i.e. the first pass from an uncompressed source), the authors used the JPEG AI reference implementation to compress these at various bitrates and tested them with high quality uncompressed images from the Raise dataset.
They trained a simple random forest on the statistical pattern of color channel correlations (particularly how residual noise in each channel matches the other channels) and compared this to a RESNET50 neural network that was directly trained on image pixels.
Detection accuracy for JPEG AI compression using hue correlation features compared at multiple bitrates. This method is most effective at lower bitrates where compression artifacts are stronger, and shows better generalizations for invisible compression levels than the baseline RESNET50 model.
RESNET50 achieved higher accuracy when the test data closely matched the training conditions, but struggled to generalize at different compression levels. The correlation-based approach is much simpler, but it turns out to be more consistent across bitrates, especially at lower compression rates, where preprocessing of JPEG AI has more powerful effects.
These results suggest that JPEG AI compression can be detected using interpretable and resilient statistical cues without deep learning.
Recompressed
Evaluate whether it is JPEG AI or not ReCompression can be reliably detected. Researchers tested rate-tolerant cues on a set of images compressed at various bitrates.
This method extracted 17-dimensional feature vectors to track how image bitrate and PSNR evolved over three compression paths. This feature set captured how much quality was lost at each step, and the potential speed and advanced rate behavior, that is, traditional pixel-based methods are not easily accessible.
Researchers trained random forests for these features and compared their performance to ResNet50, which trained them with image patches.
Results of classification accuracy for Random Forest trained on the rate distortion feature to detect whether JPEG AI images are recompressed. This method works best when initial compression is strong (i.e. low bitrate) and when it consistently outperforms pixel-based ResNet50.
Random forests have proven particularly effective when initial compression is strong (i.e. low bitrate), revealing clear differences between single and double compressed images. Like previous queues, ResNet50 iterations struggled to generalize, especially when tested at compression levels that were not seen during training.
In contrast, rate distortion features remained stable over a wide range of scenarios. In particular, queues work even when applied to another AI-based codec, suggesting that this approach generalizes beyond JPEG AI.
JPEG AI and composite images
In the final test round, the authors tested whether their quantization-based features could distinguish between JPEG AI compressed images from fully composite images produced by models such as Midjourney, Stable Diffusion, Dall-E 2, Glide, and Adobe Firefly.
To this end, the researchers used a subset of the Synthbuster dataset and mixed the actual photographs of the Raise database with images generated from various diffusion and GAN-based models.
Examples of synthetic images of Synthbuster generated using text prompts inspired by natural photographs of the Raise-1K dataset. Images are created with a variety of diffused models and have prompts designed to generate photorealistic content and textures rather than stylized or artistic renderings. Source: https://ieeexplore.ieee.org/document/10334046
Actual images were compressed using JPEG AI at several bitrate levels, and classification was raised as a two-way task: JPEG AI vs. a specific generator, or a specific bitrate vs. a stable diffusion XL.
Quantization features (correlations extracted from latent representations) were calculated from a fixed 256 × 256 region and fed to a random forest classifier. As a baseline, ResNet50 was trained with pixel patches of the same data.
Separate JPEG AI compressed images from composite images using the classification accuracy of random forests using quantization features.
In most conditions, the quantization-based approach was superior to the ResNet50 baseline, particularly at low bitrates with strong compression artifacts.
The author states:
‘The baseline ResNet50 is best suited for glide images with 66.1% accuracy, but otherwise it is generally more generalized than quantization features. Quantization features show good generalization across compressive strength and generator type.
“The importance of quantized coefficients to zero is demonstrated in the very respectable performance of the truncated (characteristic).
“However, quantization features using intermittent, perfect integers (vectors) still perform better. These results confirm that the quantization of zeros is an important clue to distinguishing AI-compressed images.
Nevertheless, it also shows that other factors contribute. The perfect vector accuracy for detecting JPEG AI is for all bitrates above 91.0%, and with stronger compression, higher accuracy occurs.
Projection of feature space using UMAP showed a clear separation between JPEG AI and composite images, with lower bitrates increasing the distance between classes. One consistent outlier was glide, whose images were clustered in different ways, with the lowest detection accuracy of the tested generators.
Two-dimensional UMAP visualization of JPEG AI compression and composite images based on quantization capabilities. The plot on the left shows that the lower JPEG AI bitrate creates a larger separation from the composite image. A proper plot, how images of different generators cluster clearly within functional space.
Finally, the authors evaluated how well the functionality is retained under typical post-processing, such as recompression and downsampling of JPEGs. Heavy processing reduced performance, but the decline was slower, suggesting that the approach retains some robustness even under degraded conditions.
Quantization Evaluation Robustness in post-processing, including JPEG recompression (JPG) and image resizing (RS), is characterized by robustness.
Conclusion
It is not guaranteed that JPEG AI will enjoy widespread adoption. For one thing, you have enough infrastructure debt at hand to impose friction. Any New codecs; and even “traditional” codecs with fine pedigree and broad consensus regarding its value, such as AV1, have struggled to remove the long-standing methods of incumbents.
Regarding potential system conflicts with AI generators, the current The generation of AI image detectors can be reduced to another type of trace, or ultimately replaced, in later systems, assuming that the AI generator always leaves the forensic residues and this is not certain.
This means that JPEG AI’s unique quantization properties may not collide with the forensic trails of the most effective new generation AI systems, along with other cues identified by new papers.
However, if JPEG AI continues to work as a in fact “AI Wash” greatly blurs the distinction between actual and generated images, making it difficult to create a compelling case for its ingestion.
First released on Tuesday, April 8th, 2025
