acceptodds
Under review as a conference paper at ICLR 2027

Flux as a Codec: Zero-Shot Lossy Compression with Flux

Abstract

We propose FluxC, a zero-shot codec that uses Flux as a compression prior. Although DiffC has shown that pretrained diffusion models can serve as zero-shot codecs, we ask how the choice of coding path and reconstruction method affects rate–distortion–perception behavior. We analyze FluxC along three axes: coding cost and bit allocation, reconstruction method, and low rate distortion. At matched signal-to-noise ratio (SNR), we show that the path choice does not affect the ideal information rate, but changes how bits are allocated across time steps. For reconstruction, we compare ancestral sampling, one-step ODE reconstruction, and many-step ODE reconstruction. Our analysis shows that ODE-based reconstruction can achieve lower distortion than ancestral sampling, especially in the low rate regime. Experiments on Gaussian sources and natural images support this prediction. Moreover, FluxC performs competitively with, and in some cases outperforms, recent neural image compression methods, suggesting that Flux can serve as a principled zero-shot codec.

Primary area:
generative models
Keywords:
Lossy Compression, Generative Models

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_9qafOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FluxC, a zero-shot codec that uses a pretrained Flux model as a compression prior without any additional training. Whereas the existing DiffC Flux reparameterizes Flux onto a VP diffusion path, FluxC communicates noisy latents via reverse channel coding directly on Flux's native OT path. At the receiver, the image is reconstructed either by one-step posterior-mean decoding (FluxC-1) or by many-step ODE integration (FluxC-ODE). The authors analyze the codec along three axes: coding cost and bit allocation over path time, reconstruction method, and low-rate distortion behavior. They show that, at matched SNR, the choice of path does not change the ideal rate but only redistributes bits over time, and that in the low-rate limit, deterministic ODE reconstruction reduces distortion faster than ancestral sampling. Experiments on Gaussian sources and natural images are reported to support these predictions, and the method is reported to be competitive with existing neural codecs in the ultra-low bitrate regime.

Critical strengths

The main strength of this paper lies not in the superiority of a new codec design, but in its analytical treatment of Flux-based zero-shot compression and its transparent experimental reporting.

  • By using the pretrained FLUX.1-dev as a compression prior without additional training, the paper extends DiffC-style zero-shot coding to the native parameterization of a modern rectified flow model, which is of practical relevance.
  • The coding cost is decomposed into an information term that depends only on the SNR and a model-mismatch term, and is further re-expressed in terms of Flux's velocity regression error and the bit allocation over path time. This clearly shows that the choice of path does not change the ideal rate.
  • The paper establishes that, in the low-rate limit, distortion decreases at order √R for deterministic ODE reconstruction and at order R for ancestral sampling, and verifies this in closed form for Gaussian sources.
  • Results that disagree with the theory are reported transparently, e.g., DiffC-A achieving higher PSNR than DiffC-F and rankings that vary across metrics. Encoding/decoding times and the effect of the number of encoding steps are also provided.
Critical weaknesses
  1. Limited theoretical gain and weak empirical support for Result 1. The paper is primarily analytical rather than demonstrating the superiority of a new technique, and Result 1 is essentially the only component that offers a theoretical gain. The cleanest test of Result 1 in Fig. 3 is DiffC Flux-F versus DiffC Flux-A, which differ only in the reconstruction method under the same backbone and path; however, the difference between them is not clearly visible. Moreover, it is questionable whether the distortion reduction below the extremely low bitrate of 2 × 10⁻⁴ bpp is of practical significance.

  2. Missing comparisons with recent generative codecs. Beyond PerCo and DiT-IC, comparisons with other recent generative codecs such as OneDC, StableCodec, and AEIC are not provided. Even if the proposed method does not outperform them, such comparisons are necessary to assess its performance level.

