acceptodds
Under review as a conference paper at ICLR 2027

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

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 compute for projection and reconstruction and 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.

Primary area:
generative models
Keywords:
image compression, video compression, real-time, diffusion models, flow models, training-free compression

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_AExBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes H-DDCM, a training-free generative image and video compression framework built upon Turbo-DDCM. The main idea is to replace dense Gaussian codebooks with implicit randomized Hadamard dictionaries, allowing efficient residual projection and reconstruction through the Fast Walsh-Hadamard Transform. The authors also propose a table-free lexicographic coding scheme to reduce the computational and memory costs of encoding selected codebook indices.

For video compression, the method is extended to a pretrained autoregressive video generation model, with a re-encoded-tail KV-cache refresh mechanism to maintain temporal context across video segments. Experiments on image and video datasets show that H-DDCM achieves reconstruction quality comparable to full-codebook Turbo-DDCM while substantially reducing memory requirements. The video implementation also achieves higher throughput than the evaluated training-free methods using the same generative backbone.

Critical strengths

I think the strongest contribution of this paper is that it addresses a real scalability problem in training-free generative compression. The memory cost of dense Gaussian codebooks becomes impractical with modern high-dimensional generative models. Replacing them with an implicit Hadamard dictionary is a simple but effective solution. In particular, the FLUX experiments show that H-DDCM can achieve reconstruction quality close to full-codebook Turbo-DDCM without materializing an extremely large dictionary. This is a meaningful improvement for the DDCM framework.

I also appreciate the attempt to extend the method to streaming video compression. The re-encoded-tail KV-cache refresh provides a practical way to preserve temporal context across segments without transmitting additional bits. The corresponding ablations are useful and show improvements over directly resetting the cache or allowing it to grow indefinitely.

Critical weaknesses

My first concern is the gap between the substantial reduction in theoretical complexity and the actual end-to-end efficiency. Replacing dense codebook operations with the Hadamard transform is clearly beneficial, but the improvement in overall runtime is much less impressive. On SD 2.1, both H-DDCM and full-codebook Turbo-DDCM require approximately 1.3 seconds for a 512×512 image, so there is almost no practical speed improvement. On FLUX, the round-trip time decreases from approximately 66 seconds to 44 seconds, which is meaningful, but still extremely slow for compressing and reconstructing a single 1024×1024 image. The situation is similar for video: the reported throughput is only around 10–16 FPS at 832×480 resolution on an NVIDIA H200 GPU. This is far from the real-time performance that modern video compression systems increasingly target.

I understand that training-free compression operates under different assumptions from trained codecs, and I do not expect H-DDCM to outperform all existing compression systems. However, recent works have made substantial progress toward practical high-speed compression, including CoD-Lite (Jia et al., ICML 2026) for generative image compression, GVC-RT (Dang et al., ACM MM 2026) for generative video compression, and strong neural video codecs such as DCVC-FM (Li et al., CVPR 2024) and DCVC-UF (Li et al., CVPR 2026). DCVC-UF is already included in the paper, but its runtime is not directly compared. I would like to see a clearer discussion of how H-DDCM fits into this rapidly developing landscape. Direct speed numbers from different hardware are not necessarily comparable, but the overall rate-distortion-perception versus computational cost trade-off is important for judging the practical significance of this work. My second and more important concern is the remaining potential of the DDCM framework itself. H-DDCM largely removes the explicit codebook memory bottleneck and achieves quality close to full-codebook Turbo-DDCM, but it does not substantially improve reconstruction quality over that baseline. Meanwhile, the overall computation is still dominated by the generative sampling process, and the method remains behind several trained compression approaches in distortion and practical speed. This makes me wonder whether the current results are approaching a practical performance ceiling of noise-codebook-based compression with frozen generative models. If the main remaining bottleneck is the repeated evaluation of a large generator, further optimizing the dictionary may provide diminishing returns. The video results also do not show a consistent advantage across all quality metrics, making the long-term competitiveness of this approach less clear.

I am not suggesting that the DDCM paradigm has necessarily reached its limit. There may still be opportunities in better noise selection, trajectory optimization, bitrate allocation, or reducing generative model evaluations. However, the paper does not provide much insight into these possibilities. Since the work mainly improves the implementation efficiency of an existing compression framework, I think understanding whether that framework still has substantial room for improvement is particularly important to assessing the significance of this contribution.

Questions for authors
  1. Practical efficiency and comparison with recent codecs

The proposed Hadamard dictionary substantially reduces codebook computation and memory, but why does this lead to almost no end-to-end speed improvement on SD 2.1? Could the authors provide a breakdown of the time spent on generative model inference, residual matching, codebook operations, and bitstream coding?

Also, how should we understand the practical efficiency of H-DDCM relative to recent fast compression methods such as CoD-Lite (Jia et al., ICML 2026), GVC-RT (Dang et al., ACM MM 2026), and strong DCVC models such as DCVC-FM and DCVC-UF? I am not necessarily requesting direct comparisons across incompatible hardware or training settings, but a clearer analysis of the quality–speed trade-off would be helpful.

  1. Fairness of the DCVC-UF comparison

Could the authors clarify the exact DCVC-UF configuration used in the experiments, including the model variant, intra period, number of coded frames, and color-space conversion? Since DCVC-UF is already included as a strong trained baseline, reporting or discussing its encoding and decoding speed under comparable conditions would help readers better understand the practical trade-offs. It would also be useful to explain how the different operating assumptions and compression objectives affect the interpretation of the perceptual and distortion results.

