{"id":"63faee13-a7a8-4627-9b81-c67dad065c57","arxiv_id":"2505.20984","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A single rate-variable generative compression model treats quantization as a forward corruption and reverses it with a two-step denoiser, outperforming prior generative codecs on perceptual quality benchmarks.","lead":"This paper presents a generative image compression model that treats rate-variable quantization as a diffusion-like forward process and trains a network to reverse it. It reconstructs photo-realistic images with only two reverse steps and one model spanning multiple bit rates, showing better perceptual metrics than prior generative compression methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The score identity in Eq. 9 relies on Gaussian Tweedie, but the corruption is uniform noise, so the sampler has no proven theoretical basis.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the score formula for uniform quantization noise is asserted without proof. My stress-test confirms that this is not merely an omitted derivation; the claimed identity is mathematically false for the corruption model the paper itself specifies. Because Eq. 10, Eq. 11, and Eq. 12 all depend on this score formula, the theoretical justification for the sampler fails. However, the paper's empirical contribution—a two-step generative compression pipeline with strong perceptual metrics—does not stand or fall on the proof of this identity. The method could still be a useful heuristic, and the experiments may be reproducible. Therefore the appropriate verdict is unchanged from the reader's CONDITIONAL: the theoretical claims must be corrected or explicitly softened before the score-based reversal narrative is accepted, and the hyperparameter β should be selected on a validation split rather than the test set. I do not see an additional concern that would move the verdict to REJECT, because the central empirical results are presented in sufficient detail to be checked independently, and an honest re-framing of the sampler as a heuristic denoising update would preserve most of the paper's practical value.","tokens_in":14299,"tokens_out":4556,"duration_ms":58990,"concrete_test":"Analytical check: in one dimension, take p_data = (1/2)δ_{−1} + (1/2)δ_{1} and q = 1. Compute p(x; q) as the convolution with U(−0.5, 0.5); it is piecewise constant on [−1.5, 1.5], so ∇_x log p(x; q) = 0 in the interior of each piece. Compute the posterior mean m(x) = E[x_0 | x] and compare m(x) − x over q with the true score; the two disagree by O(1) over most of the support. If the identity fails for this simple distribution, it is not a general score formula. A second, in-pipeline check: train the same D_θ on synthetic data from a known density corrupted by U(−q/2, q/2) for several q, then compare (D_θ − x)/q against a finite-difference estimate of ∇_x log p(x; q); an O(1) mismatch on interior points confirms that the sampler's theoretical premise is absent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the score formula in Section 3.2, immediately after Eq. 9: ∇_x log p(x; q(t)) = (x̂_0 − x_t)/q(t). This is the Gaussian Tweedie identity, but Eq. 9 defines the corruption as additive uniform noise, x_t = x_0 + U(−0.5, 0.5) · q(t). For uniform noise, the score is not the posterior-mean difference divided by q. A minimal counterexample is p_data = δ_0 with q = 1: the corrupted density p(x; q) is uniform on [−0.5, 0.5], its score is zero almost everywhere, but the proposed formula gives −x, which is nonzero inside the support. The derivative of the uniform kernel is a pair of boundary delta functions, not a multiplicative linear factor; no Tweedie-style relation connects the MMSE denoiser to the score. Since D_θ is trained by L2 minimization (Eq. 15), its optimum is the posterior mean E[x_0 | x_t], not the score. Consequently, Eq. 10 (the probability-flow ODE), Eq. 11 (the Euler sampler), and Eq. 12 (the stochastic sampler) inherit an unproven—and in fact false—assumption. The update may still work as a heuristic denoising or DDIM-like step, and the empirical results are not invalidated by this alone, but the paper's central claim that the reverse network reverses the compression process through score estimation is not established. The test-set tuning of β is a secondary methodological concern; the score identity is the primary correctness risk.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a diffusion-style generative image compression framework in which rate-variable quantization is reinterpreted as a forward diffusion process. A reverse network is trained with an L2 loss to invert quantization, and an ODE/SDE sampler is derived to reconstruct images from the compressed latent. Experiments on DIV2K, Kodak, and CLIC2020 report competitive perceptual metrics (LPIPS, FID, KID, MUSIQ, CLIPIQA) with only two reverse steps, and the method is claimed to outperform prior generative compression approaches without GAN or perceptual-loss finetuning.","tokens_in":14742,"tokens_out":6219,"duration_ms":69214,"significance":"If the theoretical framework were sound, this would be a meaningful contribution: it combines rate-variable compression with a minimal-step generative decoder and demonstrates strong perceptual results. The empirical study is broad, includes several benchmarks and metrics, and the design of the reverse network as a denoiser conditioned on the quantization