{"id":"29d72c35-6b84-4f87-98c2-1c5c33071232","arxiv_id":"2608.10706","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MMArt offers 74,234 paintings with four complementary perspectives and a unified caption, and shows the perspectives are task-asymmetric.","lead":"MMArt is a new dataset of 74,234 paintings, each paired with four separate written perspectives: narrative, formal, emotional, and historical, plus one combined caption. The paper argues no single perspective is enough by showing they help different tasks, such as retrieval versus reconstruction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The §4 complementarity results may reflect T2I/retrieval priors from metadata keywords inside perspectives (artist/style/date from ArtRAG; formal vocabulary from GalleryGPT), not the visual information each perspective encodes; no metadata-stripped or null control is reported.","rationale":"I read the paper's strongest claim as being about complementarity: that the two analyses in §4 establish that no single perspective suffices. That claim is exactly what must hold for the multi-perspective design to be justified. The reader identified the weakest link as the assumption that reconstruction fidelity measures how much visual information a perspective encodes, rather than the text-to-image model's prior and prompt-following behavior. My concern is closely aligned but more specific: the perspective texts themselves embed annotation-source metadata — artist, title, date, style terms — that can drive both the generative and the discriminative results independently of any genuine visual-content encoding. This is particularly acute for the historical perspective, which is generated by retrieval augmentation and therefore contains exactly the kind of proper nouns and movement labels on which T2I models have strong priors. The surprising inversion between historical reconstruction importance and historical retrieval failure is precisely the signature one would expect from such metadata priors, so the inversion does not by itself resolve the confound. Consequently, the complementarity conclusion is plausible but not yet cleanly established. I do not think this invalidates the dataset; conditional acceptance remains appropriate, with the added requirement that the authors report proper controls. The concrete test I propose — stripping metadata and adding a metadata-only baseline — would settle whether the task asymmetry is a property of the perspectives or of their generation provenance. This is why I keep the verdict at CONDITIONAL-equivalent (UNCHANGED relative to the reader) rather than moving to reject.","tokens_in":13319,"tokens_out":5352,"duration_ms":56878,"concrete_test":"Recompute the §4.1 generative analysis on the same 1,000-painting sample under two controlled conditions: (a) metadata-stripped versions of each perspective with all artist names, titles, dates, and style/movement terms removed, and (b) a metadata-only control (title + artist + style) with no perspective content, plus an empty-prompt baseline. Require that the leave-one-out ranking (historical as least replaceable) and the formal-vs-narrative CLIP ordering survive; if the historical leave-one-out gain collapses or the formal advantage reverses under (a), the complementarity claims are confounded by text-to-image priors rather than perspective content.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that \"no single perspective suffices for all tasks\" rests on the task-asymmetry patterns in §4.1. But the perspective conditions differ not only in perspective content; they also differ in the annotation source and in the discriminative metadata they carry. Historical texts are produced by ArtRAG from retrieved biographies, movement pages, and event documents, so they naturally contain artist names, titles, dates, and style/movement terms. Formal texts from GalleryGPT use vocabulary such as chiaroscuro, impasto, and palette, which text-to-image models are strongly conditioned on. Narrative texts from Qwen3-VL emphasize concrete scene entities and spatial layout, which retrieval embeddings favor. As a result, Fig. 4's \"formal preserves style\" and Fig. 5's \"historical is non-discriminative\" may measure T2I and retrieval-model priors over artist/style keywords, not how much visual information each perspective encodes. Section 4.1 explicitly invokes reconstruction as a \"diagnostic probe\" (Eq. 3), but no null condition (title-only, style-only, empty prompt) and no ablation removing proper nouns or metadata is reported. Because both the complementarity conclusion and the \"historical is least replaceable\" claim are read directly off these comparisons, this confound is load-bearing for the Abstract's causal interpretation, even though the dataset resource itself may still be valuable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces MMArt, a dataset of 74,234 WikiArt paintings annotated with four independently generated interpretive perspectives (narrative, formal, emotional, historical) plus a harmonized unified caption. The perspectives are produced with specialized models: Qwen3-VL for narrative, GalleryGPT for formal analysis, ArtEmis-conditioned Qwen3-VL for emotion, and ArtRAG for historical context. The paper validates the resource with semantic distinctiveness statistics, an LLM-as-judge quality assessment, a blinded human evaluation against Claude 4.5 Sonnet, and two complementarity analyses: text-to-image reconstruction and perspective-guided retrieval over the full gallery. The stated central contribution is that the two analyses show task-asymmetric usefulness and that no single perspective suffices for all tasks.","tokens_in":13592,"tokens_out":6104,"duration_ms":58383,"significance":"If the complementarity claims hold, MMArt would be a valuable and timely resource: it is the first large-scale dataset providing four aligned interpretive perspectives over the same image set, and the public release of data, code, and embeddings would