  1. Remaining potential of the DDCM paradigm

This is my most important question, and the answer would significantly affect my assessment.

H-DDCM has largely addressed the codebook scalability problem, but the overall compression performance remains limited by the expensive generative process. Do the authors believe that noise-codebook-based compression with frozen generative models is approaching its practical rate-distortion-perception and efficiency limits?

If not, what are the main remaining bottlenecks, and what concrete improvements could lead to substantial further gains beyond replacing or accelerating the codebook? For example, could optimizing the full sampling trajectory or reducing the number of generative model evaluations significantly improve the current results?

I would particularly appreciate an answer grounded in the behavior observed in this paper, rather than only general suggestions for future work. A convincing explanation of the remaining potential of this approach would make me considerably more positive about the paper.

I am currently leaning toward a rating of 6/10. However, since the review system uses a four-point scale, this would correspond to approximately 2.4/4, so I am selecting 2 for now. My assessment is still somewhat borderline, and I am open to raising my rating if the authors can convincingly address my main concerns during the rebuttal.

7 comments

Practical efficiency and comparison with recent codecs (W1/Q1):

In the tables below we report a per-step runtime breakdown of model inference, codebook operations and bitstream coding, for H-DDCM and Turbo-DDCM, on SD2.1 and FLUX at a fixed bitrate ($\sim$0.05 BPP, averaging over 5 images each). We emphasize that both methods use our bitstream implementation, so Turbo-DDCM uses our efficient implementation without its combinatorial tables bitstream implementation.

SD2.1 (Kodak24, 512x512, d = 16384):

MethodModel call (ms)Codebook compress (ms)Codebook decompress (ms)Bitstream compress (ms)Bitstream decompress (ms)Compress (s)Decompress (s)Round trip (s)
Turbo-DDCM (K=dK=d) w/ our coding27.23.012.232.193.500.540.901.44
H-DDCM27.21.510.802.223.520.500.881.38

FLUX (DIV2K, 1024x1024, d = 262144)

MethodModel call (ms)Codebook compress (ms)Codebook decompress (ms)Bitstream compress (ms)Bitstream decompress (ms)Compress (s)Decompress (s)Round trip (s)
Turbo-DDCM (K=dK=d) w/ our coding921837.5324.730.952.332.833.366.1
Turbo-DDCM (K=214K=2^{14}) w/ our coding92146.332.92.33.517.627.645.2
H-DDCM9211.950.9931.152.316.827.143.8

The gain is small on SD2.1 compared to FLUX, as the round trip is dominated by the diffusion model: on SD2.1 the network calls of the decoder alone take $\sim$0.8s of the $\sim$1.4s. The dense codebook of Turbo-DDCM costs O(dk)O(dk) per step, against O(dlog⁡d)O(d \log d) for our fast Walsh-Hadamard transform. At the small SD2.1 latent (d=214d=2^{14}) the dense codebook is only $\sim$0.5GB, and one step costs $\sim$3ms against a 27ms UNet call - the codebook accounts for just 0.09s of Turbo-DDCM's 1.44s round trip, so removing it can save at most a few percent. Therefore, when the latent dimension is small, H-DDCM's runtime advantage over Turbo-DDCM is small, and its benefit is mainly memory (dense buffer of 0.54GB vs. less than 1MB) and scalability. The dense cost grows quadratically with dd: from SD2.1 to FLUX, the Turbo-DDCM codebook step grows ∼280×\sim 280\times (3ms →\rightarrow 838ms), while the model call grows only ∼34×\sim 34\times (27ms →\rightarrow 921ms) and ours stays at $\sim$2ms. On FLUX the codebook is then a half of a Turbo-DDCM compress step and a third of its round trip (22s of 66s), whereas H-DDCM implicit codebook operations are less than 1% of the model call, resulting in 0.06s, which explains 66s for Turbo-DDCM, and 44s for H-DDCM. The advantage keeps growing with the latent size (higher resolutions, video).

We have added in App. C.1 Fig. S3 which demonstrates the runtime breakdown efficiency comparison between H-DDCM and Turbo-DDCM codebook.

Comparison with recent codecs: trained methods, such as DCVC-FM and DCVC-UF, are commonly distortion-oriented. Most trained codecs emphasize distortion-oriented optimization, although not all do: GVC-RT, for example, explicitly targets perceptual quality, however it suffers from visible artifacts, which can be observed in Fig. 4 and in our anonymous website. Training-free methods such as H-DDCM offer a complementary quality-speed trade-off by reusing a pretrained generative prior, which provides high perceptual quality, without codec-specific training.

We have evaluated CoD-Lite for comparison with H-DDCM. The following table reports the obtained metrics (PSNR, FID and round-trip time) for 2 illustrative BPP values, approximately 0.015 and 0.035, on the Kodak24 dataset. The results demonstrate that H-DDCM achieves comparable performance, without any architecture fine-tuning or training. We have added CoD-Lite comparison to the paper (see Figs. 1, S1 and S2).

