{"id":"f7307370-bb2f-4f76-9070-dccee55a4516","arxiv_id":"2506.23580","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A dataset distillation method that combines K-means image prototypes with LLM-generated text prototypes to synthesize small, high-accuracy training sets.","lead":"The authors augment dataset distillation, which compresses large image collections into tiny training sets, by adding text descriptions generated by an AI vision-language model. The resulting small datasets train image classifiers to higher accuracy than existing distillation methods on several benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The DCS-vs-GGS ablation is confounded by prompt distribution: GGS captions are out-of-distribution for the fine-tuned LDM, so the win attributed to Eq. 4 selection may only reflect distribution match.","rationale":"The paper's central claim is that DCS reaches state-of-the-art accuracy because its frequency-based text-prototype selection (Eq. 4) supplies better semantic conditioning than labels or keyword-only prompts. Table 5 is the main evidence isolating this component. But the LDM is fine-tuned on LLaVA-generated captions, and the authors explicitly state that GGS sentences were not encountered during training. DCS prompts are therefore in-distribution while GGS prompts are out-of-distribution, and diffusion models are known to degrade on out-of-distribution text prompts. The comparison cannot separate 'better prototype selection' from 'same caption distribution as fine-tuning.' The missing control is cheap and decisive: compare DCS against random same-cluster LLaVA captions or in-distribution GGS captions. If those match DCS, the method's novel selection mechanism collapses to 'use any detailed caption,' although the overall text-conditioned pipeline may still be useful; if DCS still wins, the mechanism survives. I therefore keep the reader's CONDITIONAL verdict: the core results are plausible but the load-bearing mechanism claim is not yet isolated. The reader's test-set-tuning and baseline-coverage concerns remain valid secondary issues, but this prompt-distribution confound is the most direct threat to the stated contribution.","tokens_in":14648,"tokens_out":8876,"duration_ms":100742,"concrete_test":"On ImageNette and ImageIDC at IPC=10, 20, and 50, rerun the generation pipeline with three text conditions matched for prompt distribution: (i) the DCS-selected prototype, (ii) a randomly chosen LLaVA caption from the same cluster, and (iii) a GGS-style caption that was included in the LDM fine-tuning set. If (ii) or (iii) matches DCS accuracy within 1%, the frequency-based selection in Eq. 4 is not the operative cause of the reported gains; if DCS still wins, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim is that the frequency-based text prototype selection (Eq. 4) provides better conditioning than alternatives, and Table 5 is the main ablation isolating this component. However, the LDM is fine-tuned on LLaVA-generated image-text pairs, and in §4.5 the authors state that GGS sentences are 'based on FK not encountered during the model’s training.' Thus, DCS prompts are drawn from the same caption distribution used for fine-tuning, while GGS prompts are out-of-distribution at generation time. The observed DCS advantage could be entirely due to distribution match rather than to the word-frequency matching in Eq. 4 capturing cluster-specific visual content. The paper does not include the missing control: a text prototype chosen at random from the same cluster's LLaVA captions, or a GGS-style prompt that is also included in the LDM fine-tuning distribution. Without that control, the specific contribution of the text-prototype selection algorithm over simply using any LLaVA caption is not established. This matters because the SOTA numbers are the headline claim, and the mechanism claimed to produce them is not yet isolated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a dataset distillation framework that augments image prototypes with text prototypes. The authors use LLaVA to generate per-image descriptions, fine-tune a Stable Diffusion (LDM) model on the resulting image-text pairs, cluster the latent features per class with k-means, and select, for each cluster, a 'text prototype' by ranking cluster-specific words by frequency and choosing the sentence with the highest matching score (Eq. 4). The cluster center (image prototype) and the text prototype jointly condition the LDM to synthesize distilled images. Experiments on ImageWoof, ImageNette, ImageIDC, ImageNet-1K, CIFAR-10, and CIFAR-100 report consistent improvements over prior methods across IPC settings and architectures, with the largest gains on ImageNette (e.g., +4.9 points at IPC-20) and on CIFAR-100 (+8.0 points at IPC-10 over RDED).","tokens_in":14914,"tokens_out":14347,"duration_ms":136926,"significance":"The method is a timely extension of generative dataset distillation: it