facilitate perspective-conditioned training, art VQA, and retrieval-augmented generation. The paper has notable methodological strengths, including the use of specialized generation models, two independent text-to-image generators, two architecturally distinct embedding models, an external-family LLM judge, and human evaluation. The remaining concerns are about the evidential basis for the headline cross-task claims rather than about the dataset resource itself.","major_comments":[{"comment":"The abstract's claim that leave-one-out analysis 'confirms that historical descriptions are the least replaceable perspective across both tasks' is contradicted by Fig. 5. In the Qwen3-VL-Embedding column, removing narrative (FEH) drops R@1 from 44.0 to 17.6, while removing historical (NFE) leaves R@1 at 34.8 and removing formal (NEH) leaves R@1 at 40.1. The retrieval leave-one-out results therefore identify narrative, not historical, as the least replaceable perspective for retrieval. The statement is supportable for the reconstruction metrics in Fig. 4 but not for retrieval, and it must be corrected or re-evidenced.","section":"Abstract; §4.1, Fig. 5"},{"comment":"The emotion agreement metric δ_emot is partly circular for the emotional perspective. e_emot is generated by conditioning on ArtEmis human affective utterances for the same painting, and δ_emot measures agreement between the original and reconstructed images on the ArtEmis top-1 emotion label. A high score can arise simply because the text-to-image model realizes the emotion name carried in the prompt, independent of whether the reconstructed image's visual content carries affective information. Please report emotion agreement with held-out human annotations or an independent affect classifier, and at minimum explicitly acknowledge this conditioning overlap as a limitation.","section":"§4.1, Eq. (3)"},{"comment":"The complementarity conclusions are confounded by annotation-source and metadata differences across perspectives. Historical texts are produced by ArtRAG with retrieved context and naturally contain artist names, dates, style, and movement terms; formal texts use vocabulary such as chiaroscuro and impasto; narrative texts emphasize concrete scene entities and spatial layout. Text-to-image and retrieval models may exploit these lexical and metadata priors rather than the visual information each perspective encodes. The paper should include null conditions or ablations, such as title-only prompts, metadata-only prompts, style-only prompts, and proper-noun-stripped perspectives, to show that the observed task asymmetry is not a prompt-prior artifact. This is load-bearing for the claim that 'no single perspective suffices for all tasks.'","section":"§4.1, Figs. 4 and 5"},{"comment":"The text states that all conditions are synthesized into ~80-word prompts so that differences in fidelity are attributable to perspective content rather than formatting, but the synthesis itself is performed by a prompt-following LLM. No check is reported that the synthesized prompts preserve the perspective-specific information or that the downstream generator actually follows the synthesized prompt. Reporting prompt-fidelity statistics or a small human inspection of the synthesized prompts would strengthen the diagnostic interpretation of the reconstruction results.","section":"§4.1, Fig. 4"}],"minor_comments":[{"comment":"The sentence 'We have nince annotators participated included researchers with graduate-level training in visual cultural and art understanding' contains a typo and ungrammatical phrasing; please revise and report the recruitment and rater-agreement statistics more precisely.","section":"§4.3"},{"comment":"The phrase 'which probing what visual information each perspective encodes' is ungrammatical; it should read 'probing which visual information each perspective encodes.'","section":"§4.1"},{"comment":"The figure caption and x-axis use 'Full' for what appears to be the unified caption condition, while the text sometimes refers to 'the full four-perspective condition'; please clarify whether the full condition is the unified caption or the concatenation of all four perspectives.","section":"Fig. 4"},{"comment":"Equation (1) is missing a comma between e_hist and e_unif.","section":"Eq. (1)"},{"comment":"The phrase 'elength under 30 words are nulled' should read 'explanations with length under 30 words are nulled.'","section":"§3.3"},{"comment":"The nine conditions are described as being evaluated on 'the same stratified 1,000 painting samples,' but retrieval is performed over the full 74k gallery; please state explicitly that the queries are the 1,000-painting subset while the gallery contains all 74,234 paintings.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The dataset resource itself is a solid contribution and I would not oppose publication after revision. The main fixes are to align the abstract and conclusions with the retrieval table, add controls for metadata and prompt priors in the complementarity analyses, and de-circularize the emotion metric. If the additional controls cannot be included, the paper should be reframed as a dataset resource with preliminary validation rather than as proof of perspective complementarity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: MMArt is a genuinely new resource — 74k WikiArt paintings each with narrative, formal, emotional, and historical perspectives plus a unified caption. No prior dataset covers all four for the same artworks, and the paper does a reasonable job of generating them with specialized models (Qwen3-VL, GalleryGPT, ArtRAG, ArtEmis-conditioned generation) and filtering for quality. The two-directional validation idea, reconstruction vs retrieval, is a good way to show that perspectives are not interchangeable, and the task asymmetry is interesting (narrative drives retrieval, formal drives reconstruction).