MethodBPPPSNR ↑FID ↓Round-trip time (s) ↓
CoD-Lite0.015720.7550.570.035
H-DDCM0.014820.5349.821.141
CoD-Lite0.031421.9843.640.035
H-DDCM0.034722.6640.231.422

GVC-RT is concurrent work (first released on arXiv in August 2026), which we additionally evaluated on all 5 video datasets. DCVC-UF was already evaluated in our submission. We added GVC-RT to the quantitative and qualitative comparisons (Figs. 3, 4, and S8-S14). The qualitative comparison in Fig. 4 and the videos on our anonymous website reveal grid-like artifacts in GVC-RT reconstructions. To complement these visual observations, we additionally measure source-motion-guided L2L_2 warping error following [1-3] - optical flow is estimated from the source video, used to warp adjacent reconstructed frames, and their L2L_2 discrepancy is measured after rejecting pixels that fail forward-backward flow-consistency checks. This measures consistency with source-estimated motion. At approximately 0.0056 BPP on MCL-JCV, GVC-RT achieves higher PSNR and better DOVER/FVMD than H-DDCM but exhibits worse spatio-temporal jitter artifacts, as evident by the fact that H-DDCM has lower source-flow warping error (1.6×10−31.6\times 10^{-3} versus 1.96×10−31.96\times 10^{-3}). DCVC-UF has lower warping error (0.83×10−30.83\times 10^{-3}), but at the cost of far worse perceptual quality compared to H-DDCM, as can be observed in the qualitative figures. GVC-RT also offers higher throughput compared to our method (more than 100 FPS). These results illustrate distinct quality-speed trade-offs rather than uniform superiority of one method.

MethodBPPPSNR ↑DOVER ↑FVMD ↓Warping Error (×10⁻³) ↓
GVC-RT0.006323.470.608244.791.96
DCVC-UF0.005626.440.595515.740.83
H-DDCM0.005620.940.604419.671.6

Indeed there is a cost in adapting diffusion generative models for downstream tasks due to their iterative sampling process. However, they provide a valuable prior on how the data is distributed which is exactly the point that allows us to achieve high perceptual quality. Many trained video codecs prioritize distortion, although there are some perception-oriented trained codecs such as GVC-RT. Refer to our anonymous website for comparisons between video compression methods, which show that indeed training-free methods, especially H-DDCM, can achieve high-perceptual quality results in this regime.

In contrast to task-specific trained codecs, H-DDCM can be adapted to different distributions or modalities when a compatible pretrained generator is available, without codec-specific retraining. Trained codecs typically require redesign and retraining for such changes. Additionally, trained methods could be trained on a specific dataset distribution, and adapting them to other distributions requires retraining, whereas a compatible pretrained generative model can provide the prior needed for training-free adaptation.

We have included a dedicated paragraph regarding the discussion between training-free and trained methods in the revised manuscript (see Sec. 6).

DCVC-UF configuration (Q2):

We use Microsoft's official pretrained DCVC-UF HT-S model with its corresponding intra-frame checkpoint. The configuration uses one initial I-frame (intra_period=-1), followed by 8-frame inter chunks, skip_thres=0.15, and feature-memory reset interval 128. Intra/inter QPs are equal and selected per sequence to approach the target physical bitrates using actual entropy-coded stream sizes. These settings follow the official example configuration. Inputs use 8-bit YUV420 via FFmpeg, and all methods are evaluated with the common RGB-based metric pipeline.

DCVC-UF targets rate-distortion efficiency and high throughput, whereas H-DDCM uses a generative prior to achieve high perceptual quality at very low rates. We emphasize that H-DDCM is a training-free method, which inherently follows the generative backbones bottleneck, while other trained methods such as DCVC-UF achieve higher throughput, and target low distortion at the cost of perceptual quality. DCVC-UF achieves encoding and decoding throughput of approximately a thousand FPS.

Remaining potential of the DDCM framework (W2/Q3):

We view H-DDCM as removing one bottleneck - the computation ceiling timing is the model calls, and H-DDCM almost reaches it. We believe that in terms of rate-distortion-perception for noise-codebook compression the DDCM framework can still be improved. The H-DDCM framework opens room for methods that would previously have been impractical at large latent dimensions. In particular, H-DDCM tracks the quality of complete-codebook Turbo-DDCM on the image compression task, and outperforms it on the video compression task, while avoiding its dense dictionary operations. This suggests that the choice of dictionary representations can be further improved to achieve better rate-distortion-perception tradeoff under the evaluated settings. The H-DDCM framework does not establish an intrinsic ceiling for the broader DDCM framework.

  • Faster reconstruction trajectories and fewer generator evaluations - the remaining runtime is largely governed by pretrained generator inference, particularly at smaller latent sizes. A direct path is to use fewer sampling steps or faster few-step generators, as our autoregressive video experiments already demonstrate. Another promising approach is to exploit inversion-guided noise selection (e.g., ODE inversion of the source) to make fewer step trajectories more informative; efficient inversion and sampling methods, including FireFlow [4], provide concrete algorithmic tools to explore this in compatible flow backbones.

  • Better use of the freed compute and memory for rate-distortion-perception optimization - our current selection is greedy and local to each step. With implicit O(d)O(d)-memory dictionaries and fast O(dlog⁡d)O(d\log d) transforms, the encoder can afford more sophisticated search, e.g. optimizing noise choices jointly across a sampling trajectory, adapting bit budgets across timesteps and video segments, and potentially online source-conditioned optimization. These strategies target the quality ceiling of the current selection procedure under the H-DDCM framework, rather than the already-reduced dictionary cost. They may improve quality at fixed bitrate or reduce the bits needed for a target quality. These are concrete algorithmic opportunities enabled by removing the dictionary bottleneck, and are interesting future directions.