  3. Overextension of Result 1. Result 1 characterizes the leading-order rate of distortion reduction as R → 0 from a common zero-rate distortion D₀; it does not by itself predict the absolute distortion ranking at a given rate. The paper nevertheless invokes it to explain the PSNR advantage of FluxC-1 and FluxC-ODE over a wide low-rate range in Fig. 4, which extends well beyond the asymptotic regime. Moreover, FluxC-1 is a posterior-mean decoder and falls outside the scope of Result 1. Please (i) report the empirical D₀ of each method, and (ii) verify the predicted scaling directly via the slope of log(D₀ − D) versus log R.

  4. No demonstrated RDP advantage of the OT path. The paper does not demonstrate an RDP advantage of performing RCC directly on Flux's OT path over the existing VP-normalized approach (DiffC Flux). A controlled comparison with the same backbone, step schedule, and reconstruction method (e.g., FluxC-ODE vs. DiffC Flux-F) is needed.

  5. Computational cost should be discussed as a limitation. DiffC-style methods typically require tens to hundreds of seconds per image, which is several orders of magnitude slower than conventional neural codecs (e.g., MS-ILLM). Although this issue is not specific to the proposed method, it is a critical practical challenge for this line of research and should be discussed as a limitation.

Questions for authors
  1. Can the authors provide a clearer comparison between DiffC Flux-F and DiffC Flux-A (e.g., numerical ΔMSE values at several rates or a figure in which the two are clearly distinguishable) to verify Result 1 on natural images? In addition, what is the practical significance of the distortion reduction below 2 × 10⁻⁴ bpp?
  2. Can the authors include comparisons with recent generative codecs such as OneDC, StableCodec, and AEIC to clarify the performance level of the proposed method?
  3. Can the authors (i) report the empirical zero-rate distortion D₀ of each method, and (ii) verify the predicted scaling of Result 1 directly via the slope of log(D₀ − D) versus log R? How do the authors justify using Result 1 to explain the PSNR advantage over the wide rate range in Fig. 4, given that it is an asymptotic result and that FluxC-1 falls outside its scope?
  4. Can the authors provide a controlled comparison between FluxC and DiffC Flux with the same backbone, step schedule, and reconstruction method (e.g., FluxC-ODE vs. DiffC Flux-F) to isolate the effect of performing RCC on the OT path?
  5. Can the authors discuss the computational cost of the method as a limitation, including a comparison of encoding/decoding time with conventional neural codecs?
5 comments

We thank the reviewer for the thorough and precise review. Several questions were shared with other reviewers, and we answer them in the General Response: D0D_0 and the slope of log⁡(D0−D)\log(D_0-D) versus log⁡R\log R in General Response 2, and the controlled comparison of the path and the reconstruction method in General Response 1. We answer the remaining questions below.

Can the authors provide a clearer comparison between DiffC Flux-F and DiffC Flux-A

We report the PSNR BD-rate of DiffC Flux-F relative to DiffC Flux-A within each rate below. The negative value means DiffC-F requires less rate to achieve the same PSNR than DiffC-A. Therefore DiifC Flux-F consume less rate in every rate on all datasets.

Rate decade (bpp)KodakDIV2KCLIC
[10−4,10−3)[10^{-4},10^{-3})−37.0%−32.8%−30.3%
[10−3,10−2)[10^{-3},10^{-2})−34.1%−36.9%−29.0%
[10−2,10−1)[10^{-2},10^{-1})−34.2%−34.7%−32.2%

Appendix E.1 shows that ODE reconstruction, which stays aligned with the posterior mean instead of drawing a independent posterior sample, has lower distortion than ancestral sampling at every SNR for Gaussian source. Natural images also show the same trend that with the Flux backbone, ODE reconstruction requires 22–37% less rate than ancestral sampling for the same PSNR in every rate decade from 10−410^{-4} to 10−110^{-1} bpp (Table 1).

what is the practical significance of the distortion reduction below 2 × 10⁻⁴ bpp?

