{"id":"ca8d6773-1df0-49be-b201-a6dd8bfa389f","arxiv_id":"2412.18783","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ArtNVG combines CSGO-style content/style separation with neighboring-view attention sharing to produce locally consistent stylized 3D Gaussian Splatting scenes from a single style reference image.","lead":"ArtNVG applies the artistic style of a single reference image to an existing 3D scene built by Gaussian Splatting, while keeping colors and textures consistent across neighboring views. It is a diffusion-based stylization pipeline for film and game production that reports better consistency and content preservation than two current 3D stylization baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Neighboring-View Attention of Eq. (7) concatenates K/V across views chosen by camera-center distance alone; without view-pose conditioning or overlap validation, close-but-divergent cameras can inject conflicting content, so the local-consistency claim is not yet robustly supported.","rationale":"The reader's weakest assumption identifies the same mechanism, and I agree it is the load-bearing point. The paper presents a coherent pipeline with plausible design and positive results on two baselines, including an ablation and a user study. However, the specific operation that supposedly guarantees consistency—concatenating K/V across views—has an implicit structural assumption that is not checked. This is not a matter of matching current consensus; it is an internal correctness risk: if the assumption fails, the method's headline advantage over 'w/o NV Attention' in Table 2 could reverse in an adversarial but realistic capture setting. Because the paper does provide ablations, a user study, and a clear description of the pipeline, the appropriate outcome is CONDITIONAL rather than rejection: the condition is demonstrating that NV attention remains beneficial when neighbor selection is based on actual view overlap rather than camera-center distance, and across divergent-view configurations. I therefore leave the reader's verdict unchanged.","tokens_in":14474,"tokens_out":5672,"duration_ms":52176,"concrete_test":"Run the same pipeline on a Tanks and Temples scene (e.g., 'Ignatius' or 'Horse') with two neighbor-selection rules: (a) k-nearest camera centers as described, and (b) neighbors chosen by actual 3D overlap, e.g., highest IoU of view frustums or most shared visible Gaussians. For each rule, report CLIP-DC, CFSD, and user-rated artifacts. In addition, construct an adversarial viewpoint set in which cameras are close in center distance but face away from each other (e.g., alternating inward/outward orientations around a central object) and compare the NV-attention stylized images against a no-attention baseline. If (b) substantially outperforms (a), or if divergent-view pairs show color/geometry bleeding, then the camera-center-distance criterion is insufficient and the consistency claim requires qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—superior local consistency—rests on the Neighboring-View Attention mechanism (Sec. 3.2.3, Eq. 7–8). For each view i, the self-attention query attends to the concatenated keys/values of N 'neighboring' views. This is only beneficial if the K/V features of different views are mutually compatible in the diffusion latent space; otherwise, attention mixes features from different scene content, causing blurring or ghosting. The paper selects neighbors solely by camera-center distance ('camera centers are closer'), which does not guarantee visual overlap in 360° scenes: nearby cameras can point in divergent directions (e.g., opposite sides of a statue), so their latent tokens correspond to different 3D points. No positional encoding, relative-pose conditioning, or epipolar constraint is added to the attention to align tokens across views. Moreover, the grouping algorithm is underspecified: it is not stated whether a view belongs to a single group or multiple overlapping groups, nor how overlapping groups are reconciled. The reported CLIP-DC metric samples adjacent frames along a path, which likely fall in the same neighbor group, so discontinuities at group boundaries or under divergent views are not directly measured. Thus, the consistency guarantee is not demonstrated for a plausible and untested failure regime.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ArtNVG, a 3D Gaussian Splatting (3DGS) stylization framework that applies the style of a reference image to an existing 3D scene. The method combines two main ideas: (1) Content-Style Separated Control, which uses the CSGO projection modules and a Tile ControlNet to inject content and style controls separately into cross-attention layers, and (2) Attention-based Neighboring-View Alignment, which replaces self-attention layers in the UNet and ControlNet with an attention mechanism that concatenates keys and values from nearby views (Eq. 7-8) to improve local color and texture consistency. Stylized renderings are used to finetune the original 3DGS scene with an L1 plus NNFM loss (Eq. 10). The authors report qualitative comparisons, a user study, and quantitative metrics (CFSD, CSD, CLIP-DC) against StyleGaussian and InstantStyleGaussian, claiming superior content preservation, style alignment, and local consistency.","tokens_in":14707,"tokens_out":3799,"duration_ms":37117,"significance":"If the claims are substantiated, ArtNVG would be a useful system for fast, reference-guided 3D stylization, a task with growing demand in film and games. The main technical novelty is the Neighboring-View Attention mechanism, which is simple and intuitively plausible, and the use of CSGO/Tile ControlNet for separated content and style control is a sensible application of externally validated