Neural Compression Advantages Diminish as Image Size Increases
The touted benefits of neural network-based image compression methods virtually disappear when applied to real-world, high-resolution images. This surprising finding emerged from research by a team that included Jingtong Su, a CDS PhD student and visiting researcher at FAIR, and his colleagues at Meta AI and UCLA who examined the effectiveness of various compression techniques across image dimensions.
The team published their findings in the Proceedings of the Data Compression Conference (DCC), 2025. The paper, “Enhancing and Evaluating Probabilistic Circuits for High-Resolution Lossless Image Compression,” evaluates a middle-ground approach between computationally intensive neural networks and faster traditional codecs like WebP.
“Previous papers have tried to use low-resolution datasets like MNIST, which is just 28×28 grayscale pixels, or CIFAR with 32×32 RGB pixels,” Su explained. “But real-world images are often much larger — more like 1024×1024 — which makes a huge difference when comparing models on those low-resolution toy datasets versus more realistic ones.”
The research centered on optimizing probabilistic circuits (PCs), a model class that implements mixture models with built-in computational constraints. PCs allow for efficient marginalization over input data while requiring less computational power than neural network approaches.
The team introduced two key improvements to the PC approach: invertible transforms inspired by traditional compression methods and a local mutual information estimation technique that dramatically reduced computational complexity.
“We introduced a methodology called local mutual information estimation for computing the structure of probabilistic circuits,” said Su. “We reduced the originally squared time complexity to linear while maintaining exactly the same performance compared to the previous baseline.”
This optimization enabled the team to scale up testing to much larger images than typically used in compression research.
What they discovered challenges conventional wisdom in the field.
While neural network-based approaches demonstrate dramatic improvements over traditional compression techniques when tested on small images, these advantages largely vanish with real-world, high-resolution photographs. For example, on the CIFAR dataset with 32×32 pixel images, the popular IDF++ neural approach achieved 29.2% better compression than WebP, but this advantage dropped to just 5.8% with higher-resolution images from the CLIC dataset.
The findings raise important questions about the practical applications of neural compression techniques. Traditional codecs like WebP require significantly less computational power to run, don’t need training, and have a much smaller memory footprint — yet they perform competitively on larger images.
“We used to believe that neural network-based methodology would be superior compared to traditional codecs, but this conclusion no longer holds true when we scale the image size up to real-world dimensions,” Su noted. “At approximately 1000-pixel scale, we have already observed diminishing benefits of both probabilistic circuits and neural codecs compared to the originally standard ones.”
The research also demonstrated that probabilistic circuits can serve as an effective middle ground between traditional and neural approaches. With the team’s improvements, PCs approach the compression performance of neural networks while requiring substantially less computational power.
“These probabilistic circuit-based compression tools are gradually stepping into a phase to become a new baseline in the research community and potentially for practical use with real-world datasets,” Su said. “They are easy to train, extremely lightweight compared to neural networks, and can surpass the performance of standard codecs if we have a large enough dataset.”
This work represents one of the first comprehensive studies comparing compression performance across low and high-resolution images, with important implications for future research directions. By systematically testing across different image dimensions, Su and colleagues have identified critical limitations in current neural approaches that weren’t apparent in previous studies focused only on small images.
The findings highlight the need for compression research that specifically targets high-resolution imagery, where the patterns and relationships between pixels operate differently than in smaller test images. As media content continues to grow in size and quality, developing efficient compression techniques remains an important challenge for data scientists.
By Stephen Thomas
