acceptodds
Under review as a conference paper at ICLR 2027

Calibrating Pretrained Generative Models for Practical Reverse-Channel Coding

Abstract

Pretrained diffusion and flow models can be used for image compression through reverse-channel coding (RCC), but mismatch between the deployed receiver and the source-marginalized ideal receiver can increase communication cost. We decompose ideal expected RCC cost into required source information and receiver-mismatch excess, identifying the part that calibration can reduce. This analysis leads to Calibrated Reverse-Channel Coding (CalRCC), which calibrates prediction, confidence, and path. Prediction balances coding cost and reconstruction quality, confidence adjusts receiver uncertainty, and path reallocates model evaluations along the reverse process. Experiments with controlled DiffC show that CalRCC substantially improves practical compression, with stronger fidelity and perceptual quality without training a new generative backbone. Studies with CoD and Turbo-DDCM further suggest that the principle of prediction calibration can extend across distinct diffusion codec designs. These results show that receiver calibration provides an effective way to adapt pretrained generative models for practical compression.

Primary area:
generative models
Keywords:
reverse-channel coding, relative entropy coding, generative image compression, diffusion-based compression, pretrained generative models, receiver calibration

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_GtgeOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a calibration method for adapting pretrained generative models to practical reverse-channel coding, achieving promising rate–distortion performance. However, the theoretical motivation based on latent-space KL divergence is not fully aligned with the practical RGB reconstruction objective, and further ablations are needed to clarify this discrepancy. In addition, comparisons with DiffC-F and the inclusion of CoD and Turbo-DDCM in the main rate–distortion evaluation would strengthen the experimental evidence.

Critical strengths

S1. The proposed calibration method provides a principled approach to adapting pretrained generative models for practical reverse-channel coding.

S2. The method demonstrates consistent improvements in reconstruction quality and rate–distortion performance across different calibration settings.

Critical weaknesses

W1. In Equation 8, the theoretical calibration is motivated by minimizing latent-space KL divergence, whereas the practical method optimizes RGB reconstruction loss. Is there a theoretical explanation for this mismatch?

W2. Following W1, I suggest an ablation that replaces DRGBD_{\mathrm{RGB}} with DlatentD_{\mathrm{latent}} in prediction refinement to provide direct empirical evidence of the effects of this objective mismatch.

W3. The main comparison uses the controlled DiffC-MSE baseline rather than the practical DiffC-F configuration. A direct comparison under a consistent evaluation protocol is needed to establish the benefits over the practical baseline.

W4. The results for CoD and Turbo-DDCM, along with CalRCC applied to these methods, should be included in the main rate–distortion comparison.

Questions for authors

Please refer to the Weakness

Reviewer_DQLBOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates how to improve image compression using pretrained diffusion and flow models through calibrating reverse-channel coding (RCC). The authors theoretically decompose the ideal expected RCC cost into required source information and receiver-dependent excess communication, showing that the latter can be reduced through receiver calibration.

Critical strengths
  1. I appreciate the theoretical analysis of RCC communication cost, which decomposes the communication rate into an unavoidable source information term and the cost of receiver mismatch.

  2. The authors develop practical calibrated reverse-channel coding, and further investigate transferring prediction calibration to other diffusion codecs, including CoD and Turbo-DDCM.

Critical weaknesses
  1. For RCC and the line of DiffC research, one of the major problems is the high computational complexity of the algorithm. Although Vonderfecht and Liu (2025) implemented the practical DiffC algorithm and report an acceptable coding latency, it is still much slower than other one-step methods. It seems that CalRCC indeed increases some NFE, as shown in the appendix. The authors should also report the end-to-end latency and the model size after calibration, for a fair evaluation of whether the improved performance is worth the increased latency, and comparison with baselines.

  2. Theoretical decomposition concerns ideal expected communication cost, whereas practical coding uses finite-candidate PFR and incurs additional approximation effects. The author is suggested to discuss this gap explicitly.

