acceptodds
Withdrawn from ICLR 2027

Thinking-then-Compress with Perception-Informed Spatial Budgeting for Generative Image Compression

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generative image compression introduces a generative model at the decoder to fill in details missing from the bitstream using image priors, and has become a prominent approach for perceptual compression at low bitrates. However, when a large generative decoder cannot be jointly optimized with the encoder, the encoder's compression decisions are no longer directly adapted to the decoder's reconstruction behavior. To address this issue, we propose Thinking-then-Compress, a closed-loop image compression paradigm instantiated as encoder-side search over transmitted discrete spatial schemes using reconstruction feedback from a fixed deployed decoder. The encoder first analyzes perceptually important content, proposes a regional scheme, evaluates the resulting reconstruction, and revises the scheme within a fixed feedback budget. Within this paradigm, Perception-Informed Spatial Budgeting proposes regional tier updates from image importance, codec error, and observed generative recovery, then verifies joint proposals using reconstruction feedback and their rate cost. Reconstruction-Aware Generative Enhancement further fine-tunes the generation branch and generation-side adapters for compressed inputs and uses reconstruction feedback to compose codec and generative outputs spatially, preserving reliable codec details while introducing generation where it improves recovery. Under the reported evaluation protocol, our method achieves favorable performance against existing perceptual image compression methods. Our project will be publicly available after publication.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Thinking-then-Compress, Perceptual Compression, Generative Model, Closed-Loop

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ZT4qOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a closed-loop, encoder-side optimization approach for generative image compression. The key idea is to make the encoder revise its compression scheme by observing provisional reconstructions from a fixed decoder. First, the input image is divided into regions, and regional compression-tier updates are proposed based on regional importance, codec error, and generative recovery; whether to accept each update is then decided iteratively based on provisional reconstructions from the decoder and their rate cost. Second, the codec and generative reconstructions are combined with regional mixing weights, which are selected at the encoder and transmitted. The system builds on DLF as a fixed base codec, followed by BAGEL, a fine-tuned large generative model used as a generative enhancer. The method is compared with several generative codecs, including DLF, on CLIC 2020 and DIV2K using PSNR, MS-SSIM, DISTS, TOPIQ-FR, MS-SWD, and FID.

Critical strengths

None

Critical weaknesses

W1: Recent generative codecs already train the encoder jointly with a pretrained generative model (e.g., using LoRA adapters) and achieve strong performance without particular difficulty [1–3]. In contrast, this paper keeps the encoder fixed and optimizes only the coding scheme by running the generative decoder repeatedly at the encoder side. This approach clearly increases encoder-side complexity. However, it is not clear what advantage this approach offers over joint training that would justify this cost. Therefore, the proposed approach is difficult to view as a significant contribution.

W2: It is unclear what significant advance the proposed method achieves over existing work. In the ablation (Fig. 5), removing each component changes DISTS only marginally (e.g., by 0.001–0.003 for the feedback-based scheme revision). Compared with DLF, the method also shows similar DISTS and FID around 0.01 bpp, despite using a high-complexity generative model (Fig. 3). In addition, all experiments are conducted only on a single combination of models (DLF + BAGEL), so it is unclear whether the proposed approach generalizes to other codecs or generative models.

References

[1] T. Zhang, X. Luo, L. Li, and D. Liu. StableCodec: Taming one-step diffusion for extreme image compression. In ICCV, 2025.

[2] N. Xue, Z. Jia, J. Li, B. Li, Y. Zhang, and Y. Lu. One-step diffusion-based image compression with semantic distillation (OneDC). In NeurIPS, 2025.

[3] T. Zhang, D. Liu, and C. W. Chen. Ultra-low bitrate perceptual image compression with shallow encoder (AEIC). In CVPR, pp. 12118–12128, 2026.

Questions for authors

Given the current manuscript, I do not have questions whose answers could materially change my assessment. The critical weaknesses (W1 and W2) would require additional comparisons and analyses beyond what can be clarified in a rebuttal.

Nevertheless, I believe the core idea of encoder-side optimization itself, which allocates fewer bits where the generative decoder already reconstructs well and more bits where it does not, can be a meaningful research direction. If its advantage over joint training is clearly shown, and the approach is validated beyond a single combination of codec and generative model, this work could develop into a stronger paper.

