acceptodds
Under review as a conference paper at ICLR 2027

LOONA: Linearized Objectives over Noise Atoms for Controllable Zero-Shot Diffusion Image Compression

Abstract

Zero-shot diffusion-based codecs compress images by selecting elements from a shared randomness by the encoder and the decoder. Fast methods such as Turbo-DDCM score atoms using latent residual, restricting direct control to the latent-space quadratic objective. Optimizing the zero-shot diffusion-based image compression method on a general class of objective functions has not been fully studied yet. DDCM and Turbo-DDCM explored some compression with regional priority and perception optimizing encoding methods, but those methods don't give a direction toward a scalable general compression method for arbitrary loss function. We introduce LOONA, which enables the efficient selection of noise from any differentiable objectives. LOONA does so by determining the optimal linear function for the given objective, which makes our computational complexity not scale with , avoiding exhaustive evaluation of candidate combinations. Each guided step computes one Vector-Jacobian product, with neural evaluation cost independent of codebook size and the number of selected atoms. LOONA retains Turbo-DDCM's bitstream format and frozen decoder, transmits no additional side information, and requires no retraining. We experiment on objectives as pixel-level fidelity loss, semantic preservation loss and perceptual loss and demonstrate that LOONA successfully optimizes the given objectives. From these various choice of compression objectives, LOONA improves the rate–distortion Pareto-frontier, with improved full-resolution region control, semantic preservation, and a tunable perception-distortion tradeoff.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffusion Models, Image Compression, Noise Selection, Test-time, Controllable

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

This paper proposes LOONA, a controllable encoding framework for zero-shot diffusion image compression. It models per-step noise selection as a linear surrogate of the next-step objective, derives the optimal scoring coefficient via Stein's identity, and approximates it with one backward pass, so that any differentiable objective can guide codebook atom selection without changing the bitstream format or the decoder. The same encoder supports pixel fidelity, full-resolution region-of-interest, semantic preservation, and perception-distortion traversal, achieving substantial BD-rate reductions over Turbo-DDCM.

Critical strengths

Stein's identity rigorously establishes that the optimal linear scoring coefficient equals the noise-averaged gradient, and one backward pass approximates it with a number of network evaluations independent of codebook size and atom count. This elevates a previously heuristic correlation-based score into a principled selection criterion with minimal changes, and the reasoning is clean.

Critical weaknesses
  1. The approximation chain from the theoretical optimal score to the practical gradient lacks quantitative validation. The argument in Appendix C relies on replacing the average of a product by the product of averages and treating the step Jacobian as a scaled identity, both exact only when the loss is nearly affine or single-step curvature is negligible. LPIPS, DINOv2 features, and Re-IQA are strongly nonlinear objectives, yet the paper never measures the resulting approximation error or states when the approximation breaks down.
  2. The method depends on several hyperparameters tuned on the DIV2K training set, including the gradient-start threshold, the number of gradient steps, the loss weights, and the choice of the steering quality metric, and the sensitivity of these choices across datasets and objectives is not clear.
Questions for authors
  1. Can you quantify the approximation error between the deployed gradient and the theoretical optimal score for LPIPS, DINOv2, and Re-IQA?
  2. Under what conditions does replacing the average of a product by the product of averages fail, and does the error vary across bitrates and diffusion steps?
3 comments

We thank the reviewer for the constructive comments. Both approximation steps of Appendix C are now measured (new appendix section E).

Notation. Next state xt−1=μ+σtzx_{t-1}=\mu+\sigma_t z (μ\mu: step mean, zz: injected noise). x^0∣s\hat x_{0\mid s}: denoised estimate at step ss, Jacobian JsJ_s. c(z)c(z): loss gradient ∇L\nabla L at the next estimate x^0∣t−1(μ+σtz)\hat x_{0\mid t-1}(\mu+\sigma_t z); cˉ\bar c: ∇L\nabla L at the average next estimate. E\mathbb E: average over zz.

