{"id":"64655592-377b-4f45-848b-cbad2b677f68","arxiv_id":"2505.13152","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ResCDC combines a fidelity-optimized decoder with a latent-conditioned residual denoising diffusion model to improve PSNR by up to 2 dB over CDC while retaining similar LPIPS and FID.","lead":"A new hybrid image and video compression scheme pairs a standard decoder with a diffusion model that refines the reconstruction, improving PSNR by up to 2 dB over the CDC baseline while keeping perceptual scores similar. It offers a practical way to boost fidelity in generative compression without sacrificing visual quality.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PSNR gains may be overfit to the evaluation sets; a held-out test is needed before the +2 dB claim can be accepted.","rationale":"The reader's weakest assumption focuses on the residual being approximately Gaussian. While a heavy-tailed or structured residual could challenge the single DDIM sampler, diffusion models are in principle capable of modeling arbitrary residual distributions; the more immediate threat to the paper's central claim is empirical rather than distributional. The reported numbers are point estimates on the same datasets used for hyperparameter selection, and the video section explicitly reports a PSNR drop, which contradicts the abstract's \"similar results\" and indicates that the image PSNR gain is primarily inherited from the L2-trained decoder. A held-out evaluation with multiple seeds would directly establish whether the +2 dB claim generalizes. This does not move the verdict away from CONDITIONAL, but it sharpens the condition: the authors should demonstrate robustness beyond the two benchmark sets they tuned on.","tokens_in":12117,"tokens_out":16762,"duration_ms":174874,"concrete_test":"Evaluate the released ResCDC checkpoints (and the corresponding CDC baselines) on the CLIC validation set or a held-out subset of ImageNet at the same bitrates, using the hyperparameters fixed in the paper. Report mean and standard deviation of PSNR, LPIPS, and FID over at least 3 sampling seeds, and compute BD-Rate for PSNR with a confidence interval. If the BD-Rate improvement over CDC is not significantly positive, or if LPIPS/FID are significantly worse, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim rests entirely on point estimates on Kodak and DIV2K, with hyperparameters (rho, gamma, number of DDIM steps) explicitly selected using those same datasets (Sections 4.3 and 4.4), and no error bars or multiple seeds are reported. Because the decoder alone already accounts for most of the PSNR gain over CDC (Fig. 5), it is possible that the reported +2 dB and \"comparable\" LPIPS/FID are the result of tuning the residual diffusion component to these specific benchmarks. The video experiments (Fig. 7) show the same residual-diffusion recipe decreasing PSNR by >2 dB relative to the base codecs, which is inconsistent with the abstract's \"similar results\" and suggests the image gain is dominated by the L2-trained decoder rather than a generally fidelity-improving diffusion model. If the fixed hyperparameters do not transfer, the central claim would not hold on other content.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes ResCDC, a hybrid learned image/video compression scheme that combines a distortion-optimized decoder with a latent-conditioned residual denoising diffusion model. The decoder produces an initial reconstruction, and the diffusion model, conditioned on the encoder latent, predicts a residual that is added to the decoder output for perceptual refinement. Training uses a weighted combination of bitrate, l2 decoder, l1 diffusion, and LPIPS losses, with a rho parameter trading off distortion and perceptual terms; inference uses DDIM sampling with a tunable starting noise scale gamma. On the Kodak and DIV2K benchmarks, the authors report up to +2 dB PSNR over CDC while maintaining comparable LPIPS and FID scores. The method is also extended to video by training the diffusion model on top of pretrained video codecs, with results on UVG.","tokens_in":12288,"tokens_out":3304,"duration_ms":35936,"significance":"If the reported results hold, ResCDC is a useful contribution to diffusion-based perceptual compression: it demonstrates that a jointly trained decoder plus residual diffusion can recover much of the PSNR lost by generative decoders while retaining perceptual quality. The public implementation, the clear algorithmic description, and the ablation studies on sampling steps and rho are strengths. The central image-compression claim is plausible and supported by the plotted rate-distortion curves, but the generalization of the claim is weakened by hyperparameter selection on the evaluation datasets, the absence of error bars, and an inconsistency between the abstract's video claims and the reported video PSNR results.","major_comments":[{"comment":"The