exploits LLM-generated captions for datasets that have no text annotations, and the released code and generated data support reproducibility. If the mechanism claims are confirmed, the frequency-based text prototype selection would be a useful, compute-light component for diffusion-based dataset distillation. The empirical picture is broad (six datasets, three architectures, IPC from 10 to 100), and the pipeline-level gains over strong baselines such as Minimax and D4M are consistent. The main weaknesses are that the component-level ablation (Table 5) does not isolate the proposed selection rule from a prompt-distribution confound, and the parameter analysis in Section 4.6 selects hyperparameters on the ImageIDC benchmark, which is itself reported as a headline result.","major_comments":[{"comment":"The ablation that isolates the proposed text-prototype selection (DCS vs. GGS and L+FK) is confounded by prompt distribution. The LDM is fine-tuned on LLaVA-generated captions (§3.1), and DCS text prototypes are drawn from that same caption set, whereas §4.5 states that GGS sentences are 'based on FK not encountered during the model's training'; L+FK likewise feeds the model keyword bags rather than the sentence distribution used in fine-tuning. The observed DCS advantage could therefore reflect distribution match rather than the word-frequency matching in Eq. 4. The authors should add the missing control of a text prototype selected at random from the same cluster's LLaVA captions (in-distribution but without frequency-based selection), and ideally fine-tune an LDM whose training corpus also includes GGS-style sentences so that the GGS baseline is in-distribution. Without these controls, the specific contribution of the Eq. 4 selection rule over simply using any LLaVA caption is not established.","section":"§4.5, Table 5"},{"comment":"Hyperparameters α (LOF contamination), β (nonrepresentative threshold), and k (top-k words) are analyzed on ImageIDC, and the text indicates that β=0.2 and k=35 are selected from those curves while α is 'adjusted based on the dataset characteristics.' The paper never states which values are actually used in Tables 1–4. Because ImageIDC is itself one of the reported benchmarks, choosing β and k to maximize ImageIDC accuracy and then reporting ImageIDC as a primary result is a form of test-set hyperparameter selection that can inflate the headline numbers. The authors should state the exact β, k, and α settings used in every experiment, and either select them on a held-out split or show that the same settings perform well across all datasets reported.","section":"§4.6, Fig. 4"},{"comment":"The state-of-the-art claim for the low-resolution benchmarks is not fully supported by the comparison set: CIFAR-10 and CIFAR-100 are evaluated only against SRe2L and RDED, omitting standard baselines such as DM, IDC, MTT, and DSA that report on these datasets, and ImageNet-1K (Table 3) omits generative baselines such as GLaD that report the same benchmark. The comparison should be expanded to at least the strongest available baselines for each dataset, or the SOTA claim should be explicitly qualified to the compared methods.","section":"Tables 3–4"}],"minor_comments":[{"comment":"The notation 'Concat(zc_t, τθ(Tc))' is not how Stable Diffusion conditions on text (which is via cross-attention), so Eq. (6) and Algorithm 1 should be corrected or clarified to avoid implying channel-wise concatenation of text embeddings with latents.","section":"§3.4, Eq. (6)"},{"comment":"Equation (3) defines Rw as a set of words, but Eq. (4) uses Rw[w] as an importance weight; the authors should clarify that Rw is a (word, frequency) set and that Rw[w] denotes the cluster-level frequency fc.","section":"§3.3.2, Eq. (3)"},{"comment":"The 'Full' column in Table 1 is not a competing method and should be labeled as the full-dataset reference accuracy, with a note in the caption so the bold/underlined convention is unambiguous.","section":"Table 1"},{"comment":"The claim of being the 'first work that integrates language information into visual dataset distillation' is too broad given that GLaD [4] already conditions generative dataset distillation on language; the authors should qualify the novelty to the use of LLM-generated descriptive text prototypes with frequency-based selection.","section":"Contributions, §1"},{"comment":"The paper does not specify which LLaVA checkpoint (e.g., LLaVA-1.5-7B/13B or LLaVA-NeXT) or inference settings produced the captions, nor the exact contamination values per dataset mentioned in §3.2; these details are needed for reproducibility.","section":"§3.1"},{"comment":"The total computational cost of the pipeline (LLaVA captioning, LDM fine-tuning for 8 epochs, and generation) is not reported or compared with the baselines, which matters because dataset distillation is motivated by efficiency.","section":"§4.2"},{"comment":"Some reported gains are within one standard deviation of the baseline (e.g., ImageNette IPC-10: 64.8±3.6 vs. 60.9±1.7); a sentence discussing run-to-run variability would strengthen the SOTA claims, though the consistency across IPC settings and architectures is reassuring.","section":"Table 2"},{"comment":"Reference [8] has a garbled author list ('Zongxion Geng, Jiahui andg Chen...'), and the paper inconsistently typesets 'D4M/D 4M' and 'SRe2L/SRe 2L'; these should be cleaned up.