{"id":"085af7b0-bb36-41aa-bf24-12430fa7ed2b","arxiv_id":"2411.11329","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"AutoPalette reduces the color palette of distilled images with a learned palette network and a color-guided initialization, claiming better accuracy under the same images-per-class budget.","lead":"This paper adds a color-reduction step to dataset distillation, using a learned palette network to compress each synthetic image to fewer colors and a graph-cut selection to initialize distilled data. The authors report accuracy gains on CIFAR10/100 and ImageNet subsets, but the claimed storage savings are not rigorously accounted for.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Storage accounting in §3.6 is inconsistent with the per-image palettes computed in Eqs. (4)–(5); with K=64 and IPC=10/50 the class-level 256-color budget cannot hold, and 64 colors is 6-bit, not 4-bit, so the same-storage claim is unestablished.","rationale":"The reader's weakest assumption identifies precisely the decisive flaw: §3.6 assumes a class-level color budget that would only be valid under a shared per-class palette, whereas the method in §3.3 constructs an independent palette for each image. This is not a stylistic or presentational issue; it directly undermines the quantitative storage-savings claim and the equal-budget comparisons that form the paper's central contribution. The bit-depth mislabeling (64 colors described as 4-bit instead of 6-bit) is a separate but reinforcing error. I considered whether the empirical accuracy gains could rescue the paper as an unconditional contribution independent of storage accounting, but the framing throughout is explicitly storage-efficiency-oriented, and without a valid storage model the reported IPC-based comparisons do not support the 'same storage budget' language. Since the reader already recommended REJECT and my analysis confirms that recommendation without moving it, the appropriate status is unchanged.","tokens_in":14091,"tokens_out":6810,"duration_ms":68186,"concrete_test":"Recompute the storage per class for AutoPalette using the actual representation implied by Eqs. (4)–(5): for each synthetic image, store a per-channel palette of K entries at 8 bits each plus an H×W×C index map of ⌈log2 K⌉ bits per entry, then compare with the 8-bit baseline at the same IPC for every row of Table 1. If the total is not less than 8×IPC×CHW, or if the §3.6 constraint sum_i N_i ≤ 256 fails for IPC=10 or 50 with K=64, then the same-storage comparison and the headline storage-savings claim are invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that AutoPalette achieves superior dataset distillation under the same storage budget by storing color-reduced images. That claim depends on the storage analysis in §3.6, which is inconsistent with the method as defined. Eq. (4) computes a palette per image: m̃ ∈ R^{C×K} is formed by averaging the pixel values of that image assigned to each color bucket, and Eq. (5) reconstructs each image from its own palette and its own argmax index map. Thus, storing the output B requires storing, for every image, a K-entry per-channel palette plus an H×W×C index map. Section 3.6 instead imposes a class-level packing condition sum_i N_i ≤ 256, which is only meaningful if all images in a class share a single color table. Under that condition, with n=6 bits (64 colors), at most 2^{8−6}=4 images per class can be represented, but the experiments use IPC=10 and IPC=50 with K=64. The condition is violated by the reported experimental configuration, or the per-image palettes are omitted from the accounting. Including per-image palettes for CIFAR (32×32×3, K=64) gives roughly 10×(32·32·3·6 + 64·3·8) ≈ 199,680 bits per class, compared with 245,760 bits for ordinary 8-bit images—a saving of only about 19%, not the factor implied by calling the images '4-bit'. Moreover, a 64-color representation requires 6 bits per pixel index, so the repeated '4-bit' characterization in the abstract and introduction is incorrect. Because Tables 1 and 2 compare accuracy at fixed IPC rather than at equal true storage cost, the claimed 1.7% and 4.2% same-budget gains are not established by the reported experiments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AutoPalette, a dataset-distillation framework that reduces color redundancy at two levels: a palette network maps each pixel of a synthetic image to one of K colors per channel (Eqs. 3–5), and a color-guided initialization selects diverse real images using conditional gain of a graph-cut function on median-cut quantized images. The synthetic images are stored in a reduced color space, and the paper claims in the abstract and introduction that models trained on '4-bit' images achieve competitive accuracy and that under the same storage budget AutoPalette outperforms prior methods by 1.7% and 4.2% on CIFAR-10 and CIFAR-100. Experiments are reported on CIFAR-10/100, ImageNet subsets, cross-architecture settings, and ablations.","tokens_in":14497,"tokens_out":11251,"duration_ms":99918,"significance":"The motivation—removing color redundancy in distilled images—is reasonable, and the paper contains a broad set of experiments, including compatibility with DM and TM objectives and cross-architecture evaluation, which are useful. If the storage-efficiency claim were accurately quantified, this could be a meaningful contribution to parameterization-based dataset distillation. However, the central claim is not currently supported: the storage analysis in §3.6 is