{"id":"6b078da0-8c31-4872-ab49-abca26f76da4","arxiv_id":"2412.08210","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An image set is compressed by training one conditional diffusion model to memorize the set, after which each image is transmitted as only a log2(M)-bit index while the decoder cost is amortized.","lead":"Unicorn replaces per-image bitstreams with a shared neural decoder that reconstructs each image from just a number, its index in a learned set. The decoder itself is the compressed representation, so the savings appear only when many images share one decoder.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Decoding is not shown to be deterministic: Eq. (4) and every rate-distortion comparison assume one index yields one fixed image, but the paper never fixes the initial noise, so the scheme may not be a codec at all.","rationale":"The reader's weakest assumption is exactly the load-bearing point: the method must be a deterministic codec before its bitrate can be compared. I agree that this is the most serious unresolved issue. The concern is concrete and testable, and it is fixable in principle by specifying a fixed noise or deterministic sampler and by demonstrating near-zero within-index variance in the released code. Because the paper otherwise presents a coherent set-wise compression scheme with a plausible two-part description-length accounting, the appropriate verdict remains CONDITIONAL rather than REJECT. I did not elevate capacity scaling or diversity limits to the primary concern, because even if K(q) grows with M, the method could still be a valid codec for similar-image sets; the determinism question is logically prior and currently unverified.","tokens_in":12603,"tokens_out":8897,"duration_ms":101871,"concrete_test":"Run the released implementation (or retrain per the paper) and decode a fixed set of 1000 indexes 10 times with independent random initial latent noise, identical sampler hyperparameters, and different seeds. For each index, compute the pairwise LPIPS among the 10 reconstructions and compare with the average LPIPS between different indexes. If within-index variance is comparable to between-index distance (e.g., within-index LPIPS > 0.02), the decoder is not deterministic and Eq. (4) undercounts the bitrate. Additionally, if the reported RD curves rely on a particular seed or a fixed noise template, that seed/template is part of the coded representation and its entropy must be added to the bitrate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central bitrate formula in Eq. (4), L = M log2 M + K(q), treats the index as the only per-image information. That is valid only if the decoder is a deterministic function of the index. The paper states in Sec. 1 that 'the decoding process is started from any random generated noise' and in Sec. 4.2 that 'the index controls the latent denoising diffusion process,' but it never specifies the initial noise at test time, whether a fixed seed is used, or whether a deterministic sampler such as DDIM is employed. If the receiver draws a different initial noise than the encoder used, the same index will produce a different image; the index-image mapping is then not a bijection, and the transmitted information must include the noise, breaking Eq. (4). The experimental sections report LPIPS/FID/etc. without reporting seed variance, so the reader cannot tell whether each index reliably reconstructs one image or whether the numbers are selected from favorable stochastic samples. This is not a stylistic omission: without deterministic decoding, the proposed scheme is not a compression codec, and all comparisons to ELIC, HiFiC, COIN, etc. are undefined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Unicorn, a paradigm for image compression in which a set of M images is assigned random indices and a single conditional latent diffusion model is trained to map (index, noise) to each image. The rate for the whole set is given in Eq. (4) as M log2 M + K(q), where K(q) is the encoded model size; at decode time the receiver uses the transmitted index together with randomly generated noise. A prototype, Laduree, is evaluated on 4000-image cat and hybrid sets against JPEG/HEVC/VVC, HiFiC/ELIC, and COIN/Combiner, reporting lower bitrates at matched perceptual quality and an increasing compression ratio as M grows. The paper additionally studies index embeddings, conditioning blocks, latent normalization, and weight quantization.","tokens_in":12879,"tokens_out":7059,"duration_ms":78022,"significance":"If the paradigm works as claimed, it offers a genuinely different viewpoint on dataset-specific compression, and the prototype's design explorations (CAG conditioning, GRF embedding, latent standardization) are useful engineering contributions. The paper also makes a concrete, externally testable claim: a single small conditional diffusion model can memorize thousands of index-image pairs and beat general-purpose codecs on perceptual rate-distortion for that fixed set. However, the current manuscript does not establish that decoding is deterministic, and Eq. (4) is not a rate-distortion statement as written. These gaps are central rather than cosmetic, so the paper needs substantive revision before the claims can be accepted.","major_comments":[{"comment":"The decoding procedure is underspecified with respect to initial noise. The text says the decoding process is started from 'any random generated noise' (Sec. 1) and later that 'the index controls the latent denoising