W1: the approximation chain, measured

Appendix C goes from the ideal coefficient u⋆u^{\star} (Prop. 1) to the deployed gradient gtg_t (Eq. 13) through the product of averages g~\tilde g:

u⋆=σtE[Jt−1⊤c(z)]u^{\star}=\sigma_t\mathbb E[J_{t-1}^{\top}c(z)]

g~=σtE[Jt−1]⊤cˉ\tilde g=\sigma_t\mathbb E[J_{t-1}]^{\top}\bar c

  • (a) product of averages: u⋆≈g~u^{\star}\approx\tilde g;
  • (b) scaled identity: gt=Pt⊤g~/σtg_t=P_t^{\top}\tilde g/\sigma_t by the tower property (Eqs. 12, 21), with Pt=∂μ/∂xtP_t=\partial\mu/\partial x_t close to a multiple of II.

We estimated u⋆u^{\star} and g~\tilde g by Monte Carlo (32 to 256 samples of zz) at every coded step of the deployed Kodak24 encodes. Mean cosines over the gradient steps and images; cells are 0.0128 / 0.0301 bpp; (a) and (b) are plain cosines, biased low.

objectivetotal: gtg_t, u⋆u^{\star}(a): g~\tilde g, u⋆u^{\star}(b): gtg_t, g~\tilde g
LpixelL_{\mathrm{pixel}} (with LPIPS-VGG)0.951 / 0.9640.991 / 0.9760.950 / 0.967
LpixelL_{\mathrm{pixel}} + DINOv20.931 / 0.9470.965 / 0.9600.935 / 0.959
LpixelL_{\mathrm{pixel}} - Re-IQA0.891 / 0.8980.914 / 0.9120.887 / 0.915

Q1: LPIPS, DINOv2, Re-IQA. LOONA uses these terms only inside the deployed objectives above. There the cosine is 0.89 to 0.96, and the deployed target (Eq. 14) obtains 92 to 96% of the next-step loss reduction obtained with u⋆u^{\star}. Alone, which LOONA never uses, the cosine is 0.68 to 0.81, and −gt-g_t still obtains 71 to 82% of that reduction, against -1 to 16% for Turbo-DDCM's residual.

Q2: when does the product of averages fail? The error of (a) is exactly

u⋆−g~=σtCov(Jt−1⊤,c)+σtE[Jt−1]⊤(E[c]−cˉ)u^{\star}-\tilde g=\sigma_t\mathrm{Cov}(J_{t-1}^{\top},c)+\sigma_t\mathbb E[J_{t-1}]^{\top}(\mathbb E[c]-\bar c)

Both terms vanish for affine LL (constant cc) and grow with the curvature of LL times the spread of the next estimate over candidates: (a) is 0.98 to 0.99 for LpixelL_{\mathrm{pixel}} and 0.91 to 0.97 with the DINOv2 or Re-IQA term.

  • Diffusion step. On the earlier coded steps the cosine is 0.92 for LpixelL_{\mathrm{pixel}}. It breaks down only for DINOv2 and Re-IQA alone at the highest noise (t≥869t\ge 869 of 1000), where even u⋆u^{\star} explains only 37 to 41% of the loss variance, as Sec. 3.1 anticipates; LOONA uses the gradient only at t≤390t\le 390.
  • Bitrate. No effect beyond the noise level: at the 22 timesteps shared by both rates the cosine differs by at most 0.01.

End to end. Re-encoding with a Monte Carlo u⋆u^{\star} in place of gtg_t (Kodak24, 3 seeds, same bits). Change against deployed over the three objectives and both rates (LPIPS: lower is better):

samplesPSNR (dB)MS-SSIMLPIPStime
1-0.04 to +0.00-0.003 to -0.001+0.002 to +0.011about 1x
4+0.03 to +0.08+0.001 to +0.004-0.002 to +0.0042.2x to 2.7x
64+0.07 to +0.14+0.002 to +0.009-0.008 to +0.00225x to 35x

One backward pass loses at most 0.14 dB, 0.009 MS-SSIM and 0.008 LPIPS (for LpixelL_{\mathrm{pixel}}: 0.10 / 0.07 dB, against a gain of 0.58 / 0.37 dB over Turbo-DDCM). Neither a one-sample estimate of u⋆u^{\star} nor the Taylor estimator (Table 8) is better than gtg_t. Section E contains more exp.