scale is interesting. However, the central theoretical claim—that the learned network estimates the score of the corrupted feature distribution—is not established and, in the form stated, is false for the uniform quantization noise used in the forward process. The paper may still describe a useful heuristic denoiser, but the 'estimating gradients' framing and the SDE derivation do not hold as written.","major_comments":[{"comment":"The score identity ∇_x log p(x; q(t)) = (D_θ(x,q)-x)/q(t) is not valid for additive uniform quantization noise. For a data distribution p_data = δ_0 and q = 1, the corrupted density p(x; q) is uniform on [−0.5, 0.5], so the score is zero on the interior of the support, while the proposed formula gives −x, which is nonzero. The L2-trained denoiser D_θ (Eq. 15) estimates the posterior mean E[x_0 | x_t], not the score; for uniform noise there is no Tweedie-style identity connecting the MMSE denoiser to the score. Consequently, the Euler update in Eq. (11), and the reverse ODE/SDE steps derived from it, do not follow from the stated forward process.","section":"Sec. 3.2, Eq. (9)-(11)"},{"comment":"The probability-flow ODE dx = −(dq/dt) ∇_x log p(x; q(t)) dt is asserted without derivation. In standard score-based diffusion, this ODE is derived from the Fokker–Planck equation of a Gaussian SDE. Here the forward process is defined by Eq. (6)/(9) as quantization or additive uniform noise, and no such derivation is provided; the claim that the ODE preserves the marginal distributions p(x; q(t)) is an unproven assumption rather than a consequence of the construction.","section":"Sec. 3.2, Eq. (10)"},{"comment":"The stochastic sampler parameters are tuned on the evaluation benchmark. The randomness form ω and the coefficient β are selected by evaluating on DIV2K (Fig. 3), and the same DIV2K set is used for the main comparison in Fig. 2. This test-set tuning undermines the validity of the reported DIV2K superiority and makes the 'SOTA' claim hard to interpret. Additionally, the final model description does not state whether Gaussian or uniform noise is used for the reported results, so the experiments are not fully reproducible.","section":"Sec. 4.2, Fig. 3 and Eq. (13)"},{"comment":"There is a clear implementation error in the displayed pseudocode: line 10, ¯y_{i+1} ← α(ϵ_i − d_i), overwrites the Euler update from line 8 instead of adding the stochastic term to it. The update should be ¯y_{i+1} ← ¯y_{i+1} + α(ϵ_i − d_i). This obscures the actual sampling procedure and should be corrected, along with a description of how the continuous Eq. (12) is discretized into lines 7–10.","section":"Algorithm 1, line 10"}],"minor_comments":[{"comment":"The relation between the quantization scale q(t) and the time variable t is not stated in the main text; the substitution q(t) := t appears only in Appendix A. This should be clarified in Sec. 3.2, since the notation q(t_i) − q(t_{i+1}) in Eq. (11) depends on that choice.","section":"Sec. 3.2 and Appendix A"},{"comment":"The definition of q_min as 'lower than the minimum support' is vague; it should be specified numerically or in terms of the entropy model's supported quantization range.","section":"Sec. 4.1"},{"comment":"The text contains a typo: 'minimus' should be 'minimum'.","section":"Eq. (13)"},{"comment":"The statement 'without requiring Gaussian noise initialization' is confusing because the stochastic sampler in Eq. (12) and Algorithm 1 injects noise during decoding. The intended meaning is that the initial state is the compressed latent, not pure Gaussian noise; this should be stated explicitly.","section":"Abstract and Sec. 3.2"},{"comment":"The phrase 'The code is available here' is followed by no URL or link; provide the repository address or a supplementary archive for reproducibility.","section":"Sec. 1 and reproducibility"}],"recommendation":"reject","confidential_remarks":"The paper reports strong empirical results, and the iterative denoising decoder could be a useful contribution if reframed as a heuristic post-processing or conditional diffusion method. However, the central theoretical claim—the score formula and the SDE derivation—is false as stated and cannot be repaired by a local correction within the current manuscript. A resubmission that abandons the score-based interpretation and instead presents the sampler as an empirical denoiser schedule, with proper validation-set tuning, would be more appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The empirical work is genuinely useful: a single multi-rate model whose decoder is a two-step reverse network that beats fixed-rate GIC baselines on LPIPS, FID, KID and no-reference scores, with code and latency numbers. The framing—compression as a forward diffusion path with quantization scale as the noise level, and reversing that path directly instead of using diffusion as a post-processor—is a real twist and a plausible path to practical diffusion codecs.\n\nThe theory, though, does not hold. In Sec. 3.2 the paper defines the forward corruption as xt = x0 + U(−0.5,0.5)·q(t), trains an L2 denoiser Dθ, and then writes ∇ log p(x; q) = (Dθ(x,q)−x)/q. That is not a score identity for uniform noise; it is the Gaussian Tweedie relation with the wrong denominator and the wrong noise model. The counterexample from the stress-test note is clean: for pdata = δ0, the corrupted density is uniform on [−0.5,0.5]q, whose score is zero almost everywhere, while the proposed formula is nonzero. So Eq. 10, 11, and 12 inherit an unproven—in fact false—assumption. The sampler may still work as a heuristic denoising/DDIM-like update, and the empirical results are not necessarily invalid, but the central claim that the reverse network legitimately estimates a score is not established.