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

Or Ronai; Vladimir Kulikov; Tomer Michaeli

cs.CVOct 13, 2026

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

Authors: Or Ronai, Vladimir Kulikov, Tomer Michaeli

Organizations: Technion – Israel Institute of Technology

Abstract

Training-free diffusion-based compression methods repurpose the generative priors of a diffusion or a flow model to compress data without retraining them as task-specific codecs. Recent codebook-based approaches such as Turbo-DDCM achieve strong perceptual quality at ultra-low bitrates, but their reliance on dense Gaussian codebooks remains costly in both memory and computation, and scales poorly with the large latent dimensions of modern image and video generators. Their combinatorial bitstream encoding further becomes computationally heavy at this scale. We introduce H-DDCM, a scalable generative compression framework that replaces Gaussian codebooks with implicit randomized Hadamard dictionaries. The fast Walsh-Hadamard transform allows for implicitly synthesizing full-dimensional dictionaries with $O(d \log d)$ compute for projection and reconstruction and $O(d)$ memory, without materializing the full-dimensional dictionary. We further introduce an efficient table-free lexicographic coding scheme that makes combinatorial encoding practical at larger dimensions. Beyond images, we adapt the framework to an autoregressive video flow model and develop a streaming compression scheme that periodically re-encodes recently reconstructed frames into a fresh KV-cache prefix, preserving temporal context across segments at no additional transmitted bits. This enables long-form, close to real-time generative video compression on a single GPU. Across image and video benchmarks, H-DDCM achieves competitive rate-distortion-perception performance while substantially reducing the memory and computational overhead of codebook-based training-free generative compression.

Figures & tables

Panel of Figure 1 — Figure 1: Quantitative comparison of image compression methods. We compare H-DDCM with training-free diffusion-based and other compression methods in terms of rate, distortion, and perception. Distortion is measured by PSNR, perceptual quality by FID (presented in log-scale), and runtime by round-trip encoding and decoding time. GPU memory refers to the excess memory used by training-free methods for dictionary storage, while tables memory refers to disk storage required by Turbo-DDCM’s precomputed combinatorial tables. The first row corresponds to training-free methods utilizing the SD 2.1 backbone ( $512\times 512$ , $d=2^{14}$ ) and other compression methods. The second row corresponds to FLUX backbone ( $1024\times 1024$ , $d=2^{18}$ ). The dashed gray vertical line denotes the VAE encoder-decoder bound. H-DDCM achieves competitive results on the rate-distortion-perception tradeoff, all the while being the fastest training-free method, avoiding explicit codebook materialization. Error bars represent the standard error of the mean (SEM).

Figure 2: Qualitative comparison of training-free diffusion-based image compression methods (FLUX on DIV2K). H-DDCM is the fastest and preserves visual quality and fidelity to the original image, while avoiding the memory overhead of an explicit codebook.

Figure 3: Quantitative comparison of video compression methods. We compare H-DDCM OC (

B=8

) with training-free methods using the same CF backbone and other compression baselines in terms of rate, distortion and perception. Distortion is measured by PSNR, and perceptual quality is measured by DOVER and FVMD (presented in log-scale). Encoding and decoding FPS report compression and decompression throughput, and are compared among methods using the CF backbone. H-DDCM achieves competitive rate-distortion-perception performance among the CF-based training-free methods while providing the highest measured encoding and decoding throughput, without codebooks materialization. Error bars represent the standard error of the mean (SEM).

Figure 4: Qualitative comparison of video-compression methods. The example is taken from the UVG dataset at $832\times 480$ resolution. Our method preserves source content and temporal consistency while providing the highest measured throughput among the evaluated CF-based methods. The space-time (ST) slice, which shows the temporal evolution of the pixels marked by the dashed line, reveals temporal discontinuities at ZeroGVC’s I-frame boundaries.

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1: Additional quantitative comparison of image compression methods (SD2.1). We compare H-DDCM with training-free diffusion-based and other compression methods in terms of rate, distortion and perception. Distortion is measured by PSNR, perceptual quality is measured by FID (presented in log-scale), and runtime is measured by round-trip encoding and decoding time. GPU memory refers to the excess memory used by training-free methods for dictionary storage, while tables memory denotes disk storage required by Turbo-DDCM’s precomputed combinatorial tables. The first row corresponds to DIV2K dataset, and the second row corresponds to CLIC2020 dataset. Both datasets are resized and center-cropped to resolution of

512\times 512

, corresponding to

d=2^{14}

. The dashed gray vertical line corresponds to the encoder-decoder bound. H-DDCM achieves competitive results on the rate-distortion-perception tradeoff, all the while being the fastest training-free method, avoiding explicit codebook materialization. Error bars represent the standard error of the mean (SEM).