inconsistent with the per-image palettes defined in §3.3, the '4-bit' characterization is factually incorrect for K=64 colors per channel, and the experimental comparisons do not establish equal storage budgets. These issues concern the headline contribution, so the paper's main significance is not established.","major_comments":[{"comment":"Equations (4)–(5) define a per-image color palette: for each image, m̃ is the average of that image's pixels assigned to each color bucket, and the reconstructed image b is generated from this image-specific palette. Section 3.6, however, performs the storage analysis with a class-level color budget (Σ_i N_i ≤ 256 and at most 2^{8−n} images per class for n-bit color). The experiments use K=64 colors per channel and IPC=10 or 50; with n=6 bits, the class-level condition permits only 4 images, so the reported configuration violates the analysis, and if per-image palettes are stored, their cost is omitted from the claimed savings. Concretely, for CIFAR (32×32×3, K=64), an index map costs 32·32·3·6 = 18,432 bits and the per-image palette costs 64·3·8 = 1,536 bits, for 19,968 bits per image versus 24,576 bits for 8-bit storage, a saving of roughly 19% rather than the factor implied by the paper. This accounting directly underlies the abstract's 'same storage budget' claim.","section":"§3.3, Eqs. (4)–(5), and §3.6"},{"comment":"The paper states in the abstract and introduction that the framework produces '4-bit images' and that this is the source of storage savings. With K=64 colors per channel, each pixel index requires log2(64)=6 bits per channel, so the images are 6-bit per channel, not 4-bit. Table 1 itself states that the synthetic images are reduced from 256 to 64 colors. The repeated '4-bit' label is incorrect and cannot support the claimed storage comparison.","section":"Abstract, §1, and Table 1"},{"comment":"The color transformation defined by Eqs. (4)–(5) is non-differentiable because it uses the Kronecker delta of the arg max and the arg-max selection h. The overall objective in Eq. (14) differentiates L_task with respect to S and θ_c through B=ϕ_palette(S;θ_c), and Eq. (13) updates the palette network with L_palette, which includes L_task and L_a (Eq. (12), also based on arg max). The manuscript does not describe any straight-through estimator, Gumbel-Softmax, or other gradient approximation for these discrete operations. The training procedure is therefore underspecified at a load-bearing point.","section":"§3.3 and §3.5"},{"comment":"The text in §4.3 says the ablation fixes the number of synthetic images to 10 per class 'rather than fully utilizing the available storage capacity,' which strongly suggests that the main experiments in Table 1 store more than IPC images per class within the same budget. If that is the case, Tables 1–2 are not comparisons at fixed IPC or fixed storage budget, and the headline improvements are not attributable to color reduction alone. This reading is supported by internal inconsistencies: the full AutoPalette model on CIFAR-10 IPC=10 is 74.3% in Table 1 but 66.20% in Table 3, and Table 7 reports 66.8% for 256 colors with IPC=10. The paper must state exactly how many images and how many bits per class are used in each reported result.","section":"§4.3 and Tables 1, 3, 7"},{"comment":"Equation (8) defines conditional gain as G(A|C)=G(T^Q)−G(C), which does not depend on the candidate set A and is not the conditional gain used in Appendix A.1 (where f(A|B)=f(A∪B)−f(B)). Equation (11) writes the selection objective as arg max_c G*(A) − 2Σ_{i∈A}Σ_{c∈C} Sim(i,c), but the summation and the arg max use the same symbol c, and the expression does not isolate the candidate being scored, so the selection rule is ambiguous as written. This affects the color-guided initialization, one of the two claimed contributions.","section":"Eq. (8) and Eq. (11)"}],"minor_comments":[{"comment":"The caption of Table 2 says 'All experiments are conducted on CIFAR10 with IPC=10 storage budget for parameterization methods,' but the tabulated datasets are ImageNet subsets (ImageNette, ImageWoof, etc.); the caption should be corrected.","section":"Table 2 caption and §4.2"},{"comment":"References [4] and [5] are the same paper (Cazenavette et al., 'Generalizing dataset distillation via deep generative prior'), and references [28] and [30] also appear to describe the same paper; duplicated references should be consolidated.","section":"References"},{"comment":"In Appendix A.1, the notation f(A|B) is described as the gain of adding set B to set A, but Eq. (16) is written as f(A∪B)−f(B); the notation should be aligned with the main text and with the standard definition of conditional gain.","section":"Appendix A.1"},{"comment":"The symbol C is used both for the unselected set in Eq. (11) and for the number of image channels in §3.3; this overloaded notation makes the initialization equations harder to follow and should be changed.","section":"§3.4"}],"recommendation":"reject","confidential_remarks":"The headline claim of 4-bit storage and same-budget superiority is contradicted by the method's own definitions in Eqs. (4)–(5) and by the storage analysis in §3.6. The authors should be asked to provide a corrected storage model and to clarify the actual number of images and bits per class used in each experiment. If the code is available, a quick inspection of the saved representation would resolve the ambiguity, but as written the central contribution is not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Color-space parameterization is a genuinely new direction in dataset distillation, and AutoPalette is not just a gimmick: the palette network with the two task-aware regularizers plus the graph-cut initialization produces real accuracy gains, especially on CIFAR100 (52.6% at IPC10, 53.3% at IPC50, several points above prior PDD methods). The compatibility experiments with DM and TM are also convincing evidence that the color-reduction idea helps across objectives.