W2: hyperparameter sensitivity

Tuning is not what produces the gain. In new experiments we replaced the tuned Table 1 by one untuned rule for all rates. PSNR BD-rate against Turbo-DDCM, 3 seeds (Kodak24 / DIV2K-val-100): -18.5% / -17.1% for Table 1, -20.1% / -18.4% for NgN_g = half of the coded steps, -16.0% / -14.8% for one start step (Tg=40T_g=40) at every rate.

DIV2K-val-100 was never used for tuning (development: DIV2K-train crops, six Kodak images).

  • Start threshold and gradient steps are one setting (Ng=Tg−TfreeN_g=T_g-T_{\mathrm{free}}, Table 1). On DIV2K-val-100, every tested NgN_g from 14 to 73% of the coded steps improves PSNR and LPIPS over Turbo-DDCM.
  • Loss weights (same set, 0.0301 bpp). Without any one of its four terms LOONA stays +0.18 to +0.52 dB above Turbo-DDCM with lower LPIPS. Using the Appendix B.2 weights instead of the deployed ones (per-rate values to be added) changes PSNR by 0.03 dB.
  • Steering metric. In Fig. 10, five no-reference metrics are each scored on the metrics they did not optimize; Re-IQA gives the largest held-out gain at both rates.
  • Datasets and objectives. With unchanged settings the PSNR BD-rate on CLIC2020 (new) is -18.4%; the semantic and perceptual objectives reuse the NgN_g and base weights of LpixelL_{\mathrm{pixel}}.

Thank you for the detailed responses. These responses addressed my concerns and I believe this is a worthy submission.

However, when evaluating whether this paper has greater significance, I would like to ask the authors again. Is the strategy of this paper likely to be extended to other diffusion-based image compression paradigms (such as PerCo or post-processing paradigms), or is it only applicable within the DDCM-based methods?

Thank you for the follow-up and for the positive assessment. We believe that our strategy opens a path to making other diffusion-based codecs controllable.

A direct transfer to trained codecs such as PerCo [1], or to post-processing codecs, is not trivial. Prop. 1 and the tower property (Eq. 12) assume Gaussian noise atoms and a denoiser that is the posterior-mean (Tweedie) estimator, whereas a learned codebook is not Gaussian and a jointly trained U-Net may deviate from that estimator.

The strategy itself is not entirely tied to only DDCM: rank the encoder's discrete choices by the optimal linear score of a chosen objective, obtained from one gradient on the denoised estimate. We believe it can be leveraged to design post-processing diffusion-based codecs with per-image control of the objective, a promising direction beyond this paper.

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

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

This paper presents a training-free extension of Turbo-DDCM for controllable diffusion-based image compression. It introduces an objective-dependent linear scoring rule for selecting noise atoms. The authors derive an ideal scoring coefficient using Stein’s identity and approximate it with a gradient through the current denoised estimate. In practice, the normalized descent direction is combined with the normalized latent residual to guide discrete noise selection. The method supports different differentiable objectives while retaining Turbo-DDCM’s bitstream format and decoder, without transmitting additional control information.

Critical strengths

1.The method adapts Turbo-DDCM’s discrete noise selection to different differentiable objectives while retaining its existing decoder and bitstream format. This enables application-specific control over reconstruction quality, regions of interest, and semantic preservation without retraining the diffusion model or transmitting additional control information. 2.The method reports consistent rate–distortion improvements over Turbo-DDCM on multiple datasets.

Critical weaknesses

1.Limited methodological novelty. Although LOONA provides a useful extension, its practical scoring rule is closely related to DDCM’s gradient-based noise selection. Existing methods have already established mechanisms for guiding diffusion-based generation via gradients; it appears that this paper applies such a mechanism to Turbo-DDCM. Consequently, the unique contribution of the proposed method remains to be further substantiated. 2.The paper derives an ideal noise-scoring rule but uses a simplified gradient-based approximation in practice. The paper does not sufficiently demonstrate how closely the practical approximation follows the theoretical rule. This weakens the theoretical justification for the proposed noise-selection method.