The same scalability also broadens the range of usable pretrained priors. Larger latent dimensions, new modalities, and pixel-space generative models such as JiT [5, 6] would make dense-codebook compression particularly costly, whereas H-DDCM's implicit dictionary makes adaptation more feasible without training a new codec. Thus, the novelty is not merely a constant-factor speedup - it removes a scaling obstacle and creates computational headroom for trajectory optimization, smarter rate allocation, and faster generator/inversion integration. We view these as concrete opportunities enabled by H-DDCM, while distinguishing them from the improvements already demonstrated here.

[1] Lai, Wei-Sheng, et al. "Learning blind video temporal consistency." (ECCV’18)

[2] Ceylan, Duygu, Chun-Hao P. Huang, and Niloy J. Mitra. "Pix2video: Video editing using image diffusion." (ICCV’23)

[3] Geyer, Michal, et al. "Tokenflow: Consistent diffusion features for consistent video editing." (ICLR’24)

[4] Deng, Yingying, et al. "Fireflow: Fast inversion of rectified flow for image semantic editing." (ICML’25)

[5] Li, Tianhong, and Kaiming He. "Back to basics: Let denoising generative models denoise." (CVPR’26)

[6] Vallaeys, Théophane, Jakob Verbeek, and Matthieu Cord. "Pixel-space video generation with U-JiT." (arXiv 11/26)

Reviewer_hdj2OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents H-DDCM, a training-free generative compression framework that replaces dense Gaussian codebooks with implicit randomized Hadamard dictionaries to reduce memory and computation. It also introduces efficient combinatorial coding and a KV-cache refresh scheme for streaming video compression. Experiments on image and video benchmarks report competitive rate–distortion–perception performance with reduced computational overhead.

Critical strengths

The paper is clearly presented, with a well-motivated explanation of the memory and computational costs of dense Gaussian codebooks. The proposed Hadamard dictionaries and coding scheme are introduced in a coherent manner, making the main technical ideas easy to follow.

Critical weaknesses
  1. The improvement in overall coding efficiency is relatively limited. Although the Hadamard dictionary significantly reduces codebook computation and memory, H-DDCM achieves approximately the same round-trip time as Turbo-DDCM on SD 2.1 (1.3 seconds), while reducing the time from 66 to 44 seconds on FLUX. The paper also lacks a detailed runtime breakdown of model inference, codebook operations, and bitstream coding. Such an analysis would help clarify the practical benefits of the proposed optimizations.

  2. The comparison with recent efficient generative codecs is not sufficiently comprehensive. Recent methods such as EF-LIC [1], DiT-IC [2], CoD-Lite [3], and OneDC [4] are not included in the evaluation. These methods report strong compression performance with substantially lower coding latency. Although H-DDCM focuses on training-free compression, a comparison with these methods would help establish its practical advantages and clarify the tradeoff between avoiding codec training and coding efficiency.

[1] H. Cao et al., “Efficient Learned Image Compression without Entropy Coding,” ICML 2026.

[2] J. Shi et al., “DiT-IC: Aligned Diffusion Transformer for Efficient Image Compression,” CVPR 2026.

[3] Z. Jia et al., “CoD-Lite: Real-Time Diffusion-Based Generative Image Compression,” ICML 2026.

[4] N. Xue et al., “One-Step Diffusion-Based Image Compression with Semantic Distillation,” NeurIPS 2025.

Questions for authors
  1. We would like the authors to provide a detailed runtime breakdown of model inference, codebook operations, and bitstream coding. In particular, we hope to better understand why the substantial reduction in codebook complexity leads to limited improvement in overall coding time, especially on SD 2.1.

  2. We would like the authors to provide comparisons with recent GIC methods, including EF-LIC [1], DiT-IC [2], CoD-Lite [3], and OneDC [4], in terms of rate-distortion-perception performance and coding latency under comparable evaluation settings. These results would help clarify the practical benefits of training-free compression relative to modern GIC methods.

5 comments

Our contribution is to make training-free compression more computationally and memory efficient, specifically under the DDCM framework, while retaining the flexibility of a pretrained generative prior. Specialized trained codecs are generally faster, and many optimize distortion-oriented objectives, e.g. DCVC variants. Our method instead targets scalable training-free compression. Crucially, H-DDCM can be adapted to new data distributions, generative backbones, tasks, and modalities when a suitable pretrained generator is available, without retraining a dedicated codec. In contrast, trained codecs typically require architecture redesign and retraining to support substantially different distributions, tasks, or modalities. We demonstrate this flexibility across image and autoregressive video compression; multimodal audio-video compression with a model such as LTX-2 might be a possible extension.

Coding efficiency (W1/Q1):

In the following tables we report a per-step runtime breakdown of model inference, codebook operations and bitstream coding, for H-DDCM and Turbo-DDCM, on SD2.1 and FLUX at a fixed bitrate ($\sim$0.05 BPP, averaging over 5 images each). We emphasize that both methods use our bitstream implementation, so Turbo-DDCM uses our efficient implementation without its combinatorial tables bitstream implementation. We note that we present the image compression task as a representative example to demonstrate improvement under the DDCM framework. As can be seen below, as well as in the paper, the improvement for higher latent dimensions than SD2.1 and for the autoregressive video compression is substantial.