\n\nThe dataset construction itself looks solid. Quality control is described in detail, semantic distinctiveness is checked, and the human evaluation against Claude 4.5 is decent evidence that the per-perspective descriptions are competitive.\n\nThe soft spots are in the interpretation, not the resource. First, the abstract says leave-one-out analysis shows historical is the least replaceable perspective across both tasks. That is not what the retrieval table shows: when historical is absent (NFE) R@1 is 34.8, but when narrative is absent (FEH) it drops to 17.6. Historical is least replaceable only for reconstruction, where it contributes the largest marginal gain. Second, the emotion reconstruction metric is partly circular: e_emot is generated conditioned on ArtEmis utterances, and delta_emot counts agreement with ArtEmis top-1 emotion labels, so part of the measured signal comes from the conditioning labels themselves. Third, and most important, the complementarity analyses have no control for metadata leakage into specific perspectives. Historical texts come from ArtRAG output that includes artist names, dates, and style/movement terms; formal texts use vocabulary like chiaroscuro and impasto. Those tokens can drive text-to-image priors and retrieval embeddings more than the perspective's actual visual content. A null condition (title-only, style-only) or an ablation stripping proper nouns and style terms is needed to back the claim that the asymmetry measures the perspectives themselves. Fourth, there is no direct comparison against existing single-perspective datasets like ArtEmis or SemArt as baselines, so the multi-perspective advantage over a single generalist caption is only partially demonstrated.\n\nThese are fixable. The paper deserves peer review; I would send it, but the authors need to temper the abstract and add the missing controls before the strong claims become citeable. The resource itself is valuable and likely to be used.","headline":"Worth a serious referee, but the paper oversells its complementarity evidence and needs controls for metadata leakage before its strong claims are citeable.","tokens_in":14142,"tokens_out":3183,"would_cite":true,"duration_ms":31223,"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":"The paper claims that describing a painting from four independent perspectives — narrative, formal, emotional, and historical — produces a dataset in which no single perspective is best for every art-understanding task.","keywords":["visual art understanding","multi-perspective dataset","vision-language models","perspective complementarity","image retrieval","text-to-image reconstruction","affective computing","art interpretation"],"falsifier":"Run the generative analysis with a text-to-image model never trained on paintings and a generic control prompt like 'a painting' for every artwork; if the control matches originals on the style and composition metrics as closely as the formal perspective does, the reconstruction probe is measuring generator priors rather than perspective content.","tokens_in":13105,"feed_emoji":"🎨","tokens_out":7060,"duration_ms":144154,"temperature":0.7,"pith_summary":"The paper sets out to fix a data problem: art datasets each offer only one way of describing a painting, so models learn to caption surface content rather than interpret form, feeling, and context. To change that, the authors build MMArt, a dataset of 74,234 paintings in which every work carries four independent descriptions — what is depicted, how the work is formally composed, what it feels like, and why it looks and means as it does — plus a unified caption that merges the four. The central claim is that these perspectives are genuinely complementary and task-asymmetric: narrative text is the strongest retrieval query, formal analysis best preserves composition in image reconstruction, and historical context is the least replaceable perspective when any one is removed. Validation comes from two directions: a generative probe that reconstructs paintings from text and measures style, composition, and emotional agreement, and a discriminative probe that uses each perspective to retrieve the correct painting from the full collection. A reader should care because, if right, MMArt supplies the missing supervision for models that interpret artworks across multiple registers at once.","feed_headline":"A 74,234-painting dataset annotates art four ways at once","feed_subtitle":"Narrative, formal, emotional and historical views prove task-asymmetric: retrieval favors narrative, reconstruction favors formal.","key_machinery":"The load-bearing mechanism is a four-perspective annotation scheme in which each painting is described independently by narrative, formal, emotional, and historical viewpoints, following a layered scheme of art interpretation; each perspective is produced by a model chosen for that interpretive register rather than by one generalist model. The other carrying element is a two-directional complementarity analysis. A generative probe feeds perspective text to text-to-image models and measures reconstruction fidelity with three complementary metrics (style and semantic similarity, compositional similarity, and emotional agreement), while a discriminative probe embeds all paintings into a fixed gallery and measures how well each perspective retrieves the originating painting. Together these probes convert the qualitative idea of 'different perspectives' into measurable task asymmetry that persists across multiple generators and embedding models.","core_discovery":"The discovery is that different ways of describing a painting carry different, only partially overlapping information, and the ranking of perspectives flips between tasks. In the generative direction, formal descriptions reconstruct compositional style better than other single perspectives, and the unified caption achieves the highest overall fidelity on both style and composition metrics. In the discriminative direction, narrative descriptions retrieve the correct painting from the full gallery 44.0 percent