Questions for authors

1.How closely does the practical gradient-based score follow the ideal rule derived in Proposition 1? Could the authors provide evidence supporting the accuracy of this approximation under the actual sampling settings? 2.Could the authors further clarify the key technical contributions of LOONA compared with existing gradient-based diffusion guidance methods? In particular, it would be helpful to better understand the advantages of the proposed scoring strategy over directly incorporating gradient guidance into Turbo-DDCM. 3.Could the authors provide additional datasets (e.g., CLIC2020) and comparisons with relevant generative compression methods (e.g., AEIC), under consistent evaluation settings?

3 comments

We sincerely thank the reviewer for the sharp and constructive comments. The new appendix section "Clarification on Differences Compared to DDCM's Gradient" of the revised PDF has the supporting figure and discussion. Below, we address the weaknesses and questions raised by the reviewer (W2 / Q1, Q3 and the references are in the next comment).


W1 / Q2: LOONA versus DDCM's gradient and gradient guidance

DDCM's gradient is not LOONA's: current step versus lookahead. DDCM's gradient rule [1] (their Eq. 9) picks the atom closest to the scaled score of the condition yy at the current state xtx_t:

σt∇xtlog⁡pt(y∣xt)\sigma_t\nabla_{x_t}\log p_t(y\mid x_t)

It is evaluated before the step and ignores what the step produces. For compression the score is proportional to the residual rtr_t (the rule of Turbo-DDCM [2]), with no network gradient. LOONA ranks atoms by the loss after the step (Eq. 7):

L(E[x0∣xt−1])≈L(x^0∣t−1(μθ(xt)+σtz))L(\mathbb E[x_0\mid x_{t-1}])\approx L(\hat x_{0\mid t-1}(\mu_\theta(x_t)+\sigma_t z))

By Stein's identity (Prop. 1), its best linear ranking of the Gaussian atoms is the gradient with respect to the injected noise zz, through the denoiser, averaged over candidate next states:

u⋆=Ez[∇zL(x^0∣t−1(μθ(xt)+σtz))]u^{\star}=\mathbb E_z[\nabla_z L(\hat x_{0\mid t-1}(\mu_\theta(x_t)+\sigma_t z))]

DDCM's gradient exists only where the score is closed-form (otherwise two candidates per step, no gradient: their Eq. 12); LOONA takes any differentiable loss (Eqs. 15-18). Even on DDCM's own objective, latent MSE, the rules differ: LOONA's target is built from Jt⊤rtJ_t^{\top}r_t instead of rtr_t and saves 7.0% (Kodak24) / 6.2% (DIV2K-val-100) of bits at equal PSNR. LOONA is not DDCM's rule with a new loss. In order to best clarify the difference, we added Figure 9. in our revised PDF.

Estimating the lookahead is not trivial. Exact evaluation costs a denoiser call per atom (16384 per step), and a one-sample Monte Carlo estimate of u⋆u^{\star} gains nothing (W2). By the tower property (Eqs. 12, 21), one backward pass through the denoiser at xtx_t estimates u⋆u^{\star}: gtg_t is computed at the current step, yet predicts the effect of an injected noise.

Against gradient guidance (Q2). Guidance is a design space, not one rule [9]: the gradient may pass through the denoiser at xtx_t (DPS [3], FreeDoM [4]), skip the denoiser (MPGD [7]), be smoothed by sampling (LGD [8]) or use recurrence (Universal Guidance [6]). These give different selection targets, and the choice matters: the gradient at the step mean μθ(xt)\mu_\theta(x_t) (the Taylor choice) needs one more denoiser call and lowers PSNR by 0.02 to 0.09 dB (Table 8). Prop. 1 selects the first: u⋆u^{\star} contains the denoiser Jacobian, and without it the latent-MSE target is rtr_t, i.e. Turbo-DDCM. Used directly, even that design falls below Turbo-DDCM in PSNR or LPIPS (LpixelL_{\mathrm{pixel}}, Kodak24; change at 0.0128 / 0.0301 / 0.0563 bpp):

  • −gt-g_t alone on every coded step: -0.56 / -1.18 / -1.21 dB;
  • added to the residual on every coded step: +0.53 / +0.31 / +0.25 dB, LPIPS worse by 0.008 to 0.017;
  • LOONA (residual anchor of Eq. 14, last coded steps only): +0.58 / +0.37 / +0.41 dB, LPIPS better by 0.008 to 0.033.