headline +2 dB PSNR improvement and the 'comparable LPIPS/FID' claim rest on point estimates on Kodak and DIV2K, with rho (0.1, 0.5), gamma (0.8), and the number of DDIM steps (100) explicitly selected using those same datasets. No validation split, multiple seeds, or error bars are reported, so the reported gains may reflect tuning to these specific benchmarks rather than a general property of the method. Please provide a held-out evaluation, a sensitivity analysis over the tuned hyperparameters, or at least error bars across retraining runs to support the generalizability of the central claim.","section":"Sections 4.3 and 4.4, Figures 3-6"},{"comment":"The abstract states that the video extension 'achieves similar results' and the title claims 'higher fidelity' compression, but Figure 7 shows PSNR degradation of more than 2 dB relative to the base codecs in several settings, and the conclusion explicitly acknowledges 'scoring lower on distortion metrics such as PSNR.' This contradiction is load-bearing for the paper's generalized claim. Please either report video results that support a fidelity improvement, or revise the abstract and framing to present the video extension as a perceptual-quality enhancement that trades off PSNR.","section":"Section 4.5, Figure 7, and Abstract"},{"comment":"The method assumes the residual r = x - dec(y) is approximately normally distributed, which motivates the DDIM initialization x_T ~ N(0, gamma I). No empirical evidence for this assumption is provided (e.g., histograms or normality statistics of residuals at the tested bitrates), and no sensitivity analysis with respect to gamma is shown for the image model. If the residual model is inaccurate, the reconstruction quality could degrade on other content or bitrates. Please add a distributional analysis or an ablation varying gamma to substantiate this design choice.","section":"Section 3, Algorithm 2"}],"minor_comments":[{"comment":"The phrase 'increasingthedistortion' in the abstract contains a missing-space typo; it should read 'increasing the distortion.'","section":"Abstract"},{"comment":"The footnote refers to 'DVCV-FM'; this appears to be a typo for 'DCVC-FM,' which is the model name used in the text and references.","section":"Section 4.5, footnote 5"},{"comment":"The text states that gamma = 0.8 gives the best results for models finetuned for perceptual quality, but Section 4.5 later reports gamma = 0.1 for the video models with 10 sampling steps; please clarify why the optimal gamma differs and whether this is content- or step-dependent.","section":"Section 3.3"},{"comment":"The video models show large parameter increases over the base codecs (e.g., DVC from 11.8M to 96.1M), but the paper does not discuss the associated memory or decoding-time overhead; a brief note on runtime and complexity would help practitioners assess the trade-off.","section":"Tables 1 and 2"},{"comment":"The qualitative comparison would be easier to interpret if the figure included the LPIPS/PSNR values for each panel and indicated which rho setting corresponds to the 'ours' images numbered (d)-(f), since the caption lists these but the text does not discuss the visual differences.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is clearly written and the method is well motivated, with code released and comparisons to standard baselines. The main concern is that the central claim of generalizable fidelity gains is not yet fully supported: hyperparameters are tuned on the same datasets used for reporting, no error bars are given, and the video experiments contradict the abstract's 'similar results' phrasing. These issues are fixable within the scope of the manuscript, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the headline: this paper does what it says. On Kodak and DIV2K, ResCDC—a CDC-style diffusion codec with an extra L2-trained decoder and a residual diffusion model—gets up to 2 dB higher PSNR than CDC while keeping LPIPS and FID roughly level. That's a useful result for the perceptual compression crowd. The novel bit is training the decoder and the latent-conditioned residual diffusion model jointly, so they influence each other. Ghouse et al. and Hoogeboom et al. both do residual/conditional diffusion for codec enhancement, but neither trains end-to-end with a distortion decoder. The video extension is also new, though more exploratory.\n\nThe paper is clearly written and the ablations (sampling steps, rho) are informative. They also release code, which is good.\n\nNow the soft spots. The headline +2 dB rests on single-run point estimates, with rho, gamma, and sampling steps chosen on the evaluation sets themselves. No error bars, no multiple seeds. That inflates the risk that the number is tuned to Kodak/DIV2K. The stress-test note makes a fair extra point: your own ablation (Fig. 5) shows the decoder alone contributes most of the PSNR gain, so the diffusion component's fidelity benefit is not clearly isolated. And in the video section, the same recipe drops PSNR by over 2 dB relative to the base codecs—the abstract's 'similar results' is misleading. That doesn't kill the paper's image claims, but it does suggest the diffusion model is not a universal fidelity improver. Also, the assumption that residuals are roughly normal (Sec. 3) is plausible but unexamined.