H-DDCM: Hadamard Codebooks for Fast Generative Image and Video Compression

H-DDCM. Hadamard Codebooks for Fast Generative Image and Video Compression. By Or Ronai and two others, from the Technion, Israel Institute of Technology. Published on the thirteenth of October. No code repository is linked to it.

There is a family of compression methods that need no training. They take the generative prior of an existing diffusion or flow model and use it to compress data, without retraining the model as a dedicated codec. Codebook-based approaches in this family, such as Turbo-DDCM, reach strong perceptual quality at ultra-low bitrates. Their weakness is the codebook itself. It is a dense Gaussian codebook, costly in memory and in computation, and it scales poorly to the large latent dimensions of modern image and video generators. On top of that, the combinatorial encoding of the bitstream becomes computationally heavy at that scale.

H-DDCM swaps the Gaussian codebooks for implicit, randomized Hadamard dictionaries. The word implicit matters. Thanks to the fast Walsh-Hadamard transform, a full-dimensional dictionary can be synthesized on demand, without ever materializing it. Projecting onto it and reconstructing from it cost on the order of d log d compute, where d is the dimension, and memory grows only linearly in d.

The second bottleneck, encoding, gets its own fix: an efficient lexicographic coding scheme that needs no lookup table, which makes combinatorial encoding practical at the larger dimensions.

The framework then extends beyond images. The authors adapt it to an autoregressive video flow model and build a streaming scheme around it. Periodically, the frames reconstructed most recently are re-encoded into a fresh KV-cache prefix. That keeps temporal context alive from one segment to the next, and it costs no additional transmitted bits. The result is long-form generative video compression running close to real time on a single GPU.

Across image and video benchmarks, H-DDCM achieves competitive performance on the combined measure of rate, distortion, and perception, while substantially reducing the memory and computational overhead that codebook-based, training-free generative compression usually carries. The claim, in other words, is not a jump in quality. It is holding quality steady while the cost comes down.

For the earlier codebook methods, quality was not where the trouble lay. Cost was, and it came from a dense codebook that had to exist in full. A structured dictionary that is synthesized when needed, rather than stored, is what lets the same idea stretch from images to long video on one GPU.