\n\nBut the headline claim—same storage budget, better accuracy—does not survive contact with the paper's own storage section. The method computes a separate K-color palette per image (Eqs. 4-5), so the stored representation is the index map plus per-image palette. Section 3.6 instead assumes a class-level shared color budget with sum N_i ≤ 256, which would only make sense if all images in a class share one palette. With K=64 and IPC=10 or 50, that packing condition is violated. And 64 colors per channel is 6 bits per pixel index, not 4 bits as stated in the abstract and intro. The comparison tables are at fixed IPC, not at equal true storage cost, so the promised 1.7% and 4.2% same-budget gains are not established.\n\nThere is also a technical gap: the argmax in Eqs. (4)-(5) is non-differentiable, and the paper never describes a gradient estimator. Maybe they use a straight-through trick in the code, but the text is silent. That matters because L_task is backpropagated through the palette network.\n\nNone of this kills the core idea. The accuracy numbers are real and the color-reduction approach is worth pursuing. But the paper needs a corrected storage analysis and a clearer bit-depth story before the central claim can be taken at face value.\n\nWho should read it: anyone working on parameterization-based dataset distillation. It deserves a proper peer review, but the referee should be asked to check the storage math and the differentiability carefully.","headline":"The color-parameterization idea and empirical gains are real, but the storage-efficiency claim is undercut by an inconsistent accounting and a bit-depth error.","tokens_in":15019,"tokens_out":3566,"would_cite":false,"duration_ms":32365,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Distilled images keep their accuracy when each channel is cut from 256 colors to 64.","keywords":["dataset distillation","color quantization","palette network","storage efficiency","synthetic data","trajectory matching","submodular selection","CIFAR"],"falsifier":"Compute the exact on-disk size of an AutoPalette synthetic dataset, including each image's own 64-color palette and the bit-packed pixel indices, and compare it with the 8-bit storage of the same synthetic images; if the total is close to or above 8-bit storage, the claimed 'same storage budget' comparisons are not what they appear.","tokens_in":13917,"feed_emoji":"🎨","tokens_out":7697,"duration_ms":71243,"temperature":0.7,"pith_summary":"Dataset distillation compresses a large training set into a small set of synthetic images, but those images are stored in full 8-bit color even though much of that color information is redundant. This paper argues that a distilled image can be represented with far fewer colors—64 per channel instead of 256—without losing the low-level features, such as edges and shapes, that matter for training, and that the freed storage can be spent on more diverse samples. To make that reduction work, it introduces a palette network that learns which pixels deserve which of the limited colors, plus a color-guided initialization that picks starting images with the least overlapping color patterns. If the claim holds, dataset distillation can either use much less memory for the same accuracy or use the saved budget to store more images per class.","feed_headline":"64 colors per channel suffice for distilled-image accuracy","feed_subtitle":"Storing synthetic images in 6-bit color instead of 8-bit frees budget for more images per class.","key_machinery":"The load-bearing object is the palette network, a small convolutional network that takes a synthetic image and outputs a per-pixel probability distribution over K color buckets for each channel. From these assignments the method builds a per-image palette by averaging the pixel values that fall in each bucket, then reconstructs a color-condensed image by mapping every pixel to its bucket's average color. Three auxiliary losses carry the argument: the maximum color loss pushes every bucket to be used by at least one pixel, the palette balance loss spreads pixel assignments evenly across buckets, and an alignment regularizer keeps the learned clustering close to Median Cut so that the initialization's assumptions hold. On top of that, color-guided initialization uses a generalized graph cut over gradient similarities of quantized images to pick starting images with minimally overlapping color patterns.","core_discovery":"The central claim is that the color space of distilled images is over-provisioned: most of the 256 colors per channel in a synthetic image are not needed for a downstream network to learn the discriminative structure, so the synthetic dataset can be parameterized with as few as 64 colors per channel (a 6-bit palette) and still train models that match or beat models trained on 8-bit synthetic images from other distillation methods. AutoPalette realizes this with a palette network that maps each pixel to one of K color buckets per channel, a per-image palette computed by averaging the pixels assigned to each bucket, and auxiliary