diffusion process' (Sec. 4.2), but it never fixes the initial noise, a seed, or a deterministic sampler. If the receiver draws a different initial noise than the encoder used, the same index reconstructs a different image; then the index-image mapping is not the bijection required for Eq. (4), and the rate-distortion comparisons in Secs. 5.2 and 5.3 are not well-defined. Please specify the exact decoding protocol (e.g., fixed initial noise, DDIM with a fixed seed) and report the variance of the reported metrics over decoding seeds for at least one operating point.","section":"§1, §4.2, Eq. (4)"},{"comment":"Eq. (4) counts only the index code length and the model cost; it omits any distortion term. In the LMDL framework of Eq. (1), the description length is meaningful only for a fixed distortion D, but here the model reconstructs each image only approximately and no term accounts for the reconstruction error. Therefore Eq. (4) cannot by itself support the rate-distortion claims in Sec. 5. The authors should either replace the derivation with a rate-distortion formulation that includes the distortion, or explicitly state that Eq. (4) is only the rate component at an empirically measured distortion.","section":"§4.1, Eq. (4)"},{"comment":"The 'unique bitrate superiority' scaling with M is in large part a restatement of Eq. (4). If K(q) is held constant as M grows, the per-image cost is log2 M + K(q)/M, which necessarily decreases with M. The empirical contribution should be framed as demonstrating that a trainable model can maintain a near-constant K(q) while preserving quality as M grows; as written, the paper presents a mathematical consequence of its own rate formula as a discovered property. Please separate the definitional scaling from the empirical comparison against ELIC and HiFiC.","section":"§5.3, Eq. (4)"},{"comment":"The paper does not specify exactly how K(q) is converted to bits-per-pixel in the RD curves and file-size plots. It must state whether the VAE decoder and any model-compression side information are included, how the quantized weights are counted (number of parameters times W bits, or an entropy-coded length), and whether the same accounting is applied to the INR baselines. Without this, the reported bpp values in Figures 4-6 are not auditable and the comparison with COIN and Combiner may not be fair.","section":"§5.1, §5.2, §5.3"},{"comment":"The manuscript repeatedly refers to appendices that are not present in the submitted text: Appendix 1.3 for the quantitative comparison between Unicorn and ELIC, Appendix 2.2 for the proof that the conditional reverse process reduces to an unconditional one, and Appendices 3.1, 4.1, and 4.2 for model configurations and extra results. The proof in Appendix 2.2 is load-bearing for the 'efficiency through similarity' claim, and Appendix 1.3 is load-bearing for the claim that Unicorn maintains superiority at 'normal magnitude' image counts. Please include the appendices or remove the references.","section":"§3, §4.2, §5.1, §5.2"}],"minor_comments":[{"comment":"There are typos in the abstract: 'One Nnumber Reconstruction' and 'Quantitive'; the main text also has 'rate-distrotion' in Sec. 4.2.","section":"Abstract and title"},{"comment":"The model naming convention 'Data-XHX W X' (e.g., CAT-1500H120 W 14) is introduced without a definition; please define the notation clearly before first use.","section":"§5.1"},{"comment":"The y-axis label 'Compression ratio' is not defined; specify the reference (e.g., ratio of uncompressed size to compressed size, or ratio relative to a baseline) so the reader can interpret the plot.","section":"Figure 6"},{"comment":"The sentence 'the bitrate consumption in compressing images only comes from the encoded weights of the latent diffusion model' appears to ignore the M log2 M index cost in Eq. (4); please reconcile the wording with the formal rate expression.","section":"§4.2"},{"comment":"The notation Pq(I|Y) is not defined precisely; in particular, it is unclear whether I denotes a random image, a latent feature Z, or the reconstructed image, and how the approximate nature of the reconstruction is represented in the probability model.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The core idea is close to memorization with random labels, and the paper should engage more directly with that interpretation; the codec is explicitly dataset-specific, which is a legitimate but unusual setting that needs careful framing. My main concern is the determinism issue: if the authors can fix a deterministic decoding protocol and report seed variance, the central claim becomes testable. I do not think rejection is warranted because the issue is fixable within the manuscript's scope, but the missing appendices and the rate-accounting ambiguity must also be resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is worth reading, but the central claim is currently under-specified. The idea of compressing a whole image set by memorizing index-image pairs in a conditional diffusion model is genuinely new, and the prototype (GRF embedding, CAG gating) is concrete and well engineered. The RD curves on CAT-4000 showing 20%+ savings over ELIC at matched LPIPS are real evidence that something works. The scaling behavior, bitrate falling as M grows, is also real but is just the amortization of the decoder cost; calling it a 'unique bitrate superiority' is overstatement.