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

This paper proposes Thinking-then-Compress, an encoder-side search framework for generative image compression. Using a fixed DLF codec and an adapted BAGEL enhancer, it selects regional compression tiers based on semantic importance, reconstruction error, and generative feedback. It additionally transmits quantized regional weights to blend codec and generative reconstructions. Experiments on CLIC 2020 and DIV2K report improved perceptual rate–quality performance, supported by component and search-budget ablations.

Critical strengths

1.Encoder-side spatial allocation and reconstruction blending adapt compression to a fixed enhancer, with side information included in the bitrate.

2.Results on two datasets, together with component and search-budget ablations, support improved perceptual rate–quality performance.

Critical weaknesses

1.Unclear search contribution. Equal-budget greedy/random search comparisons and an ablation removing semantic importance are missing, leaving the proposed mechanism’s contribution unclear.

2.Unvalidated proxy. Decisions optimized using 8-step reconstructions are applied to 80-step decoding without validating their consistency.

3.Missing relevant comparisons. Region Adaptive Diffusion and CADC are cited but not evaluated, limiting the evidence of superiority over existing adaptive methods.

4.Incomplete cost reporting. End-to-end latency, peak memory, and actual generator-call counts including refreshes are not reported.

5.Insufficient reproducibility details. The training dataset, batch size, and tier-anchor calibration procedure remain underspecified.

6.Limited generality evidence. Evaluation uses only DLF+BAGEL and DLF-specific compression tiers, leaving applicability to other model combinations uncertain.

Questions for authors

1.At comparable bitrates, how does performance change with equal-budget simple search or after removing semantic importance?

2.How consistently do 8-step candidate rankings and predicted gains transfer to 80-step decoding?

3.What are the measured encoding/decoding latency, peak memory, and total generator-call counts, including all refreshes?

Reviewer_99SuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces "Thinking-then-Compress," a closed-loop coding framework designed for generative image compression at extremely low bitrates. The framework consists of two modules: "Perception-Informed Spatial Budgeting," which employs iterative searching at the encoder side to achieve spatially adaptive bitrate allocation; and "Reconstruction-Aware Generative Enhancement," which fine-tunes a pre-trained generative model to improve the reconstruction quality of compressed images, while dynamically optimizing the fusion weights between traditional codec outputs and generative outputs based on regional reconstruction quality.

Critical strengths

1.The research motivation is sound. The paper addresses the challenge of jointly optimizing generative and compression models, and proposes an approach based on a generative feedback loop. 2.The methodology is well-conceived. The proposed method achieves competitive perceptual compression performance on standard benchmark datasets.

Critical weaknesses

1.Limited technical novelty. The proposed method primarily combines existing models and techniques into a relatively complex, engineering-oriented framework, without introducing sufficiently distinctive technical innovations. 2.Insufficient justification for iterative feedback-based search. Although the authors argue that jointly training large generative models is computationally expensive, they do not adequately explain why simpler alternatives, such as optimizing only the encoder or a lightweight rate-allocation network with a frozen generative model, would be insufficient. The lack of comparisons with these alternatives makes it difficult to justify the necessity and advantages of the proposed iterative search. 3.Limited performance gains and insufficient experimental validation. The proposed method achieves only modest perceptual improvements over DLF while introducing substantial computational overhead. However, the paper lacks runtime measurements, and sufficient ablations to isolate the contributions of generative model fine-tuning and feedback-based search. Comparisons with more advanced generative compression methods are also insufficient.

Questions for authors

1.What are the key advantages of the proposed iterative feedback-based search over simpler one-shot rate-allocation methods, such as training a lightweight allocation network with a frozen generative model? Can the authors provide experimental evidence supporting the necessity of iterative search? 2.Can the authors report the actual encoding and decoding runtimes and computational overhead compared to DLF? Given the relatively modest performance improvements, how can the additional computational cost be justified? 3.The paper lacks an ablation on generative model fine-tuning. Could the authors compare the pretrained and fine-tuned BAGEL models to better clarify the contributions of these components?

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

Related papers

Discussion (0)

Sign in to comment.