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The key thing to watch in revision is whether the authors can turn the Table 5 ablation into an actual isolation of their selection mechanism. If a random-caption control performs as well as DCS, the contribution reduces to 'use LLaVA captions for conditioning,' which is weaker than the paper's stated claim. The novelty framing relative to GLaD should be corrected as well. The paper otherwise fits the scope of a vision or machine learning venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this paper reports better numbers across several dataset-distillation benchmarks using LLM captions to condition a fine-tuned diffusion model, and the pipeline is simple and easy to reuse. But the main claim—that the frequency-based text prototype selection in Eq. 4 is what improves generation—is not supported by the ablations as run. The one control that would show that is missing.\n\nWhat is genuinely new: D4M already conditions the LDM on class labels; here the authors add descriptive captions produced by LLaVA, then pick one caption per image cluster as the text prototype. The word-frequency selection scheme (exclude class-wide words, rank top-k, pick best matching sentence) is a reasonable heuristic and appears to give coherent, object-preserving generations in the visualizations. The results on ImageWoof, ImageNette, ImageIDC, and ImageNet-1K are consistent, and the code/data release is a plus.\n\nNow the soft spots.\n\nFirst, the stress-test concern is correct and it matters. In Table 5, DCS beats GGS, but GGS uses GPT-generated sentences that are, as the paper itself says in Sec 4.5, not encountered during the LDM's fine-tuning. The LDM is trained only on LLaVA captions. So the DCS advantage could simply be distribution match, not the selection algorithm. The missing control is a randomly chosen LLaVA caption from the same cluster, or an in-distribution GGS-style prompt. Without that, the contribution of Eq. 4 is unproven.\n\nSecond, Sec 4.6 tunes alpha, beta, and k on ImageIDC, and ImageIDC is one of the primary reported datasets. The chosen values are not clearly stated in the main tables. That is a circularity problem, even if moderate.\n\nThird, the CIFAR comparisons are only against SRe2L and RDED, so the 'state-of-the-art' claim on CIFAR is weaker than the paper implies.\n\nFinally, the 'first to integrate language information' claim is overstated because D4M already uses class-label text; the novelty is in descriptive captions and their selection, not in using text at all.\n\nThese are fixable. The core pipeline is sound and the reported gains are plausible. A clean ablation with the right control plus a transparent hyperparameter report would make this a solid paper. I would send it to peer review, not desk-reject it.\n\nBest.","headline":"Solid DD results with LLM captions, but the key ablation is confounded by prompt distribution and hyperparameters are tuned on a reported test set.","tokens_in":15384,"tokens_out":3219,"would_cite":false,"duration_ms":31445,"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":"Adding LLM-written text prototypes to image prototypes improves dataset distillation enough to set reported accuracy highs on six benchmarks.","keywords":["dataset distillation","text prototypes","image prototypes","vision-language","diffusion models","LLM-generated captions","k-means clustering","synthetic data generation"],"falsifier":"Re-run the pipeline with the text prototype for each cluster replaced by a randomly chosen caption from a different cluster, keeping all image prototypes and diffusion settings fixed; if top-1 accuracy on ImageNette at IPC 50 does not fall materially below the reported 81.2%, then the word-frequency text selection is not what carries the claimed gain. A complementary check is to compute a standard vision-language embedding similarity between each text prototype and its own cluster's images: if prototypes are no closer to their own cluster than to other clusters, the semantic-matching mechanism fails.","tokens_in":1729,"feed_emoji":"🖼️","tokens_out":5605,"duration_ms":104041,"temperature":0.7,"pith_summary":"The paper claims that dataset distillation has been missing semantic information: prior diffusion-based methods compress only what can be seen in image features, so distilled images can lose target objects, warp anatomy, or lock in background co-occurrence biases. The proposed remedy is to give each distilled-image cluster a text prototype extracted