\n\nThe load-bearing soft spot is determinism. The paper says decoding starts from 'any random generated noise' but never specifies the initial noise or sampler. Every RD comparison assumes one index yields one fixed image. If the decoder is stochastic, then the scheme is not a codec and Eq. (4), L = M log2 M + K(q), is wrong because the transmitted information must include the noise. This is not a minor omission; it is the linchpin. The authors can probably fix it by fixing a seed or using DDIM, but as written the main claim is not established.\n\nSecond, Eq. (4) ignores distortion. It treats the model as a perfect bijection, but reconstruction is approximate. The MDL logic needs a distortion term or a proper rate-distortion analysis. Without it, the formula overstates the compression gain.\n\nThird, the baseline comparisons are not fully fair: IIC methods are evaluated at a much worse perceptual quality (LPIPS 0.35 vs 0.10 for others), which makes the claimed superiority over COIN/Combiner look artificial.\n\nWho should read this: anyone working on neural compression, especially set-wise or model-sharing schemes. It deserves a serious referee, because the idea is novel and could open a useful line. But the revision must pin down deterministic decoding and redo the RD accounting. If those are fixed, this becomes a solid contribution.\n\nMy recommendation: send it to peer review, but with the expectation of heavy revision. The core idea is more interesting than the current writeup.","headline":"Clever set-wise compression idea, but the main claim rests on an unstated deterministic decoding assumption.","tokens_in":13412,"tokens_out":2876,"would_cite":false,"duration_ms":33499,"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 shared diffusion decoder can reconstruct each image in a set from just an index number, making per-image transmission cost fall as the set grows.","keywords":["image compression","latent diffusion models","unified decoder","index-image bijection","perceptual quality","rate-distortion","neural compression","one-number reconstruction"],"falsifier":"Run the trained decoder twice with the same index and two different random initial noises and compare the reconstructions; if they diverge noticeably, the index alone is not a reproducible representation. A complementary check is to compute the actual storage cost for one image and for 4000 images, and test whether the total follows $M\\log_2 M + K(q)$ as claimed.","tokens_in":12402,"feed_emoji":"🖼️","tokens_out":11891,"duration_ms":106204,"temperature":0.7,"pith_summary":"The paper claims that an entire set of images can be compressed by training one shared decoder to memorize a random one-to-one assignment between images and index numbers. After training, each image is transmitted as only its index, and the decoder produces a visually pleasing picture from that index plus random noise; the only remaining cost is the quantized decoder, shared by all images. The central quantitative claim is that the description length of $M$ images is $M\\log_2 M + K(q)$, where $K(q)$ is the size of the compressed decoder, so the per-image online cost falls to $\\log_2 M$ bits once the decoder is installed. The paper's prototype, Laduree, is a latent diffusion transformer, and experiments on 4000-image sets report better perceptual quality at comparable bitrates than the per-image codecs ELIC and HiFiC, with savings that grow as more images are compressed together.","feed_headline":"One index per image: shared decoder cuts bitrate as set grows","feed_subtitle":"The more images stored together, the lower the per-image cost, saving up to 22% over leading codecs.","key_machinery":"The load-bearing object is the index-image pair with a shared conditional diffusion decoder: a bijection $\\{(I_i, Y_i)\\}_{i=1}^{M}$ between images and uniformly sampled indexes in $\\{1,\\dots,M\\}$, learned by a latent diffusion model that denoises random noise conditioned on $Y_i$. The identity that carries the argument is the description-length formula $L^{\\mathrm{Unicorn}}_{DL}(\\tilde S) = M\\log_2 M + K(q)$, which turns compression into memorization of random labels: the first term is the cost of transmitting $M$ indexes and the second is the quantized size of the unified decoder. Within the prototype, the cross-attention-with-gating block and the Gaussian random frequency embedding are the tailored parts chosen to introduce few extra parameters, and weight quantization supplies the model cost term $K(q)$.","core_discovery":"On its own terms, the discovery is that a single conditional diffusion decoder can learn a bijection between $M$ arbitrary indexes and $M$ images well enough that the index alone, together with a random denoising seed, determines the reconstructed image. The paper derives the description length of the whole set as $-\\log_2 P_q(I|Y) + K(q) = M\\log_2 M + K(q)$, treating the decoder as a shared codebook and the index as the codeword. The prototype trains this decoder on index-conditioned latent diffusion, using a Gaussian random frequency embedding for the index and a gated cross-attention block for conditioning, then quantizes the weights to keep $K(q)$ small. In experiments at high perceptual quality (LPIPS around 0.10), Laduree reports 21.73% bitrate savings over ELIC and 17.87% over HiFiC on a 4000-cat-image set, and similar trends on a five-category hybrid set, with the advantage increasing as the number of images grows.","pith_inferences":["An implication the authors leave implicit is that the scheme is essentially a learned lookup table, so the real scaling question is how