Extremely low rates are relevant in practice, e.g., for satellite and remote-sensing imagery, where the downlink bandwidth is severely limited [1, 2], and for low-power IoT links such as LoRa, which offer only tens of kbps [3]. When the rate is constrained to be extremely low, choosing the method with the lowest distortion at that rate is the natural criterion. Although 2×10−42\times10^{-4} bpp corresponds to only about 50 bits for a 512×512512\times512 image, a high resolution image carries proportionally more bits (about 1,700 bits for a 4K image). Given the growing use of high resolution images such as 4K and the need to store a lot of images, the distortion reduction at extremely low bpp becomes more useful as the resolution increases.

Can the authors include comparisons with recent generative codecs such as OneDC, StableCodec, and AEIC?

We added such recent works in revised version in Figure 4, 13, 14 and 15, and Table 9. At 0.01 bpp and above, these codecs achieve better distortion and perceptual quality than FluxC, which is expected since they are trained for compression. Without any training for compression, FluxC operates below 0.003 bpp, where these codecs report no operating points. Tuning the decoder for perceptual metrics is left for future work.

[1] CCSDS 122.0-B-1, Image Data Compression, Blue Book, Issue 1, Nov. 2005.

[2] Ye et al., "Map-Assisted Remote-Sensing Image Compression at Extremely Low Bitrates" (2024).

[3] De, Soumadeep, et al. "Multimedia transmission over lora networks for iot applications: A survey of strategies, deployments, and open challenges." Sensors (Basel, Switzerland) 25.23 (2025): 7128.

Reviewer_9qafOpenReview ↗

Clarification on the Contribution of Result 1

Thank you for your detailed responses. Some of my concerns have been addressed.

However, I would appreciate further clarification regarding the significance of Result 1, which appears to be one of the main contributions of this paper.

Beyond the existing understanding of posterior sampling and deterministic transport, what new insights or practical guidance does Result 1 offer for the design of generative compression systems?

Thank you for taking the time to review our revisions and share your feedback. What we are interested in is that how to reconstruct an image that is faithful to the input, in the MSE sense, from a very small amount of received information. This question motivates the low rate analysis of Result 1, which contributes both theoretically and practically.

Theoretically, Theis et al. [1] showed that ODE reconstruction achieves up to a 3 dB gain over ancestral sampling, attained at high rates. As R→0R\to0, however, both distortions approach D0D_0, so this ratio tends to 1 and no longer distinguishes the decoders. Result 1 complements their analysis at low rates: ODE reconstruction reduces the distortion at order R\sqrt{R}, whereas ancestral sampling does so at order RR. Together with Appendix E.1, which shows that ODE reconstruction is better at every SNR for Gaussian sources, this extends their comparison of the two decoders from high rates to the whole rate range.

Practically, Result 1 suggests choosing ODE reconstruction at ultra low rates (the fitted slopes agree with the predicted 1/21/2 and 11; Figure 3). Combined with Observation 1, it suggests that the choice of the decoder matters more than the choice of the path (ODE reconstruction saves 22–37% of the rate; Table 1). Combined with Appendix E.1, it suggests that ODE reconstruction is better over a wide range of rates (it needs less rate in every rate decade from 10−410^{-4} to 10−110^{-1} bpp; Table 1).

In summary, for our question, Result 1 shows theoretically and empirically that choosing ODE reconstruction as the decoder is beneficial in the MSE sense. Recent work on neural image compression is increasingly interested in decoder design at ultra-low rates [2, 3, 4], and we believe that Result 1 is useful for such designs and for practical applications.

[1] Theis, Lucas, et al. "Lossy compression with gaussian diffusion." arXiv preprint arXiv:2206.08889 (2022).

[2] Careil, Marlene, et al. "Towards image compression with perfect realism at ultra-low bitrates." International conference on learning representations. Vol. 2024. 2024.

[3]Zhang, Tianyu, Dong Liu, and Chang Wen Chen. "Ultra-low bitrate perceptual image compression with shallow encoder." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

[4] Xue, Naifu, et al. "One-step diffusion-based image compression with semantic distillation." Advances in neural information processing systems 38 (2026): 37108-37144.