SD2.1 (Kodak24, 512x512, d = 16384, 0.053 BPP):

MethodModel call (ms)Codebook compress (ms)Codebook decompress (ms)Bitstream compress (ms)Bitstream decompress (ms)Compress (s)Decompress (s)Round trip (s)
Turbo-DDCM (K=dK=d) w/ our coding27.23.012.232.193.500.540.901.44
H-DDCM27.21.510.802.223.520.500.881.38

FLUX (DIV2K, 1024x1024, d = 262144)

MethodModel call (ms)Codebook compress (ms)Codebook decompress (ms)Bitstream compress (ms)Bitstream decompress (ms)Compress (s)Decompress (s)Round trip (s)
Turbo-DDCM (K=dK=d) w/ our coding921837.5324.730.952.332.833.366.1
Turbo-DDCM (K=214K=2^{14}) w/ our coding92146.332.92.33.517.627.645.2
H-DDCM9211.950.9931.152.316.827.143.8

The gain is small on SD2.1 compared to FLUX, as the round trip is dominated by the diffusion model: on SD2.1 the network calls of the decoder alone take $\sim$0.8s of the $\sim$1.4s. The dense codebook of Turbo-DDCM costs O(dk)O(dk) per step, against O(dlog⁡d)O(d \log d) for our fast Walsh-Hadamard transform. At the small SD2.1 latent (d=214d=2^{14}) the dense codebook is only $\sim$0.5GB, and one step costs $\sim$3ms against a 27ms UNet call - the codebook accounts for just 0.09s of Turbo-DDCM’s 1.44s round trip, so removing it can save at most a few percent. Therefore, when the latent dimension is small, H-DDCM's runtime advantage over Turbo-DDCM is small, and its benefit is mainly memory (dense buffer of 0.54GB vs. less than 1MB) and scalability. The dense cost grows quadratically with dd: from SD2.1 to FLUX, the Turbo-DDCM codebook step grows ∼280×\sim 280\times (3ms →\rightarrow 838ms), while the model call grows only ∼34×\sim 34\times (27ms →\rightarrow 921ms) and ours stays at $\sim$2ms. On FLUX the codebook is then about half of a Turbo-DDCM compress step and about a third of its round trip (22s of 66s), whereas H-DDCM implicit codebook operations are less than 1% of the model call, resulting in 0.06s, which explains 66s for Turbo-DDCM, and 44s for H-DDCM. The dictionary-computation advantage becomes more important as latent dimensionality grows (higher resolutions, videos), although end-to-end gains remain dependent on generator inference.

Bitstream coding takes 2-4 ms per step on SD2.1 and 31-52 ms on FLUX, i.e. less than 6% of a model call, and runs in parallel to the diffusion model, so it is hidden behind inference and never the bottleneck for either method (with Turbo-DDCM's official table-based coding, the FLUX K=dK=d table alone takes approximately 23 min to build for a small M-value, see App. E).

We have added in App. C.1 Fig. S3 which demonstrates the runtime breakdown efficiency comparison between H-DDCM and Turbo-DDCM codebook.

Additional comparisons (W2/Q2):

We thank the reviewer for suggesting EF-LIC, DiT-IC, CoD-Lite, and OneDC. We evaluated all four on Kodak24, DIV2K, and CLIC2020, measuring PSNR, SSIM, LPIPS, FID, and encoding/decoding runtime. The results show comparable rate-distortion-perception performance in the evaluated regime, while specialized trained codecs are faster. This distinction is expected, as H-DDCM incurs the inference cost of its frozen generative backbone, but avoids training a task-specific codec and can be transferred to other distributions and modalities given an appropriate pretrained generator, whereas trained codecs typically need redesign and retraining for such changes. Within the evaluated training-free baselines and backbones, H-DDCM improves computational efficiency while retaining comparable rate-distortion-perception behavior. Specialized trained codecs remain substantially faster. Its low dictionary memory and computational complexity also make adaptation to the rapidly progressing larger pretrained generators for different modalities and latent dimensions more feasible. Moreover, pixel-space generative models, such as JiT [1, 2], have become more popular recently. Adapting a training-free compression scheme on these models is much more feasible with our contribution.

We added these baselines to Figs. 1, S1, and S2. The table below illustrates the comparison at approximately 0.03-0.05 BPP on Kodak24. Rates are not identical, therefore we refer to Fig. 1 to observe the curves obtained by each method.

MethodBPPPSNR ↑FID ↓Round-trip time (s) ↓
EF-LIC0.035622.2847.760.023
DiT-IC0.046122.6139.560.252
CoD-Lite0.031421.9843.640.035
OneDC0.025821.9142.290.266
H-DDCM0.034722.6640.231.422

[1] Li, Tianhong, and Kaiming He. "Back to basics: Let denoising generative models denoise." (CVPR’26)

[2] Vallaeys, Théophane, Jakob Verbeek, and Matthieu Cord. "Pixel-space video generation with U-JiT." (arXiv 11/2026)

We kindly refer reviewer #hdj2 to our latest response to reviewer #AExB regarding the extension of H-DDCM to the audio domain. Single-forward-pass codecs trained on a specific data distribution, such as [1-4], cannot easily be adapted to other data distributions, such as audio. This is an advantage of training-free methods when a suitable generative prior is available, as in this case. Specifically, we show that H-DDCM achieves higher perceptual quality than the evaluated methods while achieving at least comparable distortion.