from LLM-generated captions, and to condition the diffusion model on both the image prototype and that text prototype. The result, the paper reports, is surrogate datasets that are more logically coherent and that achieve the best top-1 accuracy it compares against on ImageWoof, ImageNette, ImageIDC, ImageNet-1K, CIFAR-10, and CIFAR-100. A sympathetic reader would care because the recipe adds semantic grounding to a purely visual compression task using only open-source components, and it works on datasets that have no pre-existing text annotations.","feed_headline":"LLM text prototypes lift distilled-dataset accuracy on six benchmarks","feed_subtitle":"Pairing word-frequency-chosen captions with k-means image prototypes yields gains of up to 4.9% over prior methods.","key_machinery":"The load-bearing object is a cross-modal prototype pair. Image prototypes are cluster centers produced by k-means on latent features after Local Outlier Factor removal of low-density samples. Text prototypes are built per cluster by tokenizing LLM-generated captions, deleting words that appear in more than a threshold fraction of the class, ranking the surviving words by frequency, and selecting the sentence whose matching score, computed by summing the frequencies of its top-k words, is highest. The image prototype and the text prototype are then fed together into a fine-tuned latent diffusion model, with the text embedding conditioning the denoising step, so that every synthesized image inherits both the cluster's visual center and its semantic description.","core_discovery":"The central discovery is that the missing ingredient in diffusion-based dataset distillation is semantic, not visual. Once each image prototype, obtained by k-means clustering in the latent space of a pretrained autoencoder, is paired with a text prototype selected from LLM-generated descriptions by word-frequency matching, the latent diffusion model produces synthetic images that contain the intended objects, avoid structurally implausible outputs, and train downstream classifiers to higher top-1 accuracy than image-only distillation on every benchmark tested. Concretely, the paper reports gains of 3.9%, 4.9%, and 3.5% over the best prior method on ImageNette and 2.9%, 4.2%, and 2.5% on ImageIDC at IPC 10, 20, and 50, plus an 8.0% gain over RDED on CIFAR-100 at IPC 10. The claim is that this is the first integration of language information into visual dataset distillation for classification, and that it generalizes across low- and high-resolution datasets and across test architectures.","pith_inferences":["The mechanism suggests text prototypes could be edited directly to steer distilled data, for example rewriting captions to remove co-occurrence biases such as fish always appearing with plants; the paper does not explore this editing use.","Word-frequency matching is one proxy for semantic closeness; a direct test would replace Equation 4 with embedding-similarity selection of the caption per cluster, and if that matches or beats the proposed method, the frequency ranking is serving as a proxy rather than the essential ingredient.","The same prototype-plus-caption scheme may extend beyond classification to detection or segmentation distillation by prompting the LLM for layout-aware descriptions, which the paper only lists as future work.","Because the text stage is a single LLM pass, the two-stage pipeline likely keeps caption-generation cost low relative to iterative pixel optimization on very large source sets, making the semantic addition cheap at scale."],"forward_implications":["Datasets without any existing text annotations can now be distilled with semantic guidance, because the LLM supplies the captions in a one-off pass.","The paired text prototype fixes two failure modes of earlier diffusion-based distillation: synthesized images that contain no target object and synthesized images with implausible structures such as extra or missing limbs.","Per-cluster text prototypes split a class into distinct visual contexts, so the distilled dataset preserves within-class diversity rather than collapsing to a single class-level description.","The gains are largest at small images-per-class settings, for example 8.0% on CIFAR-100 at IPC 10, suggesting semantic information is most valuable when the number of synthesized images is tight.","Because the whole pipeline uses only public datasets, an open-source LLM, and a public diffusion model, the reported results establish a reusable baseline for future vision-language dataset distillation."],"supporting_citations":[{"why":"Supplies the minimax diffusion baseline the method must beat and the evaluation code used for fair comparison, and motivates the semantic addition by showing image-only outputs without target objects.","marker":"[9]"},{"why":"Provides the open-source vision-language model that generates the descriptive text for datasets lacking annotations.","marker":"[17]"},{"why":"Provides