\n\nNone of this is fatal for a subfield where this evaluation style is common. The central claim—ResCDC improves PSNR at similar perceptual quality on standard benchmarks—appears to hold for those benchmarks. It deserves a serious referee, and you'd want the authors to add error bars, hold out a test set for hyperparameter selection, and be clearer about video's PSNR trade-off.\n\nFor me, it's a cite-in-passing paper, not a must-read. Take it to a reading group if you want to discuss evaluation practices in generative compression.","headline":"ResCDC gives a believable +2 dB PSNR over CDC on image benchmarks, but the evaluation is too thin to call it definitive—still worth a referee.","tokens_in":12823,"tokens_out":2685,"would_cite":true,"duration_ms":25215,"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":"A hybrid codec that combines a distortion-optimized decoder with a latent-conditioned residual diffusion model raises PSNR by up to 2 dB over diffusion-only compression while keeping perceptual metrics stable.","keywords":["diffusion models","image compression","perceptual quality","residual prediction","latent conditioning","video compression","DDIM sampling","learned codec"],"falsifier":"Measure the empirical distribution of the residual r = x - dec(y) for Kodak and DIV2K at the tested bitrates: if the residuals are strongly non-Gaussian (e.g., heavy-tailed or bimodal) for typical content, the central modeling assumption fails and the reported PSNR/perceptual advantage should shrink or vanish; a systematic comparison of residual histograms against sampling quality would settle it.","tokens_in":11899,"feed_emoji":"🖼️","tokens_out":5823,"duration_ms":48484,"temperature":0.7,"pith_summary":"Generative compression with diffusion models produces images that look good but distort the original. This paper claims that adding a conventional distortion-optimized decoder before the diffusion stage closes much of that fidelity gap. The decoder produces a first reconstruction; the diffusion model, conditioned on the same latent, predicts only the residual between that reconstruction and the original. On Kodak and DIV2K the hybrid scheme gains up to 2 dB PSNR over the diffusion-only baseline while keeping LPIPS and FID roughly unchanged, and the same recipe lifts perceptual quality of learned video codecs.","feed_headline":"Hybrid diffusion codec gains 2 dB PSNR at equal perceptual quality","feed_subtitle":"A decoder plus a residual denoising model closes the fidelity gap of generative codecs, in images and video.","key_machinery":"The load-bearing mechanism is the decomposition of decoding into a deterministic fidelity-optimized reconstruction and a stochastic perceptual refinement. A decoder network, trained jointly with the rest of the model under an $\\ell^2$ loss, produces an initial image from the compressed latent; a DDIM-style denoising network, conditioned on the same encoder latent, predicts the residual r = x - dec(y) rather than the full image. Because the residual is assumed to be roughly Gaussian, only 100 sampling steps (and gamma=0.8 initialization noise) are needed, and both the bitrate loss and a perceptual LPIPS loss balance the two components through parameters $\\lambda$ and rho. In the video extension, the same residual diffusion is trained on top of a frozen pretrained codec, with the previous reconstructed frame as extra conditioning, using v-prediction and only 10 steps.","core_discovery":"The paper's central claim is that a latent-conditioned residual denoising diffusion model, trained jointly with a conventional decoder, can recover most of the distortion performance of pure autoencoder codecs while retaining the perceptual gains of diffusion-based generative compression. The decoder output is refined by adding a predicted residual, so the diffusion model only needs to model the difference between the autoencoder reconstruction and the ground truth, which the authors argue is approximately Gaussian and therefore cheap to sample. Compared with the conditional diffusion codec (CDC) baseline, the resulting scheme, ResCDC, reports up to +2 dB PSNR on Kodak and DIV2K at comparable LPIPS and FID, with BD-rate savings of about 13–17% against the noise-predicting CDC variant; applied to pretrained video codecs such as Scale-Space Flow and DVC, it improves LPIPS substantially at some PSNR cost.","pith_inferences":["If the residual Gaussianity assumption is the real driver, then content with highly structured residuals (e.g., text, sharp edges) should show smaller gains; one could