components rather than a fundamentally new model. The paper's strengths include a clear system description, a user study, supplementary extended results, and an ablation that isolates the contribution of Neighboring-View Attention to local consistency. However, the quantitative evaluation is thin and does not yet establish the central consistency claim: the grouping mechanism is underspecified, the evaluation metrics are aggregated without error bars or significance tests, and the Content-Style Separated Control contribution is never ablated. These issues require further evidence before the paper's claims can be accepted.","major_comments":[{"comment":"The Neighboring-View Attention mechanism selects neighbors solely by camera-center distance ('camera centers are closer'), but in unbounded 360-degree scenes, cameras with close centers can point in divergent directions and thus have little or no visual overlap. In such cases, concatenating K and V from different views (Eq. 7) mixes tokens that correspond to different 3D content, which can blur details or inject conflicting content into the attention output. The paper does not add any relative-pose conditioning, positional encoding, epipolar constraint, or overlap validation to the attention mechanism, nor does it specify whether the neighboring-view groups are disjoint or overlapping and how the groups are reconciled. This is load-bearing because the central claim of improved local consistency relies on mutual compatibility of K/V features across views. Please specify the grouping algorithm precisely, validate compatibility for views with large baselines and divergent orientations, and report consistency metrics at group boundaries and across randomly selected view pairs rather than only along adjacent frames that likely fall in the same group.","section":"Sec. 3.2.3, Eq. (7)-(8)"},{"comment":"The quantitative comparison reports only three aggregate metrics averaged over eight scenes, with no per-scene breakdown, no standard deviations, and no significance tests. The reported differences are small in several cases (e.g., CLIP-DC 0.81 vs. 0.79), so it is unclear whether the claimed superiority is robust or within run-to-run variation. Please provide per-scene results, error bars or confidence intervals, and, ideally, paired significance tests across scenes and styles. This is necessary to support the abstract and Section 4 claim that ArtNVG 'surpasses existing methods' on all three metrics.","section":"Sec. 4.4, Table 1"},{"comment":"The ablation study validates the Neighboring-View Attention and the NNFM loss, but it does not ablate the Content-Style Separated Control, which is listed as Contribution 2 and is central to the claimed reduction of 'information leakage.' Without a variant that replaces the CSGO/Tile ControlNet content and style controls with an inversion-based content preservation approach, the experiments do not directly support the claim that separated control reduces leakage or improves content fidelity. The 'Train from Scratch' variant ablates the pre-trained scene geometry, not this proposed component. Please add an ablation that isolates the effect of Content-Style Separated Control.","section":"Sec. 4.5, Table 2"},{"comment":"The CLIP-DC metric in Eq. (16) measures the consistency of CLIP edit directions between adjacent frames on a rendered camera path. The paper does not specify the camera path used, and it is likely that adjacent frames on such a path are precisely the views that are clustered together in the same neighboring-view group. As a result, CLIP-DC may not detect discontinuities at group boundaries or between views with larger baselines. Please describe the path and, more importantly, evaluate consistency across group boundaries and across pairs of views that are not in the same group, since those are the cases where the proposed alignment mechanism is most at risk.","section":"Sec. 3.2.3 and Sec. 4.4"}],"minor_comments":[{"comment":"Equation (6) uses the predicted noise epsilon_t as the added noise in the DDIM update, which is not the standard DDIM sampling or inversion step and appears to be a typo; please clarify the actual sampling/inversion procedure used.","section":"Eq. (6)"},{"comment":"The notation around Neighboring-View features is difficult to follow: K_NV and V_NV are defined as concatenations of K_i and V_i for N views, but the figure does not clearly indicate whether the query comes from a single view and whether the same group is used for all self-attention layers; please clarify with a concise formal description of the grouping and attention shapes.","section":"Sec. 3.2.3 and Fig. 2"},{"comment":"The 'Ours' row in Table 2 (CFSD 0.12, CSD 0.15, CLIP-DC 0.80) differs from the 'Ours' row in Table 1 (CFSD 0.14, CSD 0.13, CLIP-DC 0.81), which is confusing because the reader may expect identical numbers; please state explicitly that the tables report averages over different scene subsets.","section":"Sec. 4.4 and Sec. 4.5"},{"comment":"In the NNFM loss definition, the min operation is written as 'min Dcos(Fr(i,j), Fs(i,j))' without specifying the index over which the minimum is taken; please write the index explicitly, following the ARF formulation.","section":"Supplementary, Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a systems-style contribution that combines CSGO, Tile ControlNet, and a straightforward attention-sharing mechanism. The central consistency claim is plausible but currently rests on an underspecified grouping mechanism and a thin quantitative evaluation. I did not find evidence of circularity or inappropriate fitting of metrics; the issue is insufficient validation, not methodological dishonesty. With the requested ablations and a more rigorous evaluation, the paper could be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nArtNVG is a solid system paper for 3DGS stylization, and the central claim survives a close read. The authors split content and style control using CSGO's projection modules plus a Tile ControlNet, inject them into different cross-attention layers of SDXL, and replace self-attention with a Neighboring-View attention that concatenates K/V from camera-adjacent views. That combination is new relative to StyleGaussian and InstantStyleGaussian, and the qualitative results show genuinely cleaner local detail than either baseline. The best evidence is the ablation: dropping NV attention degrades all three metrics, and the N-sweep shows the choice of 15 came from a real curve, not a guess. The components are externally validated, so there is no circularity worry.