Panel of Figure 6 — Figure S2: Complementary quantitative comparison of Figs. 1 and S1 of image compression methods. We compare H-DDCM with training-free diffusion-based and other compression methods in terms of rate, distortion and perception. Distortion is measured by SSIM and LPIPS, runtime is measured by encoding (compression) time, decoding (decompression) time, and round-trip time (compression and decompression time). PSC is omitted due to its slow runtime. The first, second and third rows correspond to Kodak24, DIV2K and CLIC2020 datasets, respectively, evaluated on SD2.1 ( $512\times 512$ , corresponding to $d=2^{14}$ ). The fourth row corresponds to DIV2K dataset, evaluated on FLUX ( $1024\times 1024$ , corresponding to $d=2^{18}$ ). The dashed gray vertical line corresponds to the encoder-decoder bound. H-DDCM achieves competitive quality, at the same time of being the fastest training-free method, without materializing the dictionary. Error bars represent the standard error of the mean (SEM).

Figure S3: Qualitative comparison of training-free diffusion-based image compression methods (SD2.1). The images are taken from the Kodak24 dataset (resized and center-cropped to resolution of $512\times 512$ ).

Panel of Figure 8 — Figure S4: Qualitative comparison of training-free diffusion-based image compression methods (SD2.1). The images are taken from the DIV2K dataset (resized and center-cropped to resolution of $512\times 512$ ).

Panel of Figure 9 — Figure S5: Qualitative comparison of training-free diffusion-based image compression methods (SD2.1). The images are taken from the CLIC2020 dataset (resized and center-cropped to resolution of $512\times 512$ ).

Figure S6: Qualitative comparison of training-free diffusion-based image compression methods (FLUX). The images are taken from the DIV2K dataset (resized and center-cropped to resolution of $1024\times 1024$ ).

Figure S7: Quantitative comparison of video compression methods. We compare H-DDCM OC (

B=8

) with training-free methods using the same CF backbone, and other compression baselines in terms of rate, distortion and perception. Distortion is measured by PSNR, and perceptual quality is measured by DOVER and FVMD (presented in log-scale). Encoding and decoding FPS report compression and decompression throughput, respectively, and are compared among methods using the CF backbone for fair comparison. The first row corresponds to HEVC-B dataset, the second row corresponds to HEVC-C dataset, and the last row corresponds to HEVC-D dataset. H-DDCM achieves competitive rate-distortion-perception performance among the CF-based training-free methods while providing the highest measured encoding and decoding throughput, without codebooks materialization. Error bars represent the standard error of the mean (SEM).

Figure S8: Complementary quantitative comparison of Figs. 3 and S7 of video compression methods. We compare H-DDCM OC (

B=8

) with training-free methods using the same CF backbone and other compression baselines in terms of rate, distortion and perception. Distortion is measured by MS-SSIM, LPIPS and DISTS, perceptual quality is measured by FID and FVD (presented in log-scale). The first, second and third rows correspond to HEVC-B, HEVC-C and HEVC-D datasets, respectively. H-DDCM achieves competitive rate-distortion-perception performance among the CF-based training-free methods. Error bars represent the standard error of the mean (SEM).

Figure S9: Qualitative comparison of video-compression methods. The presented video is taken from MCL-JCV ( $832\times 480$ ) dataset. Our method produces comparable results on the rate-distortion-perception tradeoff, while being the fastest method utilizing the CF backbone. The space-time (ST) slice, which shows the temporal evolution of the pixels marked by the dashed line, reveals temporal discontinuities at ZeroGVC’s I-frame boundaries.

Figure S10: Qualitative comparison of video-compression methods. The presented video is taken from HEVC-C ( $832\times 480$ ) dataset. Our method produces comparable results on the rate-distortion-perception tradeoff, while being the fastest method utilizing the CF backbone. The space-time (ST) slice, which shows the temporal evolution of the pixels marked by the dashed line, reveals temporal discontinuities at ZeroGVC’s I-frame boundaries.

Figure S11: Qualitative comparison of video-compression methods. The presented video is taken from MCL-JCV ( $832\times 480$ ) dataset. Our method produces comparable results on the rate-distortion-perception tradeoff, while being the fastest method utilizing the CF backbone. The space-time (ST) slice, which shows the temporal evolution of the pixels marked by the dashed line, reveals temporal discontinuities at ZeroGVC’s I-frame boundaries.

Figure S12: Qualitative comparison of video-compression methods. The presented video is taken from HEVC-B ( $832\times 480$ ) dataset. Our method produces comparable results on the rate-distortion-perception tradeoff, while being the fastest method utilizing the CF backbone. The space-time (ST) slice, which shows the temporal evolution of the pixels marked by the dashed line, reveals temporal discontinuities at ZeroGVC’s I-frame boundaries.

