Turbo-DDCM: Fast and Flexible Zero-Shot Diffusion-Based Image Compression
New method runs 'substantially faster' than state-of-the-art diffusion models, offering flexible region-based and quality-targeted compression.
A research team led by Amit Vaisman, Guy Ohayon, and Tomer Michaeli has introduced Turbo-DDCM, a breakthrough in diffusion-based image compression. Published and accepted for ICLR 2026, the method directly addresses the primary bottleneck of existing zero-shot diffusion compressors: their notorious slowness and high computational cost. Turbo-DDCM modifies the foundation of Denoising Diffusion Codebook Models (DDCMs), which compress images by sequentially selecting noise vectors from codebooks to guide a denoiser. The key innovation is efficiently combining a large number of these noise vectors at each denoising step, which significantly cuts down the total number of required operations, leading to substantially faster compression times while maintaining performance on par with the best existing techniques.
Beyond raw speed, Turbo-DDCM introduces crucial flexibility for real-world applications. The team developed two novel variants. The first is a priority-aware version that allows users to specify certain regions of an image (like a face or text) for higher-fidelity compression, intelligently allocating more bits to these areas. The second is a distortion-controlled variant that compresses based on a target Peak Signal-to-Noise Ratio (PSNR), a direct measure of reconstruction quality, rather than a target Bits Per Pixel (BPP). This lets users define the desired output quality and lets the algorithm determine the compression rate, a more intuitive workflow for many professionals. Comprehensive experiments position Turbo-DDCM not just as a research curiosity but as a compelling, practical framework for next-generation image compression.
- Builds on DDCMs but combines many noise vectors per step, 'significantly reducing' denoising operations for major speed gains.
- Introduces a priority-aware variant for user-specified regions and a distortion-controlled variant that compresses to a target PSNR instead of BPP.
- Maintains state-of-the-art reconstruction quality, making it a practical candidate for real-world use, with code publicly available.
Why It Matters
Makes advanced AI-powered image compression fast and flexible enough for practical use in media, design, and storage.