Reviewer_gSdmOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes H-DDCM, a training-free generative image and video compression framework that addresses the scalability bottlenecks of Turbo-DDCM. It replaces explicit Gaussian codebooks with implicit randomized Hadamard dictionaries, enabling fast projection and reconstruction without quadratic dictionary storage. It also introduces group-factorized combinatorial coding and an overcomplete extension using multiple randomized bases. For video compression, the framework adapts a causal autoregressive flow model through a re-encoded-tail KV-cache refresh that carries reconstructed temporal context across segments without retransmitting it. Experiments demonstrate the effectiveness and efficiency across image and video benchmarks.

Critical strengths
  1. The Hadamard construction directly targets the two operations required by codebook-based compression: residual projection and selected-atom synthesis. It reduces the computation while preserving full-dimensional candidate coverage. The FLUX experiments provide useful evidence that this avoids the fidelity loss of capped Gaussian codebooks without incurring the cost of explicitly processing a complete codebook.
  2. The paper addresses the combinatorial bitstream bottleneck, which would otherwise remain after accelerating codebook operations. The table-free ranking implementation and group-factorized protocol form a coherent approach to large-dimensional coding. The video extension refreshes the cache from reconstructed frames, allowing both endpoints to recover temporal context without additional transmission. The cache ablations show advantages over hard resets and unrestricted cache growth.
Critical weaknesses
  1. Group-factorized coding reduces the size of individual combinatorial operations. Its count-vector payload and separately rounded group ranks change the support-coding cost relative to a single global rank. Appendix E specifies this cost, yet the experiments do not clearly quantify the resulting bitrate overhead or how the chosen group count balances rate against runtime. Moreover, Appendix G uses balanced per-basis counts to configure rate-matched overcomplete variants, whereas the actual counts depend on the selected support.
  2. The video experiments report average FPS on an H200, but processing three-latent-frame chunks and refreshing the cache at continuation boundaries can introduce buffering and periodic computation that average throughput does not reveal. The reported operating range also remains below conventional 24–30 FPS playback rates. The cache ablations support temporal continuity across segments, but startup latency, refresh-boundary latency, and quality over repeated refreshes are not clearly characterized. Consequently, the evidence is stronger for efficient context-bounded continuation than for broadly practical, near-real-time streaming.
Questions for authors
  1. How much bitrate overhead does grouping introduce compared with global combinatorial coding?
  2. Appendix C.7 shows that better aggregate metrics can coincide with worse temporal alignment. The perceptual gains over the baselines also preserve the source motion and event timing?
  3. How does the number of randomized Hadamard bases affect reconstruction quality and end-to-end runtime at a matched bitrate? The reason for using eight bases to tackle video compression can be further ablated.
  4. Some related works are not discussed and compared, such as [1] Extreme image compression using fine-tuned vqgan models, 2023. [2] Multi-realism image compression with a conditional generator, CVPR, 2023 [3] Once-for-All: Controllable Generative Image Compression with Dynamic Granularity Adaption, ICLR, 2025
4 comments

Group-factorized coding & Balanced per basis-counts (W1/Q1):

We thank the reviewer for raising this point. Group-factorized coding introduces a small bitrate overhead in exchange for faster encoding and decoding of the selected indices.

Specifically, grouping requires transmitting the number of selected atoms in each group, at a cost of ⌈log⁡2(M+G−1G−1)⌉\lceil \log_2 \binom{M+G-1}{G-1} \rceil bits per step. This overhead is partially offset by the shorter combinatorial ranks within each group.

To quantify this trade-off, we evaluated G∈1,2,4,8G\in \\{1,2,4,8\\} using the actual supports selected by H-DDCM, with G=1G=1 corresponding to global coding. The measured overhead is small, reaching at most 2% of the bitrate, or 0.002 BPP, for G=8G=8. Moreover, it decreases as the bitrate increases.

BackboneMMOverhead for G=2/4/8G=2/4/8 (%)Extra BPP (G=8G=8)Rank / unrank (ms), G=1G=1Rank / unrank (ms), G=8G=8
SD2.1 (d=214d=2^{14})1020.31 / 0.72 / 1.30+0.00112.2 / 3.61.9 / 3.2
3000.19 / 0.42 / 0.73+0.00203.0 / 4.42.1 / 3.4
FLUX (d=218d=2^{18})2190.14 / 0.33 / 0.66+0.000350 / 7232 / 53
10630.05 / 0.12 / 0.23+0.0005124 / 15639 / 61
46410.02 / 0.04 / 0.08+0.0007342 / 39567 / 90

Grouping becomes particularly beneficial at larger latent dimensions. For FLUX, it reduces the combined ranking and unranking time by 4.7×4.7\times at M=4641M=4641. Since bitstream coding overlaps with model inference, this additional efficiency keeps coding below the model-call time. For the smaller SD2.1 latent, the benefit is limited. We therefore use G=1G=1 for SD2.1 and G=8G=8 for FLUX. We added this analysis to App. E.