W2 / Q1: ideal coefficient versus deployed gradient

On the deployed Kodak24 encodes we estimated u⋆u^{\star} by Monte Carlo (64 samples). Mean cosine over the steps using gtg_t; cells: 0.0128 / 0.0301 bpp.

LpixelL_{\mathrm{pixel}}with DINOv2 termwith Re-IQA term
cosine of gtg_t and u⋆u^{\star}0.951 / 0.9640.931 / 0.9470.891 / 0.898

Re-encoding with a Monte Carlo u⋆u^{\star} in place of gtg_t (LpixelL_{\mathrm{pixel}}, 3 seeds, same bitstream length); change against ours:

samplesPSNR (dB)MS-SSIMLPIPSencoding time (ours: 1x)
1-0.01 / +0.00-0.0022 / -0.0012+0.011 / +0.0021x / 1x
4+0.05 / +0.04+0.0020 / +0.0012+0.004 / -0.0012.2x / 2.7x
32+0.09 / +0.07+0.0049 / +0.0018+0.001 / -0.00216x / 20x
64+0.10 / +0.07+0.0054 / +0.0019+0.002 / -0.00226x / 34x

The PSNR gap saturates at 0.10 / 0.07 dB, against a gain of 0.58 / 0.37 dB over Turbo-DDCM, and LPIPS (lower is better) ends within 0.002: one backward pass (6 s per image) keeps about 85% of the ideal PSNR gain. These measurements are in the new appendix section "Measured Approximation Error of the Deployed Gradient".


Q3: CLIC2020 and AEIC [5]

CLIC2020 test set (428 images), protocol of Table 5, codebook seed 42, official AEIC checkpoints. BD-rate in % against Turbo-DDCM:

methodtrainingPSNRMS-SSIMLPIPS
LOONAnone-18.4-14.3-13.8
AEIC-MEtrained-14.3-18.6-42.6
AEIC-SEtrained+9.7-6.0-37.3

AEIC (one model per rate) leads on LPIPS and patch-FID; LOONA (one frozen model) on PSNR, also on Kodak24 and DIV2K-val-100 (-18.6 / -17.3 against -6.4 / -14.9 for AEIC-ME). Q3 will be added to the paper.


References

[1] Guy Ohayon, Hila Manor, Tomer Michaeli, Michael Elad. Compressed Image Generation with Denoising Diffusion Codebook Models. ICML 2025.

[2] Amit Vaisman, Guy Ohayon, Hila Manor, Michael Elad, Tomer Michaeli. Turbo-DDCM: Fast and Flexible Zero-Shot Diffusion-Based Image Compression. ICLR 2026.

[3] Hyungjin Chung, Jeongsol Kim, Michael T. McCann, Marc L. Klasky, Jong Chul Ye. Diffusion Posterior Sampling for General Noisy Inverse Problems. ICLR 2023.

[4] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, Jian Zhang. FreeDoM: Training-Free Energy-Guided Conditional Diffusion Model. ICCV 2023.

[5] Tianyu Zhang, Dong Liu, Chang Wen Chen. Ultra-Low Bitrate Perceptual Image Compression with Shallow Encoder. CVPR 2026.

[6] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, Tom Goldstein. Universal Guidance for Diffusion Models. ICLR 2024.

[7] Yutong He, Naoki Murata, Chieh-Hsin Lai, Yuhta Takida, Toshimitsu Uesaka, Dongjun Kim, Wei-Hsiang Liao, Yuki Mitsufuji, Zico Kolter, Ruslan Salakhutdinov, Stefano Ermon. Manifold Preserving Guided Diffusion. ICLR 2024.

[8] Jiaming Song, Qinsheng Zhang, Hongxu Yin, Morteza Mardani, Ming-Yu Liu, Jan Kautz, Yongxin Chen, Arash Vahdat. Loss-Guided Diffusion Models for Plug-and-Play Controllable Generation. ICML 2023.

[9] Haotian Ye, Haowei Lin, Jiaqi Han, Minkai Xu, Sheng Liu, Yitao Liang, Jianzhu Ma, James Zou, Stefano Ermon. TFG: Unified Training-Free Guidance for Diffusion Models. NeurIPS 2024.