decoder cost $K(q)$ and reconstruction quality behave as $M$ and image diversity grow, not whether the bijection can be memorized.","The paper never specifies a fixed initial noise per index or demonstrates same-index, same-seed determinism; a testable extension would fix a seed for each index and report reconstruction variance across restarts, since all bitrate comparisons assume the decoder is a deterministic function.","If the approach scales, it suggests a general 'compression by memorization' strategy in which the decoder is a shared codebook and the index is the codeword, naturally combinable with stronger model-compression and pre-trained priors.","At small $M$, the decoder cost $K(q)$ is not amortized and likely dominates, so the claimed advantage over per-image codecs could reverse for very small collections; the paper's comparisons start at 1000 images."],"forward_implications":["If the claim holds, online transmission of an image becomes a $\\log_2 M$-bit index lookup once the shared decoder is installed, so the per-image bitrate shrinks as the collection grows.","Total compressed size grows sublinearly in the number of images, so the compression ratio improves with $M$ instead of staying flat as it does for per-image codecs.","The scheme is strongest for semantically similar image sets, because the diffusion model can reuse shared structure across the memorized images; the cat set shows larger savings than the five-category hybrid set.","Quantizing the decoder weights to roughly 14 bits keeps quality losses acceptable, giving a concrete handle on trading decoder cost against distortion.","At the same time, the paper only matches the GAN-based codec HiFiC in PSNR, consistent with a rate-distortion-perception trade-off, so the gains are perceptual rather than pixel-fidelity gains."],"supporting_citations":[{"why":"Supplies the lossy minimum-description-length principle used to set up the bitrate formula.","marker":"Madiman, Harrison, and Kontoyiannis 2004"},{"why":"Provides the uniform conditional distribution $P(I|Y)=1/M$ and the model description length that appear in Eq. (4).","marker":"Blier and Ollivier 2018"},{"why":"Supports the premise that neural networks can overfit random labels, which justifies learning the index-image bijection.","marker":"Zhang et al. 2021"},{"why":"Gives the latent diffusion model and pre-trained VAE used to generate images from noise conditioned on an index.","marker":"Rombach et al. 2022"},{"why":"Supplies the diffusion transformer backbone and the conditioning and embedding baselines that the prototype adapts.","marker":"Peebles and Xie 2023"},{"why":"ELIC is the neural end-to-end baseline whose bitrate the prototype reports undercutting by 21.73%.","marker":"He et al. 2022"},{"why":"HiFiC is the GAN-based perceptual baseline compared in the rate-distortion curves.","marker":"Mentzer et al. 2020"},{"why":"COIN defines the implicit-neural-representation compression baseline that motivates the IIC comparison.","marker":"Dupont et al. 2021"},{"why":"Combiner is the Bayesian INR baseline used to show that a unified decoder needs fewer bits than per-image models.","marker":"Guo et al. 2023"}],"fun_headline_variants":["One number per image: shared diffusion decoder shrinks bitrate","Index-only image compression: bitrate drops as dataset grows","Shared decoder codebook: one number reconstructs any image","Bitrate savings grow with image count via one-number code"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The decoding procedure assumes that the same index together with the same initial random noise always yields the same image; the paper does not specify the initial noise or verify this determinism, yet every bitrate comparison depends on it.","fun_headline_variants_meta":{"raw":{"variants":["One number per image: shared diffusion decoder shrinks bitrate","Index-only image compression: bitrate drops as dataset grows","Shared decoder codebook: one number reconstructs any image","Bitrate savings grow with image count via one-number code"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000723,"raw_usage":{"total_tokens":3299,"prompt_tokens":1056,"completion_tokens":2243,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":672,"completion_tokens_details":{"reasoning_tokens":2175}},"tokens_in":672,"tokens_out":2243,"duration_ms":17221,"temperature":1.0,"reasoning_tokens":2175,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:05:55.463305+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained decoder twice with the same index and two different random initial noises and compare the reconstructions; if they diverge noticeably, the index alone is not a reproducible representation. A complementary check is to compute the actual storage cost for one image and for 4000 images, and test whether the total follows $M\\log_2 M + K(q)$ as claimed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the lossy minimum-description-length principle used to set up the bitrate formula."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the uniform conditional distribution $P(I|Y)=1/M$ and the model description length that appear in Eq. (4)."},{"cited_title":"D.; Tschannen, M.; and Agustsson, E","cited_arxiv_id":null,"evidence_quote":"HiFiC is the GAN-based perceptual baseline compared in the rate-distortion curves."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Combiner is the Bayesian INR baseline used to show that a unified decoder needs fewer bits than per-image models."}],"review_version":1}