losses that force every bucket to be used and used in balance. It further reduces redundancy between images by initializing the synthetic set with real images selected for low color-pattern replication, measured by submodular information gain on Median-Cut-quantized copies. With the same storage budget, the paper reports gains of 1.7% on CIFAR-10 and 4.2% on CIFAR-100 over existing parameterization-based distillation methods.","pith_inferences":["The paper's abstract says the outputs are '4-bit,' but the experiments use 64 colors per channel, which is 6-bit; the pixel-storage saving is about 25% relative to 8-bit, not the 50% a 4-bit format would give. (Inference from the experimental setting.)","The storage analysis in Section 3.6 budgets one palette per class, while Eq. (4) builds a different palette per image; if those per-image palettes are stored, the reported memory savings shrink or disappear. (The paper does not address this accounting.)","The same palette-network idea could be combined with spatial downsampling or frequency-domain parameterization to compress distilled datasets along multiple axes at once; the paper only varies color depth.","A natural follow-up is a shared class-level palette that makes the storage budget and the initialization approximation match the training-time color mapping; the paper does not evaluate that variant."],"forward_implications":["Pixels in distilled images can be stored in 6 bits instead of 8 (64 colors per channel instead of 256), a 25% cut in pixel storage, with accuracy at least matching 8-bit synthetic images.","At a fixed storage budget, AutoPalette outperforms other parameterization methods by 1.7% on CIFAR-10 and 4.2% on CIFAR-100, meaning the freed budget can be spent on more images per class.","The palette network is framework-agnostic: combining it with trajectory matching or distribution matching improves accuracy in every tested IPC setting, with gains up to 15.7 points on CIFAR-100.","Distilled images remain informative at very low color depth: dropping from 256 to 8 colors per channel costs only about 2 points on CIFAR-10 at IPC=10."],"supporting_citations":[{"why":"Supplies the trajectory-matching distillation objective that AutoPalette uses as its task loss and as the primary comparison baseline.","marker":"[3]"},{"why":"Provides the difficulty-aligned trajectory matching implementation that AutoPalette follows for training and evaluation.","marker":"[12]"},{"why":"Median Cut color quantization is used to approximate the palette network during initialization and serves as a comparison baseline for color reduction.","marker":"[15]"},{"why":"Generalized graph cut defines the submodular conditional gain used by the color-guided initialization to select diverse images.","marker":"[2]"},{"why":"Introduces dataset distillation and the bi-level optimization formulation that AutoPalette extends with color parameterization.","marker":"[40]"},{"why":"Distribution matching serves as the second distillation framework in the compatibility experiments, showing the palette network is plug-and-play.","marker":"[46]"}],"fun_headline_variants":["64 colors per channel beat 8-bit for distilled data","AutoPalette: less color, more distilled accuracy","Distilled images thrive with 6-bit color palettes","Color-aware initialization and palettes boost distillation","Shrink synthetic color space, expand model performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed storage savings depend on the assumption that a single color palette can be shared by all images in a class, but the method as described computes a separate palette per image, so palette storage is not charged to the budget.","fun_headline_variants_meta":{"raw":{"variants":["64 colors per channel beat 8-bit for distilled data","AutoPalette: less color, more distilled accuracy","Distilled images thrive with 6-bit color palettes","Color-aware initialization and palettes boost distillation","Shrink synthetic color space, expand model performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000241,"raw_usage":{"total_tokens":1517,"prompt_tokens":939,"completion_tokens":578,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":503}},"tokens_in":555,"tokens_out":578,"duration_ms":5701,"temperature":1.0,"reasoning_tokens":503,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:39:33.554250+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the exact on-disk size of an AutoPalette synthetic dataset, including each image's own 64-color palette and the bit-packed pixel indices, and compare it with the 8-bit storage of the same synthetic images; if the total is close to or above 8-bit storage, the claimed 'same storage budget' comparisons are not what they appear.","supporting_citations":[{"cited_title":"Color image quantization for frame buffer display","cited_arxiv_id":null,"evidence_quote":"Median Cut color quantization is used to approximate the palette network during initialization and serves as a comparison baseline for color reduction."},{"cited_title":"Graph cuts in vision and graphics: Theories and applications","cited_arxiv_id":null,"evidence_quote":"Generalized graph cut defines the submodular conditional gain used by the color-guided initialization to select diverse images."},{"cited_title":"Dataset condensation with distribution matching","cited_arxiv_id":null,"evidence_quote":"Distribution matching serves as the second distillation framework in the compatibility experiments, showing the palette network is plug-and-play."}],"review_version":1}