TLDR: Turbo-DDCM is a new zero-shot diffusion-based image compression method that significantly improves speed, achieving 1.5 seconds per image for compression and decompression, an order of magnitude faster than previous methods. It builds on Denoising Diffusion Codebook Models (DDCMs) by efficiently combining many noise vectors per step and introducing an improved encoding protocol. The method also offers flexible variants for priority-aware compression (focusing on user-specified regions) and distortion-controlled compression (targeting specific image quality instead of bitrate), making it a practical and high-performing solution.
Image compression is a crucial technology in our digital world, enabling faster sharing and storage of visual content. While traditional methods have long dominated this field, recent advancements in artificial intelligence, particularly diffusion models, are opening new avenues for more efficient and higher-quality compression.
A new research paper introduces Turbo-DDCM, an innovative method designed to significantly speed up zero-shot diffusion-based image compression. Zero-shot methods are particularly appealing because they don’t require specific training for compression; instead, they leverage pre-trained diffusion models that can be used for various tasks like image generation, restoration, and editing. This flexibility means that improvements to the core diffusion model benefit all these applications simultaneously.
However, a major hurdle for existing zero-shot diffusion compression techniques has been their slow processing times, often taking several minutes to compress and decompress a single image. This makes them less practical compared to faster, training-based alternatives. Turbo-DDCM addresses this challenge head-on, achieving a remarkable round-trip compression-decompression time of just 1.5 seconds per image, without relying on specialized hardware optimizations. This represents an order of magnitude speedup over the fastest existing zero-shot approaches.
Turbo-DDCM builds upon a prior method called Denoising Diffusion Codebook Models (DDCMs). DDCMs work by selecting specific “noise vectors” from reproducible random codebooks during the image denoising process, guiding the model to reconstruct the target image. The key innovation in Turbo-DDCM is its ability to efficiently combine a large number of these noise vectors at each denoising step. This drastically reduces the total number of denoising operations required, leading to the substantial speed improvements.
The researchers also introduced an improved encoding protocol. Traditional DDCM methods, especially when combining many noise vectors, suffered from redundancies in how the selected noise indices were stored. Turbo-DDCM’s new protocol eliminates these redundancies, making the compression even more efficient and reducing the bits-per-pixel (BPP) needed to represent an image.
Beyond speed, Turbo-DDCM offers enhanced flexibility with two notable variants:
Priority-Aware Compression
This variant allows users to specify regions of an image that are more important, such as a face in a portrait or text in a document. The system then allocates more bits to these “priority-aware” regions, ensuring higher reconstruction quality where it matters most, even at very low bitrates. This capability is particularly useful for applications like medical imaging or video conferencing.
Also Read:
- Condition Preference Optimization: A New Approach for Precise Image Generation Control
- GEWDiff: A Novel Approach to Hyperspectral Image Super-resolution
Distortion-Controlled Compression
Typically, compression methods aim for a target bitrate, but the actual visual distortion can vary significantly across different images at that fixed bitrate. Turbo-DDCM’s distortion-controlled variant allows users to specify a target PSNR (Peak Signal-to-Noise Ratio), a common measure of image quality, instead of a target BPP. This ensures a more consistent level of quality across different images, making the output more predictable.
The paper presents comprehensive experiments demonstrating that Turbo-DDCM not only achieves state-of-the-art runtime but also maintains competitive performance in terms of image quality and perceptual fidelity when compared to other advanced compression techniques, both zero-shot and training-based. It offers a compelling, practical, and flexible solution for image compression. For more in-depth technical details, you can read the full research paper available at arXiv:2511.06424.