the latent diffusion model whose text encoder, U-Net, and decoder carry the prototype-conditioned synthesis.","marker":"[28]"},{"why":"Supplies the image-prototype clustering idea and the D^4 baseline that the proposed text prototypes are added to.","marker":"[31]"},{"why":"Supplies Local Outlier Factor, used to remove low-density samples before clustering so that image prototypes are not distorted by outliers.","marker":"[1]"},{"why":"Provides the ImageNette benchmark where the reported gains of 3.9%, 4.9%, and 3.5% are measured.","marker":"[10]"},{"why":"Provides the ImageIDC benchmark and the IDC-1 baseline, and is used to test the method on fine-grained classes with high inter-class similarity.","marker":"[11]"},{"why":"Provides the RDED baseline compared on ImageNet-1K and CIFAR, anchoring the method's large-scale and low-resolution results.","marker":"[33]"}],"fun_headline_variants":["LLM text prototypes boost dataset distillation accuracy","Semantic captions lift distilled-data performance","Vision-language prototypes enhance dataset distillation","Pairing LLM text with image clusters gains up to 8%","Language-aware distillation beats image-only on six sets"],"cache_read_input_tokens":17664,"weakest_assumption_plain":"The method assumes the LLM-generated captions are accurate and consistent enough that the sentence picked by word-frequency matching truly describes what the cluster's images show, but caption quality is never measured directly, only the final downstream classifier accuracy.","fun_headline_variants_meta":{"raw":{"variants":["LLM text prototypes boost dataset distillation accuracy","Semantic captions lift distilled-data performance","Vision-language prototypes enhance dataset distillation","Pairing LLM text with image clusters gains up to 8%","Language-aware distillation beats image-only on six sets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000206,"raw_usage":{"total_tokens":1409,"prompt_tokens":970,"completion_tokens":439,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":368}},"tokens_in":586,"tokens_out":439,"duration_ms":4676,"temperature":1.0,"reasoning_tokens":368,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:36:51.887780+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the pipeline with the text prototype for each cluster replaced by a randomly chosen caption from a different cluster, keeping all image prototypes and diffusion settings fixed; if top-1 accuracy on ImageNette at IPC 50 does not fall materially below the reported 81.2%, then the word-frequency text selection is not what carries the claimed gain. A complementary check is to compute a standard vision-language embedding similarity between each text prototype and its own cluster's images: if prototypes are no closer to their own cluster than to other clusters, the semantic-matching mechanism fails.","supporting_citations":[{"cited_title":"Efficient dataset distillation via minimax diffusion","cited_arxiv_id":null,"evidence_quote":"Supplies the minimax diffusion baseline the method must beat and the evaluation code used for fair comparison, and motivates the semantic addition by showing image-only outputs without target objects."},{"cited_title":"Visual instruction tuning","cited_arxiv_id":null,"evidence_quote":"Provides the open-source vision-language model that generates the descriptive text for datasets lacking annotations."},{"cited_title":"High-resolution image 9 synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Provides the latent diffusion model whose text encoder, U-Net, and decoder carry the prototype-conditioned synthesis."},{"cited_title":"Dˆ 4: Dataset distillation via disentangled diffusion model","cited_arxiv_id":null,"evidence_quote":"Supplies the image-prototype clustering idea and the D^4 baseline that the proposed text prototypes are added to."},{"cited_title":"A review of local outlier factor algorithms for out- lier detection in big data streams","cited_arxiv_id":null,"evidence_quote":"Supplies Local Outlier Factor, used to remove low-density samples before clustering so that image prototypes are not distorted by outliers."},{"cited_title":"A smaller subset of 10 easily classified classes from imagenet, and a little more french, 2020","cited_arxiv_id":null,"evidence_quote":"Provides the ImageNette benchmark where the reported gains of 3.9%, 4.9%, and 3.5% are measured."},{"cited_title":"Dataset condensation via efficient synthetic- data parameterization","cited_arxiv_id":null,"evidence_quote":"Provides the ImageIDC benchmark and the IDC-1 baseline, and is used to test the method on fine-grained classes with high inter-class similarity."},{"cited_title":"On the diversity and realism of distilled dataset: An efficient dataset distilla- tion paradigm","cited_arxiv_id":null,"evidence_quote":"Provides the RDED baseline compared on ImageNet-1K and CIFAR, anchoring the method's large-scale and low-resolution results."}],"review_version":1}