of the time at rank one, while formal descriptions, despite their reconstruction strength, reach only 7.8 percent, and historical descriptions are nearly non-discriminative. Leave-one-out analysis shows historical context is the hardest perspective to replace across both tasks. The paper concludes from this task asymmetry that no single perspective suffices for all art-understanding tasks, which directly motivates the dataset's multi-perspective design.","pith_inferences":["The paper leaves implicit that the same reconstruction-plus-retrieval protocol could be reused as a generic diagnostic for any dataset that annotates one image in several registers, not only art.","Because all pairwise perspective embedding similarities are below 0.55, a testable extension is to measure whether substituting one perspective for another actually changes downstream question-answering answers; if it does not, some of the claimed distinctiveness may not translate into task-relevant gains.","A practical extension is to ablate the human affective utterances from the emotional-perspective pipeline; if judge-assessed fidelity drops sharply, that would demonstrate where human grounding contributes inside a synthetic annotation pipeline."],"forward_implications":["Models can be trained with explicit perspective conditioning, so a user can ask for a formal analysis, an emotional reading, or a historical account of the same painting.","Art visual question answering can be organized around perspective, because different questions depend on different annotation registers.","Retrieval systems can select query perspective by task: narrative text for finding a specific painting, formal and historical text for browsing by style or period.","Retrieval-augmented generation over art gains query-side historical grounding, since historical descriptions carry facts that are not visible in the image.","Benchmark designers can use the unified caption as a single-text upper bound, since it slightly outperforms every single perspective on retrieval while matching the best reconstructive conditions."],"supporting_citations":[{"why":"Supplies the human affective utterances that ground the emotional perspective.","marker":"[1]"},{"why":"Provides the specialized formal-analysis model whose output forms the formal perspective.","marker":"[5]"},{"why":"Provides the retrieval-augmented historical generation that forms the historical perspective.","marker":"[37]"},{"why":"Provides the vision-language model used for narrative, emotional, and unified-caption generation.","marker":"[2]"},{"why":"Defines the CLIP similarity metric used to score style and semantic reconstruction fidelity.","marker":"[26]"},{"why":"Defines the DINOv3 similarity metric used to score compositional reconstruction fidelity.","marker":"[32]"},{"why":"Provides one of the two retrieval embedding models that rank the full gallery.","marker":"[17]"},{"why":"Provides the other retrieval embedding model used to confirm results are model-agnostic.","marker":"[20]"},{"why":"Supplies one of the two text-to-image generators used in the reconstruction probe.","marker":"[6]"},{"why":"Supplies the second text-to-image generator that confirms reconstruction rankings are not generator-specific.","marker":"[42]"}],"fun_headline_variants":["Art dataset reveals perspective flips between retrieval and reconstruction","Four art views: narrative beats retrieval, formal beats reconstruction","74k paintings, four lenses: which art annotation wins depends on task","One painting, four annotations: task decides which art view matters","MMArt: multi-view art data shows no single description fits all tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The complementarity conclusions rest on treating text-to-image reconstruction fidelity as a direct measure of how much visual information a perspective encodes; if the generators rely on their own style priors or obey some prompts better than others, the resulting task rankings would not actually be measuring the perspectives.","fun_headline_variants_meta":{"raw":{"variants":["Art dataset reveals perspective flips between retrieval and reconstruction","Four art views: narrative beats retrieval, formal beats reconstruction","74k paintings, four lenses: which art annotation wins depends on task","One painting, four annotations: task decides which art view matters","MMArt: multi-view art data shows no single description fits all tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000722,"raw_usage":{"total_tokens":3249,"prompt_tokens":961,"completion_tokens":2288,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":2201}},"tokens_in":577,"tokens_out":2288,"duration_ms":17793,"temperature":1.0,"reasoning_tokens":2201,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:56:21.166833+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the generative analysis with a text-to-image model never trained on paintings and a generic control prompt like 'a painting' for every artwork; if the control matches originals on the style and composition metrics as closely as the formal perspective does, the reconstruction probe is measuring generator priors rather than perspective content.","supporting_citations":[{"cited_title":"In: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021)","cited_arxiv_id":null,"evidence_quote":"Supplies the human affective utterances that ground the emotional perspective."},{"cited_title":"In: Proceedings of the 32nd ACM International Conference on Multimedia","cited_arxiv_id":null,"evidence_quote":"Provides the specialized formal-analysis model whose output forms the formal perspective."},{"cited_title":"In: Proceedings of the 33rd ACM International Conference on Multimedia","cited_arxiv_id":null,"evidence_quote":"Provides the retrieval-augmented historical generation that forms the historical perspective."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies one of the two text-to-image generators used in the reconstruction probe."}],"review_version":1}