Vonderfecht and Liu, “Lossy Compression with Pretrained Diffusion Models,” ICLR, 2025

Questions for authors

See the weakness part.

Reviewer_soTLOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CalRCC, a receiver calibration framework for image compression based on reverse-channel coding (RCC) with pretrained generative models. The authors first decompose the ideal communication cost into the required source information and an excess cost caused by receiver distribution mismatch. Based on this analysis, they introduce three calibration strategies: (1) prediction calibration using LoRA with rate-distortion and perceptual objectives, (2) confidence calibration through a closed-form adjustment of the receiver covariance, and (3) path calibration by optimizing the evaluation schedule along the reverse process.

The proposed method is implemented using a pretrained FLUX model with finite-candidate PFR coding. Experiments on DIV2K, Kodak, and MCL-JCI demonstrate improvements over a controlled DiffC baseline, including approximately 1.5–1.6 dB PSNR gains at matched bitrates and around 50% BD-rate reduction. Additional experiments investigate the individual calibration components and the transferability of prediction calibration to CoD and Turbo-DDCM.

Critical strengths
  1. Clear and well-organized presentation. The paper is generally well written and follows a logical structure, progressing from the theoretical motivation to the three calibration components and experimental evaluation. The main ideas are clearly explained, and the figures help illustrate the proposed framework.

  2. Detailed technical and experimental documentation. The paper provides extensive supplementary materials, including mathematical derivations, implementation details, and experimental protocols. The distinction between ideal KL-based communication costs and actual serialized bitrates is also clearly stated, which helps readers understand and assess the reported results.

Critical weaknesses
  1. Limited methodological novelty and insufficient attribution of improvements. The proposed framework mainly combines existing techniques, including LoRA-based fine-tuning, Gaussian covariance adjustment, and timestep schedule optimization. The communication cost decomposition follows standard KL and mutual information identities, while related scheduling strategies have already been explored in Practical DiffC. More importantly, the paper lacks a key baseline that fine-tunes the pretrained model using only reconstruction loss (without the RCC-specific rate term). Therefore, it remains unclear how much of the reported improvement comes from the proposed RCC-aware calibration rather than conventional reconstruction-oriented fine-tuning.

  2. Insufficient evidence of practical advantages over existing methods. The main performance gains are reported against a controlled DiffC baseline using single-pass reconstruction and a fixed reference schedule, rather than a fully optimized Practical DiffC configuration. Although the appendix includes additional reconstruction comparisons, the overall advantage over stronger configurations remains insufficiently established. Furthermore, path calibration substantially increases the number of model evaluations, while end-to-end encoding/decoding time and computational overhead are not adequately evaluated. These limitations weaken the claim of practical compression improvements.

Questions for authors
  1. Contribution of the RCC-specific training objective. The prediction calibration combines a rate-related KL term with reconstruction losses. However, it is unclear how much of the improvement comes from the RCC-specific objective rather than conventional LoRA fine-tuning. Could the authors provide a controlled comparison with reconstruction-only LoRA training (lambda_R=0), using the same training budget and codec configuration? This would help clarify the necessity and contribution of the proposed rate-aware calibration.

  2. Comparison with a stronger DiffC baseline. The main results are based on a controlled DiffC baseline with single-pass reconstruction and a reference timestep schedule. Although iterative reconstruction is partially evaluated in Appendix F.1, could the authors further clarify how CalRCC compares against a fully optimized Practical DiffC configuration, including its timestep optimization and iterative reconstruction? This comparison is important for assessing the actual improvement over prior work.

  3. Practical computational efficiency. Path calibration substantially increases the number of model evaluations, while the paper mainly reports NFE and candidate counts rather than actual runtime. Could the authors provide end-to-end encoding and decoding times, or a more detailed computational cost comparison with DiffC? This is important for determining whether the bitrate savings justify the additional computational overhead, particularly given the paper's emphasis on practical compression.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.