We sincerely thank the reviewer for carefully considering our response and for raising the score. The questions on the relation to DDCM's gradient and to gradient guidance made the positioning of our method much clearer. Thank you for your time and effort.

Reviewer_1MMSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes LOONA, which builds on Turbo-DDCM. LOONA changes the noise scoring rule to combine the original residual with the negative gradient of the loss on decoded images. The method keeps the original bitstream format and frozen decoder. It adds backpropagation at selected sampling steps. It supports region prioritization, semantic preservation, and control over perceptual quality.

Critical strengths
  • The authors use loss gradients from decoded images to select noise. This lets the encoder directly account for the quality of the reconstructed image. The idea is straightforward and makes sense.
  • The writing and presentation of the paper are excellent, with a clear and logical flow.
Critical weaknesses
  • Gradient computation adds noticeable latency compared with Turbo-DDCM. Figure 8 shows a longer round-trip time even at low bitrates.
  • Faces and text are difficult to preserve in image compression. Figure 2 shows better face reconstruction with LOONA. Please also show its performance on text regions.
  • The authors use different starting steps and numbers of gradient steps for different bitrates. However, Figure 10 shows that more gradient steps do not always improve quality. Do these settings need to be tuned again for different bitrates or images from different scenes? How does the method generalize across different cases?
Questions for authors

See weakness.

2 comments

We thank the reviewer for the thorough, constructive feedback and the positive assessment of our work. References [1-3] are included in the comment.

W1: latency

At matched encoding time, LOONA is better than Turbo-DDCM [1]. Kodak24, 0.01227 bpp for all three rows:

encoding timePSNR (dB)LPIPS
Turbo-DDCM, released (T=30, M=50)0.90 s20.100.334
Turbo-DDCM, T=18, 8 coded steps, M=360.58 s20.500.349
LOONA, schedule of row 2 + 2 gradient steps0.82 s20.630.319
  • Against the released settings, LOONA encodes faster and is better in PSNR and LPIPS.
  • Row 2 isolates the gradient: on the same schedule, 2 gradient steps add 0.24 s and give +0.13 dB and LPIPS -0.030. LOONA is the best of the three in both metrics.

During the discussion period we will add matched-time results at more bitrates. In the setting of Fig. 8 (T=60 for both codecs; one A100, median per image), the extra time is confined to the encoder:

  • Encoding, Turbo-DDCM / LOONA: 3.26 / 5.25 s at 0.0128 bpp and 3.60 / 9.99 s at 0.1164 bpp, about 0.22 s per gradient step.
  • **Decoding is identical. ** Median difference is within 0.08 s at every rate.
  • The cost is a knob. NgN_g is chosen by the encoder; the decoder does not need it. At 0.0301 bpp, Ng=6N_g=6 adds 1.25 s and already gives +0.24 dB and LPIPS -0.014 over Turbo-DDCM.

W2: text regions

COCO-Text [2]: 100 images, 364 word boxes. A fixed recognizer (PaddleOCR [3], PP-OCRv4 English model) reads each ground-truth box on the decoded image. Retention: share of the 238 words read correctly on the original that are still read correctly. Visual results: Appendix Figs. 25 and 26.

Without a region map (the setting of Fig. 2), same bits; cells are Turbo-DDCM / LOONA:

bppretentionPSNR in boxes (dB)PSNR full image (dB)
0.01280.08 / 0.1416.68 / 17.2619.33 / 19.94
0.03010.27 / 0.3219.10 / 19.7221.72 / 22.25

LOONA gains 0.6 dB in the text boxes, as in the whole image, and reads more words at both rates.

With a region map on the text boxes (foreground weight 4 for both codecs). Turbo-DDCM's priority map reweights the latent residual at 64x64. LOONA keeps this map and adds the region loss of Eq. 16, at full pixel resolution. Cells are Turbo-DDCM / LOONA:

bppretentionPSNR in boxes (dB)PSNR outside boxes (dB)
0.01280.35 / 0.4018.36 / 18.9918.58 / 19.14
0.03010.51 / 0.5520.56 / 21.3221.29 / 21.80
  • At 0.0301 bpp LOONA retains 0.55 of the words at a full-image PSNR of 21.73 dB; Turbo-DDCM without a map has the same PSNR (21.72 dB) and retains 0.27.
  • Over all weights tested, no Turbo-DDCM point is Pareto-optimal (retention, background PSNR) at 0.0301 bpp; at 0.0128 bpp only its largest weight is, with a 12.0 dB background.
  • The region loss alone (no residual map) retains 0.50 / 0.24 at 0.0301 / 0.0128 bpp: the two are complementary.