Reviewer_1bFROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FluxC, a zero-shot image codec using Flux and reverse channel coding along the OT path. It analyzes coding costs, bit allocation, and ODE-based reconstruction, with experiments showing competitive low-bitrate performance.

Critical strengths
  1. Applying zero-shot compression to Rectified Flow models is interesting.
  2. The theoretical analysis is useful, especially the Gaussian-source results.
  3. Competitive results on several image benchmarks.
Critical weaknesses
  1. The benefit of using the native OT path is not clearly established. Different bit allocation alone is not a strong contribution.
  2. The improvements may mainly come from the decoder rather than the coding path. These effects are not well separated.
  3. The low-rate theory relies on idealized assumptions and does not directly apply to FluxC-1. Figure 3 does not convincingly verify the claimed scaling.
  4. The novelty over recent diffusion and flow-based compression methods needs clearer justification.
Questions for authors
  1. Can the authors isolate the benefit of OT over VP using the same backbone and decoder?
  2. Does the low-rate theoretical result apply to FluxC-1? If not, how does it explain its experimental advantage?
  3. What is the main technical difference from recent Rectified Flow-based compression methods?
  4. How many independent image patches were used for FID evaluation?
3 comments

We thank the reviewer for the constructive comments. Several questions were shared with other reviewers, so we answer them in the General Response above: the low rate theory and the verification of Figure 3 in General Response 2, isolating the effect of OT from VP with the same backbone and decoder in General Response 1, and the difference from prior work in General Response 3. We answer the remaining questions below.

The benefit of using the native OT path is not clearly established. Different bit allocation alone is not a strong contribution.

We agree, and the new controlled comparison confirms it that with the same Flux backbone and the same reconstruction, the PSNR BD-rate between OT and VP has no consistent sign across rate decades and datasets (General Response 1). This agrees with our analysis where at matched SNR, the two paths have the same ideal rate (Observation 1) and the same ideal reconstruction (Appendix B.1). We therefore no longer claim an advantage of the OT path itself. Our contribution is that FluxC defines a valid codec on the native OT path of Flux, and that the reconstruction method, not the path, determines the rate–distortion performance (22–37% BD-rate per rate decade).

Does the low-rate theoretical result apply to FluxC-1? If not, how does it explain its experimental advantage?

Result 1 does not apply to FluxC-1, only to FluxC-ODE (and ancestral sampling). The advantage of FluxC-1 comes instead from the fact that it outputs the posterior mean, which minimizes the MSE. This is why FluxC-1 achieves the highest PSNR in our experiments. We have revised the corresponding explanation in Section 4.

What is the main technical difference from recent Rectified Flow-based compression methods?

The main difference from recent rectified flow based methods is whether the model is tuned for image compression. DiT-IC [1] trains a flow matching model specifically for image compression, whereas we use Flux as an image codec without any tuning. DiffC Flux [2] also treats Flux as a zero-shot codec, but it does not analyze Flux as a codec in detail, such as how the reconstruction method affects the rate-distortion-perception behavior, or the asymptotic improvement of ODE reconstruciton at low rates. [1] Shi, Junqi, et al. "DiT-IC: Aligned Diffusion Transformer for Efficient Image Compression." (2026). [2] Liu, Feng. "Lossy compression with pretrained diffusion models." (2025).

How many independent image patches were used for FID evaluation?

We used 2100 independent patches from DIV2K and CLIC datasets. We recomputed FID using each patch exactly once and additionally report KID, which is less biased for small sample sizes. We report it in Figure 15 and find that the trends are mostly the same with the one of reported in main paper.

Thank you for the clarifications. I maintain my rating.

Thank you for your response. We would appreciate it if you could let us know which of your concerns remain insufficiently addressed so that we can respond within the discussion period. Thank you again for your time and consideration.