Regarding balanced per-basis counts, we clarify that these are used only to estimate the support budget MM for rate matching. The actual selection is unconstrained, and the encoder transmits the realized count of each basis. All reported BPP values are measured from the physical bitstreams.

We further verified this approximation over 500 video/rate experiments, finding a maximum bitrate deviation of only 0.465%.

Computations, latency, refresh overhead, and continuity (W2):

We agree that average FPS alone does not characterize latency. Across 500 H200 video/rate runs, decoder processing from the post-unpack handoff to reconstructed pixels averages 0.519 s for the first nine-frame chunk, 0.795 s for ordinary twelve-frame chunks, and 0.809 s for refresh-boundary chunks. Cache refresh requires 176 ms of mean component work at both the encoder and the decoder, including tail VAE encoding, cache reset and prefix commit - this is already included in the FPS computation. Complete encoder active work, including VAE processing, averages 0.945 s per chunk.

Following the Self-Forcing/Causal-Forcing generation setting, which reports that the generation is near-real-time, H-DDCM achieves nearly their generation throughput - encoding/decoding are 12.52/15.13 FPS versus the 15.79 FPS generation reference.

To quantify motion continuity, we additionally report source-flow-guided L2L_2 warping error [1-3] in the revised video evaluation. An optical flow is estimated from the source sequence and used to warp neighboring reconstructed frames. Then the discrepancy between the warped and reconstructed frames is evaluated and aggregated across the video. This is done only for pixels passing forward-backward flow-consistency checks in the source video. Importantly, a refresh-induced discontinuity would be captured by this error. To further assess quality across continuations, we refer the reviewer to the video comparisons on our anonymous website and the paper's ST-slice comparisons, including the cache-refresh ablation in App. C.7. These provide qualitative evidence of temporal continuity over the evaluated sequences.

The measured 12.52/15.13 FPS is close to the backbone throughput of 16 FPS but remains below 24-30 FPS playback. Since our method scales closely to the backbone's sampling throughput, we believe that 24 and higher FPS will become feasible in the near future. We have revised App. C.7 accordingly to include these measurements and characterizations.

Perceptual gains and source motion/event timing (Q2):

Improved aggregate perceptual metrics alone do not establish temporal alignment. We therefore added source-flow-guided L2L_2 warping error to the revised video evaluation (described above), in Figs. 3 and S8. This evaluates consistency with source motion rather than merely smoothness according to reconstructed flow. We complement it with the qualitative evaluations and ST-slices in the paper, e.g. Fig. 4, as well as qualitative videos playback comparison in our anonymous website. This provides quantitative and qualitative evidence for motion correspondence, providing a more complete evaluation when combined with the existing perceptual metrics.

Number of Hadamard bases (Q3):

We refer to App. C.5, in which we performed an ablation between different numbers of Hadamard bases (ranging from B=1B=1 to B=8B=8 bases) at a matched bitrate on all 30 MCL-JCV samples. We selected B=8B=8 empirically as a favorable quality-runtime operating point among the evaluated configurations, not as a theoretically optimal or saturated choice. For example, at approximately 0.0056 BPP, B=8B=8 improves PSNR by 0.17 dB and reduces FID by 8.2% relative to B=1B=1, while measured encoding and decoding throughput remains comparable across basis counts - approximately 12.5 FPS and 15.2 FPS, respectively. The provided qualitative result supports this ablation, and App. C.5 now includes the runtime measurement.

Related works (Q4):

We thank the reviewer for these references. We have expanded the related work section to discuss these papers and clarify their relationship to our method. Additionally, we have further expanded our experimental comparisons with CGIC [R3], the most recent work suggested by the reviewer, as in [R3] they have evaluated against [R1] and [R2], and achieved better performance. The following table reports the obtained metrics (PSNR, FID and round-trip time) for 2 illustrative BPP values, approximately 0.07 and 0.1, on the Kodak24 dataset. The results demonstrate that H-DDCM achieves competitive performance, without any architecture fine-tuning or training. We have updated Figs. 1, S1 and S2 (quantitative results) to include this additional baseline.

MethodBPPPSNR ↑FID ↓Round-trip time (s) ↓
CGIC0.07122.0475.70.43
H-DDCM0.06624.1034.61.66
CGIC0.10323.2755.30.44
H-DDCM0.10125.0032.01.84

[1] Lai, Wei-Sheng, et al. "Learning blind video temporal consistency." (ECCV’18)

[2] Ceylan, Duygu, Chun-Hao P. Huang, and Niloy J. Mitra. "Pix2video: Video editing using image diffusion." (ICCV’23)

[3] Geyer, Michal, et al. "Tokenflow: Consistent diffusion features for consistent video editing." (ICLR’24)

I appreciate the authors' response. Some of my concerns have been addressed. However, there are still some issues:

  1. As mentioned, for SD2.1, the benefit related to G is limited, and the authors finally set G=1. This poses a concern about the necessity and applicability of the group factorization in generative image compression with diverse diffusion-based backbones.
  2. The models with different B values seem to perform unstably in DOVER. Could you analyze this phenomenon?
  3. Though CGIC performs better than the other two methods, this does not mean that these methods can be ignored. Accordingly, in my view, the authors should provide a clearer comparison.