\n\nWhat is new is mostly the combination. The components—rate-variable entropy model, L2-trained denoiser, Euler sampling, EDM-style U-Net—are existing tools; the paper's contribution is the specific construction, not a new theoretical object. That is okay, but it means the SOTA claim depends on a correct theory or at least an honest 'heuristic' framing.\n\nSoft spots, in order: (1) the false score identity, load-bearing; (2) β and the noise form are selected on the DIV2K test set and the same benchmark is used for the headline comparisons, so the reported SOTA numbers are not fully independent; (3) the stochasticity schedule is justified only empirically. The citation pattern is fine, including the self-citations to the authors' own entropy model and EDM; no issue there.\n\nBottom line: this deserves a serious referee, not a desk reject. A good review should require either a correct derivation of the score relation (or an explicit statement that the update is a heuristic denoiser) and re-tuning β on a validation split before the SOTA claim is accepted. I would read the revised version.","headline":"A genuinely useful multi-rate generative codec with a plausible and promising empirical design, but the central score identity is asserted without proof and is false for uniform quantization noise, so the theoretical claims and the SOTA numbers should be treated as provisional.","tokens_in":15166,"tokens_out":3199,"would_cite":true,"duration_ms":34918,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Compression itself can be the forward diffusion, so one reverse network handles all bit rates in two steps.","keywords":["generative image compression","score-based diffusion","stochastic differential equations","rate-variable quantization","quantization scaling","learned image compression","reverse diffusion","perceptual reconstruction"],"falsifier":"Estimate $\\nabla \\log p(x; q)$ by Monte Carlo for a fixed latent distribution corrupted by $U(-0.5, 0.5)\\cdot q$ and compare it with the paper's proxy $(D_\\theta(x,q) - x)/q$; a systematic discrepancy would indicate that the Euler update in Eq. (11) is not reversing the forward process, breaking the claimed link between the compression schedule and the score.","tokens_in":14113,"feed_emoji":"🖼️","tokens_out":13432,"duration_ms":113165,"temperature":0.7,"pith_summary":"This paper tries to establish that the lossy-compression pipeline itself can serve as the forward corruption process of a diffusion model, so that a single reverse network can undo compression at any chosen bit rate and produce photo-realistic images. In this framing, rate-variable quantization is modeled as adding uniform noise of width $q(t)$, and the task of reversing that noise is treated as a score-matching problem. The authors train a denoiser to estimate the score of the rate-variable feature distribution and use it in an Euler reverse sampler, adding a small stochastic term to improve perceptual quality. If the claim is right, generative image compression no longer needs a separate denoising pass after decoding; the reverse network is the generative model, and only two reverse steps are needed for any rate.","feed_headline":"Compression as diffusion: one model, all rates, two steps","feed_subtitle":"Quantization becomes the forward noise; a reverse network restores photo-realistic images at any rate in two steps","key_machinery":"The central object is the rate-variable compression forward process: scaling a latent $y$ by $q$ before quantization and rescaling after produces $y + U(-0.5, 0.5)\\cdot q$, so the quantization scale $q(t)$ plays the role of diffusion time. The reverse machinery is the ODE $dx = -(dq/dt)\\nabla \\log p(x; q(t))\\,dt$, discretized by Euler's method, with the score replaced by the trained denoiser through $\\nabla \\log p(x; q(t)) \\approx (D_\\theta(x, q) - x)/q$. This substitution is what carries the argument: it converts the compression-rate schedule into a reverse-time trajectory from compressed to original features, and the stochastic sampler in Eq. (12) with $\\alpha_t = \\beta\\sqrt{q_t - q_{\\min}}$ injects controlled randomness to improve perceptual fidelity.","core_discovery":"The paper claims that compression, implemented as quantization scaling of a latent feature, is a legitimate forward diffusion path described by a stochastic differential equation, and that reversing that path with an estimated score function yields the reconstructed image directly. The forward process is approximated by $x_t = x_0 + U(-0.5, 0.5)\\cdot q(t)$, the reverse ODE is $dx = -(dq/dt)\\nabla \\log p(x; q(t))\\,dt$, and the trained denoiser $D_\\theta$ supplies the score through $\\nabla \\log p(x; q(t)) = (D_\\theta(x_t, q_t) - x_t)/q_t$. Discretizing with Euler's method gives the update $x_{i+1} = x_i + \\frac{q_i - q_{i+1}}{q_i}(D_\\theta(x_i, q_i) - x_i)$, and adding a noise term with amplitude $\\beta\\sqrt{q_t - q_{\\min}}$ yields the stochastic sampler. The reverse network is trained with a plain L2 reconstruction loss against the original latent, without adversarial or