Reviewer_jhJmOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FluxC, a zero-shot lossy compression method based on diffusion models. Its main idea is to exploit shared randomness between the encoder (sender) and decoder (receiver) to enable efficient compression. The main contributions include: (1) the paper evaluates the coding cost and signal-to-noise ratio (SNR) of two diffusion paths: the variance-preserving (VP) path and the optimal transport (OT) path, (2) it compares the performance of FluxC with that of DiffC and (3) it investigates three reconstruction strategies: one-step reconstruction, multi-step ODE reconstruction, and an ancestral-sampling baseline and (4), it develops an information-theoretic analysis of the coding cost and characterizes the distortion behavior in the low-rate regime.

Critical strengths
  1. The paper asks welll-motivated research question. The paper asks whether a pretrained flow-matching model can serve as a zero-shot lossy codec, and whether using its native OT path is advantageous compared with a VP diffusion path.

  2. The coding-cost expression separates an ideal mutual-information contribution from a denoiser-mismatch term (observation 1). This provides a useful conceptual distinction between the information theoretically required rate and the additional cost induced by an imperfect learned estimator.

  3. The paper considers one-step estimation, ODE reconstruction, and ancestral sampling, and recognizes that pixelwise distortion and perceptual quality need not rank these methods identically.

  4. The Gaussian-source analysis offers a controlled setting, while the natural-image experiments use Kodak, DIV2K, and CLIC data and report several metrics, including PSNR, MS-SSIM, LPIPS, and FID.

Critical weaknesses
  1. The novelty is limited, as DiffC already explores diffusion-based compression with shared randomness.

\item The relationship between ideal rate and practical coding cost needs clarification. The ideal cost is the KL divergence between pp and qq. In prior work, this cost is motivated by rejection sampling (Communication Complexity of Correlation'') and Poisson processes (Strong Functional Representation''), with entropy coding used to achieve compression. The paper should therefore clarify how this ideal cost translates into practical bitrate, accounting for coding overhead, quantization, stopping information, and side information, and explicitly distinguish the ideal rate from the realized bitrate.

  1. About the low-rate distortion results, $ D_{\mathrm{ODE}}(R)=D_0-c_{\mathrm{ODE}}\sqrt{R}+o(\sqrt{R}), \quad D_{\mathrm{A}}(R)=D_0-c_{\mathrm{A}}R+o(R), $ depend on assumptions about ODE marginal preservation and endpoint estimator behavior. The paper should distinguish these theoretical assumptions from properties verified for the learned FLUX model and test whether the predicted scaling holds at practical rates.

  2. Comparisons should use matched realized bits per pixel (bpp) and report encoding/decoding time, hardware, model size, and inference steps to establish practical trade-offs.

  3. Evaluation is limited to center-cropped images. Broader resolutions, aspect ratios, and input-specific semantic-consistency metrics would strengthen the results, as FID alone does not establish fidelity to individual inputs.

  4. The interpolation Zt=(1−t)X+tϵZ_t=(1-t)X+t\epsilon specifies OT-style marginals but not a unique Markov process. The paper should clarify the relationship among the marginal path, Markov transitions, RCC conditional distribution, and ODE velocity field, including the validity of any Gaussian transition formulas.

  5. The stopping point near t=0.2t=0.2 may affect rate-distortion performance. An ablation over stopping times is needed to assess robustness.

Questions for authors

The following additions would make the paper stronger:

See all the weaknesses above. Additionally, the paper would benefit from

  1. Report realized bitrate with a complete accounting of transmitted information and overhead.
  2. Evaluate sensitivity to stopping time and ODE solver steps.
  3. Use controlled ablations to isolate the effects of the OT path and reconstruction method.
  4. Improve the readability of graph annotations in Figure 1.
  5. Clarify that the conditional mean is optimal for squared-error distortion under the assumed regression setting, not necessarily for perceptual metrics.
3 comments

We thank the reviewer for the detailed comments. Several points were shared with other reviewers: the controlled ablation of the path and the reconstruction method in General Response 1, the assumptions of the low-rate result and the test of its scaling in General Response 2, and novelty and computational cost in General Response 3. We answer the remaining points below.

The novelty is limited, as DiffC already explores diffusion-based compression with shared randomness.