Figure S13: Qualitative comparison of video-compression methods. The presented video is taken from UVG ( $832\times 480$ ) dataset. Our method produces comparable results on the rate-distortion-perception tradeoff, while being the fastest method utilizing the CF backbone. The space-time (ST) slice, which shows the temporal evolution of the pixels marked by the dashed line, reveals temporal discontinuities at ZeroGVC’s I-frame boundaries.

Figure S14: Quantitative comparison ablation between H-DDCM and H-DDCM OC (

B\in\{2,4,8\}

). The overcomplete dictionary variant provides better performance across all distortion and perceptual quality metrics. Error bars represent standard error of the mean (SEM).

Figure S15: Qualitative comparison ablation between H-DDCM and H-DDCM OC ( $B=8$ ). The presented video is taken from HEVC-B ( $832\times 480$ ) dataset, for $\text{BPP}=0.01$ . Both proposed methods achieve compelling results, however the perceptual quality and video consistency of the overcomplete dictionary is better.

Figure S16: Quantitative comparison ablation between H-DDCM OC utilizing the transpose variant, and H-DDCM OC utilizing the pseudo-inverse, both with

B=8

. Both variants achieve similar performance, and the cosine similarity between their selected noise vectors remains close to

1

, i.e . , the selected

\mathbf{z}_{t}^{\star}

is almost aligned. Pseudo-inverse is denoted by

(\cdot)^{\dagger}

. Error bars represent standard error of the mean (SEM).

Figure S17: Quantitative comparison ablation between H-DDCM OC (

B=8

), H-DDCM OC (

B=8

) without the re-encoded-tail mechanism for both hard cache reset – “w/o re-encoded-tail”, and without cache reset – “w/ continuous cache”, and H-DDCM OC (

B=8

) without the first timestep initialization mechanism. H-DDCM OC (

B=8

) achieves better with the re-encoded tail mechanism. Without the first timestep initialization mechanism provides better metrics, however as can be seen in Fig. S18 , it struggles to consistently align with the input video. Error bars represent standard error of the mean (SEM).

Figure S18: Qualitative comparison ablation between H-DDCM OC ( $B=8$ ), H-DDCM OC ( $B=8$ ) without the re-encoded-tail mechanism for both hard cache reset – “w/o re-encoded-tail”, and without cache reset – “w/ continuous cache”, and H-DDCM OC ( $B=8$ ) without the first timestep initialization mechanism. The presented video is taken from MCL-JCV ( $832\times 480$ ) dataset, for $\text{BPP}=0.01$ . While H-DDCM OC ( $B=8$ ) is aligned with the input video, without the re-encoded-tail mechanism when a new segment begins and a KV cache refresh occurs – there is a temporal discontinuity which can be observed in the spatiotemporal slice, and artifacts which can be observed in the second presented frame. For the continuous cache variant, i.e . , without cache reset, there are severe artifacts caused as this size of a cache/context is out-of-distribution for the model, which could be observed in the third presented frame. Additionally, without the first timestep initialization the method struggles to consistently align with the input video.

$K$	Mechanism	One-time setup (s)	Compress (s)	Decompress (s)	Total (s)
$2^{14}$	Table ( Vaisman et al., 2026 )	61.1	19.05	29.75	48.80
$2^{14}$	Recurrence (ours)	0	18.06	27.68	45.73
$2^{18}=d$	Table ( Vaisman et al., 2026 )	1386.8	37.09	38.14	75.23
$2^{18}=d$	Recurrence (ours)	0	33.81	34.69	68.50

Appendix

Table S1: Turbo-DDCM round-trip bitstream cost at

M=219

(FLUX, with

d=2^{18}

). The per-call cost is close between the two mechanisms; the one-time table build is what the original protocol pays fresh for every

(K,M)

it is run at, and which the recurrence never pays.

$M$	Recurrence rank (s)	Recurrence unrank (s)	Hockey-Stick rank (s)	Hockey-Stick unrank (s)	Rank ratio	Unrank ratio
43	0.034	0.049	0.0001	0.0008	0.004 $\times$	0.02 $\times$
67	0.033	0.050	0.0003	0.0022	0.008 $\times$	0.04 $\times$
84	0.039	0.055	0.0004	0.0035	0.01 $\times$	0.06 $\times$
128	0.041	0.060	0.0010	0.0091	0.02 $\times$	0.15 $\times$
219	0.051	0.071	0.0034	0.0320	0.07 $\times$	0.45 $\times$
373	0.063	0.083	0.0117	0.1153	0.19 $\times$	1.4 $\times$

Appendix

Table S2: Recurrence vs. hockey-stick, at

K=d=2^{18}

standalone rank/unrank cost, without model calls. Ratios are hockey-stick over recurrence.