Group-factorized coding:

As noted above, for a small latent space such as that of SD2.1, a larger number of groups brings no advantage, so we use G=1G=1 there. For the larger latents of FLUX and CF we use G=8G=8. The benefit of group-factorized coding grows with the latent dimension, so it becomes more relevant as generators grow. This is already happening: pixel-space generative models such as JiT [1, 2] are becoming popular, and for such models we expect group-factorized coding to reduce the coding cost.

DOVER metric:

The DOVER curves are largely flat, with small non-monotonic variations, and do not show a systematic degradation as the number of bases increases. Across the evaluated basis counts and rate points, the mean scores range approximately from 0.597 to 0.606, with per-point standard errors around 0.004–0.006 across videos.

DOVER is a no-reference quality measure. Consequently, improving reconstruction fidelity (which is what increasing the number of bases targets) does not necessarily increase DOVER. A plausible explanation is that the diffusion prior already produces perceptually plausible videos at low rates for this metric. We therefore interpret these results as broadly similar DOVER performance across basis counts, rather than using the small fluctuations to rank the variants.

Related works:

We evaluated MRIC [R2] on Kodak24, DIV2K and CLIC2020. Its official code is not public, so we used the TensorFlow 2 reimplementation with its released λ=0.128\lambda=0.128 checkpoint; MRIC-D and MRIC-R denote the decoder-side realism parameter β=0\beta=0 and β=2.56\beta=2.56, respectively. We could not evaluate [R1]: the weights of its official implementation are on a Google Drive that requires permission. We have requested access without success so far, and will add the comparison if we get it.

The table compares H-DDCM and MRIC on Kodak24 at a similar rate. We take the largest MM whose rate does not exceed that of MRIC. The results on DIV2K and CLIC2020 are in the revised manuscript (Figs. S1 and S2).

MethodBPPPSNR ↑LPIPS ↓FID ↓Round trip (s) ↓
MRIC-D0.13627.800.19778.60.12
MRIC-R0.13626.720.07852.00.12
H-DDCM0.12825.400.09331.02.06

H-DDCM achieves better FID when compared to both variants, which aligns with higher perceptual quality. Both MRIC variants are faster and slightly better in PSNR. MRIC-R is better in LPIPS. We have updated Figs. 1, S1 and S2 to include MRIC.

[1] Li, Tianhong, and Kaiming He. "Back to basics: Let denoising generative models denoise." (CVPR’26)

[2] Vallaeys, Théophane, Jakob Verbeek, and Matthieu Cord. "Pixel-space video generation with U-JiT." (arXiv 11/2026)

Comments on the paper

3 comments

As the discussion period is about to end, we would like to thank all reviewers again for their thoughtful feedback and engagement. We have uploaded an updated manuscript and accompanying anonymous project page reflecting all requested analyses and experiments, and we believe we have thoroughly addressed every question and concern raised. We summarize the key updates below:

  • Runtime Breakdowns and Coding Scalability: We added per-step runtime decompositions on SD2.1 and FLUX (Fig. S3), group-factorization rate-runtime trade-offs (App. E), and chunk-level video latency and cache-refresh measurements (App. C.7).
  • Expanded Image and Video Comparisons: We added rate-distortion-perception and runtime evaluations against 6 new trained methods - CGIC, MRIC, EF-LIC, DiT-IC, CoD-Lite, and OneDC, across Kodak24, DIV2K, and CLIC2020 (Figs. 1, S1, S2), demonstrating that H-DDCM achieves comparable performance to trained image codecs. Additionally, we compared against the concurrent trained method GVC-RT and added a source-flow-guided L2L_2 warping error measurement across our video benchmarks (Figs. 3, 4, S8-S14, and our anonymous webpage), showing that H-DDCM achieves strong perceptual quality and temporal consistency while being the fastest training-free method; pairing H-DDCM with a faster generative prior would directly yield higher throughput.
  • Cross-Modality Transfer Proof-of-Concept: To directly demonstrate the practical advantage of scalable training-free compression over modality-specific trained codecs, we added a zero-shot audio compression evaluation using frozen Stable Audio Open against EnCodec and Opus on SQAM (App. H and our webpage), achieving less than half the FAD of EnCodec without any codec training.
  • Discussion and Ablations: We expanded the discussion comparing training-free and trained paradigms (Sec. 6), clarified DCVC-UF settings, and added basis-count runtime (App. C.5).

We would greatly appreciate the opportunity to address any remaining concerns or points of clarification before the discussion period closes.

Dear reviewers, towards the end of the discussion phase, we trust that our responses have successfully addressed your inquiries. We look forward to receiving your feedback regarding whether our replies sufficiently resolve any concerns you may have, or if further clarification is needed.

Area_Chair_QGtYOpenReview ↗

Please join the discussion before it closes

Thank you to the authors for the rebuttal and to the reviewers for engaging in the discussion. As the discussion period draws to a close, I encourage everyone to keep the conversation going. I’d appreciate a brief final assessment from each reviewer. After reading the rebuttal and the discussion so far, what is the main reason you lean toward acceptance or rejection, and what evidence supports that view? Please note which concerns the rebuttal resolved and which remaining concerns, if any, are decisive for your recommendation.

If your assessment differs from another reviewer’s, please point to the specific claim or evidence you interpret differently. Authors are welcome to clarify any factual points that remain open. This will help us identify where we agree and understand the reasons for any remaining disagreement.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.