DiffC defines its codec on the VP path. We define a valid codec on the native OT path of Flux and show that, at matched SNR, the path changes neither the ideal rate nor the ideal reconstruction, whereas the reconstruction method changes the distortion. We show this theoretically (Observation 1, Result 1) and empirically (Table 1, Figure 3). Please also see General Response 3.

explicitly distinguish the ideal rate from the realized bitrate.

We thank the reviewer for this insightful comment. The submission reported the length of the index stream without the side information. We have updated all bpps of FluxC and DiffC to the realized rate. This update does not change the trends or the conclusions of the paper. We have also added a discussion of the ideal rate and the realized rate to Appendix E.2 (Table 4).

report encoding/decoding time, hardware, model size,

Table 9 reports the encoding and decoding times on the same hardware (NVIDIA RTX 3090) and the model sizes of FluxC compared with recent neural codecs. Please also see General Response 3.

inference steps to establish practical trade-offs

We conducted an experiments by changing the number of inference steps and we found 10 steps take the best trade-offs between performance (distortion) and processing times. See Figure 12 and Table 8 in the revised paper.

Broader resolutions, aspect ratios,

We added the Figure 13 and 14 for higher resolution images (longest side of 1024), which shows the ranking of the method is almost the same with the one of center crops, and FluxC can perform well at low bit rate (around 0.001 bpp) even for high resolution.

input-specific semantic-consistency metrics would strengthen the results.

In revised version, we added DISTS [1] in Figure 4 and 13. The advantage of FluxC at low rates is consistent with the trends in the other perceptual distortion metrics.

The paper should clarify the relationship among the marginal path, Markov transitions, RCC conditional distribution, and ODE velocity field, including the validity of any Gaussian transition formulas.

Section 3.1 and Appendix B.1 now define the Gaussian Markov process that FluxC uses for coding, and show that its marginal coincides with that of the OT path at every time. Since this process is a Gaussian Markov process, the posterior used for RCC is derived in the same way as in DDPMs. The sender uses this posterior, q(zs∣zt,x0)q(\mathbf{z}_s \mid \mathbf{z}_t, \mathbf{x}_0), and the receiver uses its approximation pθ(zs∣zt)p_\theta(\mathbf{z}_s\mid\mathbf{z}_t), which replaces x0\mathbf{x}_0 by the denoising prediction obtained from the Flux velocity field.

An ablation over stopping times is needed to assess robustness.

We added the ablation study changing the stopping times in encoding in Figure 11. It shows moving the stopping time closer to the data endpoint extends the curve to higher rates, but the rate of each additional step grows rapidly which is consistent with Observation 2.

Improve the readability of graph annotations in Figure 1.

We revised Figure 1 with larger fonts, a shared legend and one curve per codec.

Clarify that the conditional mean is optimal for squared-error distortion ..., not necessarily for perceptual metrics.

We clarified this points in Section 4 (highlighted in blue).

[1] Ding, Keyan, et al. "Image quality assessment: Unifying structure and texture similarity." (2020)

Thank you for the clarifications. I maintain my rating.

Thank you for your response. We would greatly appreciate it if you could let us know which concerns remain unresolved. Any specific suggestions would be very helpful for improving the paper. Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful reading. We answer shared concerns here and the remaining points individually. Revisions are highlighted in blue.

1. Separating the effects of the path and the reconstruction method (1bFR, 9qaf, jhJm)

Reviewers were concerned that the experiments did not separate the effect of the path from that of the reconstruction method. To address this, we fixed the backbone to Flux and changed one factor at a time: (i) the path (OT vs. VP) with the reconstruction fixed, and (ii) the reconstruction (ODE vs. ancestral sampling) with the path fixed. Table 1 reports the PSNR BD-rate.

Path. With the reconstruction fixed, OT vs. VP has no consistent sign across rate decades and datasets (−8.9% to +15.3%). This agrees with our analysis that at matched SNR, the two paths have the same ideal rate (Observation 1), and their latents are related by a deterministic rescaling, so ideal decoders yield the same distortion (Appendix B.1). The remaining differences come from the different step schedules and the network's approximation error.