W3: gradient steps, re-tuning and generalization

Why more gradient steps do not always help (Fig. 11; Fig. 10 of the original submission). LOONA ranks atoms by the residual rtr_t (source latent minus denoised estimate; Turbo-DDCM's target) on every coded step and adds the loss gradient only on the last NgN_g (Eq. 14). In Fig. 11, and in its repeat on DIV2K-val-100 (never used for tuning), the best NgN_g lies in between, at one third to three quarters of the coded steps:

  • Too few: the gradient carries the objective, which the residual lacks, so each added step helps.
  • Too many: the added steps are the earliest, noisiest ones, where the denoised estimate is a heavily smoothed posterior mean and the loss gradient on it is a poor guide for selection, while the residual stays informative (Sec. 3.1). The gradient is also least aligned with the residual there (Fig. 19, 0.0301 bpp): their cosine similarity is 0.35 to 0.5 on the first coded steps and rises to 0.6 to 0.75 on the steps LOONA uses.

This trade-off comes from the noise level of the diffusion steps, not from the image content, so it generalizes. Across bitrates, the best NgN_g scales with the number of coded steps: Table 1 uses 36 to 61% of them at all nine rates, and in a new experiment the single rule NgN_g = half of the coded steps is slightly better than Table 1 in PSNR and LPIPS BD-rate (Kodak24, DIV2K-val-100). Across objectives, the semantic and perceptual objectives reuse the NgN_g of the pixel objective.

No re-tuning for different images. The settings are fixed per bitrate, never per image: Table 1 (Appendix B.1) has one setting per bpp (start step and NgN_g are tied: Ng=Tg−TfreeN_g=T_g-T_{\mathrm{free}}), used for every image of every dataset. With Table 1 unchanged, PSNR improves on at least 98.6% of the (image, seed) pairs at every rate (Kodak24, DIV2K-val-100), and LOONA also improves on CLIC2020 and COCO-Text (new in this rebuttal).

References

[1] Amit Vaisman, Guy Ohayon, Hila Manor, Michael Elad, Tomer Michaeli. Turbo-DDCM: Fast and Flexible Zero-Shot Diffusion-Based Image Compression. ICLR 2026.

[2] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, Serge Belongie. COCO-Text: Dataset and Benchmark for Text Detection and Recognition in Natural Images. arXiv:1601.07140, 2016.

[3] Yuning Du, Chenxia Li, Ruoyu Guo, Xiaoting Yin, Weiwei Liu, Jun Zhou, Yifan Bai, Zilin Yu, Yehua Yang, Qingqing Dang, Haoshuang Wang. PP-OCR: A Practical Ultra Lightweight OCR System. arXiv:2009.09941, 2020. We use the PP-OCRv4 English recognition model of the PaddleOCR toolkit (release 2.8.1, https://github.com/PaddlePaddle/PaddleOCR); PP-OCRv4 is documented in the toolkit and has no separate paper.

Comments on the paper

1 comment

Summary of the revision

We thank all reviewers for their constructive feedback. The revised PDF contains the following additions:

  • Relation to DDCM's gradient and to gradient guidance (Reviewer CCvM): new Appendix D, "Clarification on Differences Compared to DDCM's Gradient".
  • Approximation error of the deployed gradient (Reviewers CCvM and puTR): new Appendix E, "Measured Approximation Error of the Deployed Gradient", with the measured alignment and the end-to-end comparison against the ideal coefficient.
  • Text regions (Reviewer 1MMS): new qualitative results on COCO-Text (Figs. 25 and 26).

The additional experiments requested by the reviewers (CLIC2020 with AEIC, matched encoding time, text-region recognition, hyperparameter robustness) are reported in the individual responses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.