test by measuring residual kurtosis on such images and correlating it with the PSNR gap.","The video results suggest that a large share of perceptual improvement can be obtained by post-processing a frozen codec, which may make the method practical as an enhancement layer rather than a full replacement codec.","The paper leaves open whether the decoder and diffusion could be trained with a single unified noise schedule; a joint optimization with a learned initialization might further reduce steps.","The method's dependence on a fixed gamma=0.8 suggests that content-adaptive initialization noise could squeeze out additional quality."],"forward_implications":["Diffusion-based codecs can close the fidelity gap to autoencoder codecs without sacrificing perceptual metrics, making them viable for bandwidth-sensitive applications that require both low bitrate and faithful detail.","Residual prediction reduces the required number of sampling steps, lowering the decoding cost that has limited diffusion codecs in practice.","Because the recipe works on top of pretrained video codecs, it offers a plug-in perceptual enhancement path for existing learned video compression systems.","At higher bitrates the PSNR advantage grows, so the hybrid design is most attractive for quality-focused streaming scenarios."],"supporting_citations":[{"why":"The conditional diffusion codec that ResCDC extends; provides the baseline architecture, loss, and the main comparison point.","marker":"[49]"},{"why":"Introduces residual prediction in a diffusion model for deblurring, the mechanism ResCDC adapts for compression.","marker":"[44]"},{"why":"Demonstrates residual diffusion for codec augmentation; motivates modeling the residual instead of the full image.","marker":"[13]"},{"why":"Defines the denoising diffusion probabilistic model and the epsilon-prediction training objective used here.","marker":"[17]"},{"why":"Provides the DDIM sampler that makes 100-step (and 10-step video) decoding possible.","marker":"[39]"},{"why":"The MS-Hyperprior autoencoder backbone whose decoder is trained jointly with the diffusion model.","marker":"[30]"},{"why":"HiFiC is the perceptual-optimized GAN baseline compared in the rate-distortion-perception evaluation.","marker":"[28]"},{"why":"The Vimeo90k septuplet dataset used for training the image and video models.","marker":"[45]"}],"fun_headline_variants":["Hybrid diffusion codec: +2 dB PSNR at equal perceptual quality","Residual diffusion refinement boosts fidelity by 2 dB PSNR","Latent-conditioned residual diffusion improves codec fidelity by 2 dB","Video and image codec: diffusion residual adds 2 dB PSNR, same LPIPS","Diffusion codec models the residual, not the noise, to save PSNR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The residual between the decoder reconstruction and the original image is close enough to a Gaussian that a single DDIM sampler with fixed noise level and 100 steps can reconstruct it well across bitrates and content.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid diffusion codec: +2 dB PSNR at equal perceptual quality","Residual diffusion refinement boosts fidelity by 2 dB PSNR","Latent-conditioned residual diffusion improves codec fidelity by 2 dB","Video and image codec: diffusion residual adds 2 dB PSNR, same LPIPS","Diffusion codec models the residual, not the noise, to save PSNR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000735,"raw_usage":{"total_tokens":3271,"prompt_tokens":915,"completion_tokens":2356,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":2251}},"tokens_in":531,"tokens_out":2356,"duration_ms":16831,"temperature":1.0,"reasoning_tokens":2251,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:18:05.617998+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the empirical distribution of the residual r = x - dec(y) for Kodak and DIV2K at the tested bitrates: if the residuals are strongly non-Gaussian (e.g., heavy-tailed or bimodal) for typical content, the central modeling assumption fails and the reported PSNR/perceptual advantage should shrink or vanish; a systematic comparison of residual histograms against sampling quality would settle it.","supporting_citations":[{"cited_title":"Advances in Neural Information Processing Systems36 (2024)","cited_arxiv_id":null,"evidence_quote":"The conditional diffusion codec that ResCDC extends; provides the baseline architecture, loss, and the main comparison point."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces residual prediction in a diffusion model for deblurring, the mechanism ResCDC adapts for compression."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Demonstrates residual diffusion for codec augmentation; motivates modeling the residual instead of the full image."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The MS-Hyperprior autoencoder backbone whose decoder is trained jointly with the diffusion model."}],"review_version":1}