Reconstruction. With the path fixed, ODE reconstruction needs 22–37% less rate than ancestral sampling in every rate decade from 10−410^{-4} to 10−110^{-1} bpp. Thus the reconstruction method, not the path, determines the rate–distortion performance. Appendix E.1 explains this gap where ancestral sampling draws a posterior sample, whereas ODE reconstruction is a deterministic function of the observation aligned with the posterior mean.

2. Result 1: zero-rate distortion D0D_0 and the low-rate scaling (1bFR, 9qaf, jhJm)

Reviewers asked us (i) to report D0D_0, (ii) to verify the predicted scaling directly through the slope of log⁡(D0−D)\log(D_0 - D) versus log⁡R\log R, and (iii) to examine whether the assumption of Result 1, that the reconstruction is a sample from the source marginal, holds in the Flux experiments.

We added a dedicated experiment (Appendix E.2) on all 24 Kodak images and 30 images each from DIV2K and CLIC, averaged over four shared seeds.

D0D_0 and the assumption. Table 5 reports D0D_0 of each decoder (32 seeds) and its ideal value under PXP_X, estimated from pairs of distinct images. The measured D0D_0 is 8–27% smaller, so the marginal-preservation assumption of Result 1 holds approximately for Flux.

Scaling. The new Figure 3 plots log⁡(D0−D)\log(D_0-D) versus log⁡R\log R. Between 2 bits and 10−410^{-4} bpp, the fitted slopes are 0.35–0.56 for ODE reconstruction and 0.80–1.53 for ancestral sampling, close to the predicted 1/2 and 1.

FluxC-1. As reviewers pointed out, Result 1 covers ODE reconstruction and ancestral sampling, not FluxC-1. We now attribute the PSNR advantage of FluxC-1 to the posterior mean minimizing the MSE (Section 3.3).

3. Computational cost, difference from prior work, and novelty (1bFR, 9qaf, jhJm)

Cost. FluxC is 1.5–2.5 orders of magnitude slower in encoding and decoding than recent neural codecs (Table 9), because RCC transmits the latent step by step and each step requires a forward pass of the 12B backbone. We state this as a limitation.

Prior work and novelty. Codecs such as PerCo, DiT-IC and OneDC are trained end to end for compression and naturally achieve better perceptual quality at 0.01 bpp and above. FluxC uses Flux without any training, which shows how far a generative prior alone can go and isolates the factors that determine performance. Our contributions are:

  1. Low-rate behavior of the reconstruction method. Theoretically and empirically, as R→0R\to0, ODE reconstruction reduces the distortion at order R\sqrt{R}, ancestral sampling at order RR.
  2. The reconstruction method, not the path, determines the performance. With the path fixed, ODE reconstruction requires 22–37% less rate than ancestral sampling, whereas the difference between the paths has no consistent sign.
  3. Performance without training for compression. FluxC outperforms PerCo and DDCM, and it operates below 0.003 bpp, where recent trained codecs report no operating points.

We leave the acceleration of RCC, FluxC with the distilled foundation model and tuning for perceptual metrics to future work.

Main revisions

The main revisions are as follows:

  • Clarified contributions and limitations (Abstract, introduction, and conclusion).
  • Explanation of ODE vs. ancestral sampling beyond the low-rate limit (Appendix E.1).
  • All bpp of FluxC and DiffC include RCC side information (Table 4, Appendix E.2).
  • New experiments:
    • path vs. reconstruction (Table 1m Section 4).
    • seed dependence of DiffC-F and DiffC-A (Table 8 and 9, Appendix E.2).
    • low-rate test of Result 1 (Figure 3, Table 5, and Appendix E.2).
    • higher resolution images (Figure 13-14 and Appendix E.3).
    • the number of decoding steps (Figure 12, Table 8 and Appendix E.2).
    • stopping times (Figure 11 and Appendix E.2).
    • computational cost (Table 9 and Appendix E.3).
    • report with new metrics (DISTS and KID).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.