Technion's Turbo-DDCM Compresses Images With Diffusion 34x Faster
A new zero-shot diffusion compressor hits 1.8 second round trips while matching state-of-the-art quality, without any custom hardware kernels.
Takeaways
- Turbo-DDCM cuts zero-shot diffusion image compression to 1.8 seconds round trip, no custom kernels required.
- Replaces DDCM's greedy matching pursuit with a closed-form thresholding solution, reducing diffusion steps by 92 percent.
- New lexicographic bit protocol removes factorial redundancy in atom-index encoding, saving about 40 percent of bits per step.
- Priority-aware variant supports per-pixel region-of-interest compression, a first for zero-shot diffusion codecs.
- Distortion-controlled variant targets a specific PSNR instead of BPP by regressing against JPEG file size.
- Accepted at ICLR 2026; builds on DDCM using Stable Diffusion 2.1 Base as the backbone.
Diffusion models keep pushing image compression to new lows in bitrate, but they carry a nasty catch: zero-shot approaches that reuse a pretrained diffusion backbone as a codec are painfully slow, taking anywhere from a handful of seconds with custom CUDA kernels to several minutes per image. A team from the Technion and the Flatiron Institute has released Turbo-DDCM, an ICLR 2026 paper that shrinks that round trip to about 1.8 seconds on a single A40 GPU while matching the rate-distortion-perception quality of the previous state of the art.
The trick has nothing to do with another custom kernel. The authors rethink how the compressor picks its noise vectors at each denoising step, back it with a cleaner bitstream protocol, and expose two new user-facing knobs: pick which regions get more bits, or pick a target PSNR instead of a target bitrate.
Why zero-shot diffusion codecs are stuck
Neural image compression has moved from purpose-trained autoencoders (BPG, HiFiC, ILLM) to diffusion models, either trained from scratch, fine-tuned, or used zero-shot. The zero-shot route is appealing because the same diffusion backbone can be reused across compression, restoration, editing, and generation, so any improvement to the backbone lifts every task.
The problem is speed. Existing zero-shot methods like PSC, DDCM, and DiffC need hundreds or thousands of denoising steps to reconstruct a single image. Even DiffC, which was recently accelerated with a hand-written CUDA kernel, still takes around 10 seconds per image and shows large variation in both bitrate and runtime across different inputs.
Turbo-DDCM builds directly on DDCM (Denoising Diffusion Codebook Models). DDCM replaces the standard Gaussian noise sampling in reverse diffusion with a selection from pre-defined codebooks of fixed iid Gaussian vectors, and it retains sample quality even for tiny codebooks. To compress an image, the encoder picks the codebook noise vector at each step that best pushes the denoiser toward the target image, and transmits only the indices. The decoder, sharing the same random seed, rebuilds the codebooks and replays the choices.