perceptual losses, and the authors report that their method outperforms existing generative compression approaches on perceptual distortion, statistical fidelity, and no-reference quality metrics across standard benchmarks.","pith_inferences":["The score formula $(D_\\theta(x,q) - x)/q$ is proven for Gaussian noise, not for uniform noise; a variant using the exact score of a uniform-noise distribution would test whether the sampler's theoretical basis matters for the reported results.","Because the entropy model already estimates the distribution of quantized features, one could combine that learned prior with the estimated score to compute an exact likelihood, enabling rate control under an information-theoretic objective.","The randomness-injection amplitude $\\alpha_t = \\beta\\sqrt{q_t - q_{\\min}}$ is set empirically; a principled derivation from the reverse SDE's stationary distribution would turn this heuristic into a testable schedule.","The method's two-step decoding suggests the effective forward corruption is mild relative to Gaussian diffusion, so the same reverse network might transfer to other bounded-noise corruption processes, such as dropout or JPEG-style quantization."],"forward_implications":["A single trained model covers all bit rates by varying the quantization scale $q$, removing the need to retrain separate models per rate.","Only two reverse steps are needed at decode time, making the diffusion-based decoder faster than prior diffusion-based compression systems.","No GAN or perceptual-loss finetuning is required; the score-based reverse alone is claimed to achieve competitive perceptual quality.","The framework applies to any learned compression pipeline that can be expressed as additive uniform quantization noise.","The rate-distortion curve becomes continuously adjustable through the quantization scale, enabling smooth rate control in deployment."],"supporting_citations":[{"why":"supplies the score-based SDE/ODE framework that the paper adapts to compression as forward corruption","marker":"[40]"},{"why":"provides denoising score matching, the training objective used to learn the score of corrupted features","marker":"[39]"},{"why":"establishes the VAE-based learned compression framework and the uniform-noise model of quantization","marker":"[5]"},{"why":"gives the end-to-end rate-distortion optimization and the uniform noise formulation for quantization scaling","marker":"[4]"},{"why":"supplies the reverse network architecture and the stochastic-sampler design that the paper adopts","marker":"[21]"},{"why":"provides the entropy model architecture used for the rate-variable compression stage","marker":"[15]"},{"why":"supplies the pretrained latent autoencoder whose feature space hosts the compression and reverse processes","marker":"[36]"}],"fun_headline_variants":["Diffusion-based compression: one model, any rate, two steps","Reimagining compression as diffusion: smooth rates, minimal steps","Score-estimated reverse diffusion for generative compression","Compression via SDE: no Gaussian noise, just two sampling steps","Rate-variable generative compression with a single denoiser"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that the gradient of the log-probability of a compressed feature is exactly $(D_\\theta(x, q) - x)/q$, an identity proven for Gaussian noise but assumed without proof for the uniform quantization noise that actually corrupts the features; if that equality is wrong, the reverse ODE and stochastic sampler are not following the true score, and the method's theoretical grounding collapses.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion-based compression: one model, any rate, two steps","Reimagining compression as diffusion: smooth rates, minimal steps","Score-estimated reverse diffusion for generative compression","Compression via SDE: no Gaussian noise, just two sampling steps","Rate-variable generative compression with a single denoiser"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000757,"raw_usage":{"total_tokens":3361,"prompt_tokens":941,"completion_tokens":2420,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":2336}},"tokens_in":557,"tokens_out":2420,"duration_ms":19203,"temperature":1.0,"reasoning_tokens":2336,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:43:06.778982+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Estimate $\\nabla \\log p(x; q)$ by Monte Carlo for a fixed latent distribution corrupted by $U(-0.5, 0.5)\\cdot q$ and compare it with the paper's proxy $(D_\\theta(x,q) - x)/q$; a systematic discrepancy would indicate that the Euler update in Eq. (11) is not reversing the forward process, breaking the claimed link between the compression schedule and the score.","supporting_citations":[{"cited_title":"Advances in neural information processing systems 32 (2019) 1, 2, 3, 4, 5","cited_arxiv_id":null,"evidence_quote":"provides denoising score matching, the training objective used to learn the score of corrupted features"},{"cited_title":"Advances in neural information processing systems35, 26565–26577 (2022) 4, 5, 6, 7","cited_arxiv_id":null,"evidence_quote":"supplies the reverse network architecture and the stochastic-sampler design that the paper adopts"},{"cited_title":"In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"supplies the pretrained latent autoencoder whose feature space hosts the compression and reverse processes"}],"review_version":1}