Reference H-DDCM $M_{r}$	Overcomplete $M_{\mathrm{oc}}$	H-DDCM bits/slot	Overcomplete bits/slot
16	13	276	272
32	26	523	523
64	52	984	985
128	103	1,844	1,843
256	204	3,436	3,438
512	402	6,365	6,367

Appendix

Table S3: Overcomplete dictionary rate-matched configurations at

B=8

\tilde{d}=2^{19}

. The overcomplete dictionary always needs strictly fewer atoms than the reference it matches, since each of its indices costs more to name.

Walsh-Hadamard Transform

See all 11 papers

Jul 24, 2026

Generative Video Compression with Adaptive Score Distillation

Diffusion models provide strong generative capabilities for video compression at ultra-low bitrates. Existing diffusion-based video codecs adapt base models originally developed for text-conditioned generation, whereas diffusion models designed and trained specifically for compression remain unexplored. To fill this gap, we introduce our Generative Video Codec (GenVC), built on a video diffusion model trained from scratch for compression. To our knowledge, this is the first compression-oriented video diffusion model. We realize this model directly in pixel space with a global-to-local hierarchy that recovers fine spatio-temporal details, enabling high-quality generative reconstruction from compressed representations. To accelerate inference, we distill the multi-step model into one step using distribution matching distillation (DMD). Applying DMD directly, however, drives the student toward motion-stalled reconstructions. We trace this to a teacher-side guidance failure: once student-induced perturbations leave the frozen teacher's training region, its guidance can become misleading, causing DMD updates to reinforce rather than correct the student drift. To break the resulting feedback loop, we propose Adaptive Score Distillation, which gates DMD updates according to their alignment with the ground-truth direction, enabling high-quality reconstruction with coherent motion. Experimental results show that GenVC achieves state-of-the-art perceptual quality at ultra-low bitrates, with average bitrate savings of 62.5% at matched LPIPS and 71.3% at matched FID over GLVC. Unlike prior codecs that inherit billion-scale pretrained backbones, our diffusion model has only 478.0M parameters and decodes 1080p video in a single step at 15.1 fps on an A100 GPU.

Naifu Xue, Zhaoyang Jia, Haosen Li +7

Communication University of China, Beijing, China · Microsoft Research Asia · University of Science and Technology of China, Hefei, China +1

Neural Video Compression Diffusion Model Distillation Pixel-Space Diffusion Models Video Diffusion Models Generative Video Compression Score Distillation Sampling+1

Ask PDF

Jan 26, 2026

Advances in Diffusion-Based Generative Compression

Popularized by their strong image generation performance, diffusion and related methods for generative modeling have found widespread success in visual media applications. In particular, diffusion methods have enabled new approaches to data compression, where realistic reconstructions can be generated at extremely low bit-rates. This article provides a unifying review of recent diffusion-based methods for generative lossy compression, with a focus on image compression. These methods generally encode the source into an embedding and use a diffusion model to iteratively refine it during decoding, so that the reconstruction approximately follows the true data distribution. The embedding can take various forms and is typically transmitted via an auxiliary entropy model, and recent methods also explore the use of diffusion models themselves for information transmission via channel simulation. We review representative approaches through the lens of rate-distortion-perception theory, highlighting the role of common randomness and connections to inverse problems, and identify open challenges.

Yibo Yang, Stephan Mandt

Chan Zuckerberg Biohub · University of California Irvine

Learned Image Compression Generative Image Compression Diffusion Models Diffusion Model Compression

Ask PDF

Mar 8, 2026

Compression as Adaptation: Implicit Visual Representation with Diffusion Foundation Models

Modern visual generative models acquire rich visual knowledge through large-scale training, yet existing visual representations (such as pixels, latents, or tokens) remain external to the model and cannot directly exploit this knowledge for compact storage or reuse. In this work, we introduce a new visual representation framework that encodes a signal as a function, which is parametrized by low-rank adaptations attached to a frozen visual generative model. Such implicit representations of visual signals, e.g., an 81-frame video, can further be hashed into a single compact vector, achieving strong perceptual video compression at extremely low bitrates. Beyond basic compression, the functional nature of this representation enables inference-time scaling and control, allowing additional refinement on the compression performance. More broadly, as the implicit representations directly act as a function of the generation process, this suggests a unified framework bridging visual compression and generation.

Zongyu Guo, Jiajun He, Zhaoyang Jia +6

Microsoft Research Asia · University of Cambridge

Generative Video Compression Video Compression Implicit Neural Representations

Ask PDF

Abstract

Walsh-Hadamard Transform

Explore similar work

Generative Video Compression with Adaptive Score Distillation

Advances in Diffusion-Based Generative Compression

Compression as Adaptation: Implicit Visual Representation with Diffusion Foundation Models