\n\nThe soft spots are in the evaluation, not the idea. Eight scenes, single grouped averages, no error bars, no significance tests. CLIP-DC measures editing-direction consistency in CLIP space, not pixel-level color or texture agreement, so the paper's marquee local-consistency claim is carried mainly by the user study and images, not the tables. The stress-test concern about camera-center distance is fair as a robustness gap: camera center proximity does not guarantee visual overlap in 360-degree scenes, there is no pose conditioning or overlap check on the shared K/V, and the grouping algorithm — disjoint clusters or overlapping windows? — is underspecified. I read it as an untested failure regime rather than a demonstrated flaw; nothing here contradicts the mechanism working on standard captures, and the evaluation never exercises the divergent-view case. For reproduction, the missing code and underspecified neighbor-selection details matter more. Minor: two baselines is a narrow comparison, and InstantStyleGaussian is a self-declared faithful reproduction.\n\nWho this is for: people working on diffusion-based 3D stylization or multi-view consistency in 3DGS editing, and production pipelines given the 20-minute run. It deserves a serious referee. I would send it to review and ask for per-scene results with variance, a precise grouping procedure, a divergent-view test case, and code before publication.","headline":"A genuine, moderately novel 3DGS stylization system whose central claim holds, but the local-consistency evidence is thinner than the tables suggest and its key mechanism has an untested divergent-view regime.","tokens_in":15282,"tokens_out":6950,"would_cite":true,"duration_ms":57326,"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":"ArtNVG stylizes 3D scenes in 20 minutes by separating content and style controls and sharing attention across nearby views.","keywords":["3D Gaussian Splatting","3D stylization","diffusion models","content-style separation","neighboring-view attention","multi-view consistency","style transfer","zero-shot stylization"],"falsifier":"Run the pipeline on a scene captured with wide-baseline cameras, so neighboring camera centers are far apart and overlapping content is small, then measure CFSD and CLIP-DC against the no-NV-attention variant; if consistency scores do not improve, or if visible seams appear where clusters meet, the key-value consistency assumption fails.","tokens_in":1670,"feed_emoji":"🎨","tokens_out":1633,"duration_ms":54552,"temperature":0.7,"pith_summary":"This paper introduces ArtNVG, a pipeline that applies the artistic style of a reference image to an existing 3D Gaussian Splatting scene in about 20 minutes. The core claim is that two design choices fix the weaknesses of prior 3D stylization: separating content and style controls in a diffusion model (using the CSGO projection modules and a Tile ControlNet) instead of inverting the content image, and replacing ordinary self-attention with a Neighboring-View Attention that lets nearby views share key and value features during denoising. Together these are said to preserve scene content, match the reference style, and keep local colors and textures consistent across views. The paper reports better CFSD, CSD, and CLIP-DC scores than StyleGaussian and InstantStyleGaussian, plus higher user preference.","feed_headline":"Two controls and shared attention stylize 3D scenes in 20 minutes","feed_subtitle":"ArtNVG keeps content and style apart and joins nearby views, beating prior 3D stylization on fidelity, style, and consistency.","key_machinery":"The load-bearing mechanism is the Attention-based Neighboring-View Alignment layer: for N views, the keys and values are concatenated as $K_{\\mathrm{NV}} = [K_1, \\ldots, K_N]^T$ and $V_{\\mathrm{NV}} = [V_1, \\ldots, V_N]^T$, and each view's query attends to all neighboring keys and values, with $\\mathrm{NVAttn}_i = \\mathrm{Attn}(Q_i, K_{\\mathrm{NV}}, V_{\\mathrm{NV}})$. This replaces every self-attention layer in the UNet and the content control, so the denoising of each view sees shared local features. Paired with Content-Style Separated Control (the CSGO style projection for style and the CSGO content projection plus Tile ControlNet for content), the diffusion process keeps content and style on separate channels, reducing information leakage.","core_discovery":"The paper's central discovery is a zero-shot 3D stylization method built on 3D Gaussian Splatting that avoids two known failure modes: information leakage between content and style, and local inconsistency among neighboring views. Content and style are injected through separate cross-attention controls, and self-attention layers in both the UNet and the content control module are replaced by NV attention, which concatenates the keys and values of N camera-center-close views into one attention computation. The stylized renderings then finetune the original 3DGS scene, with an added NNFM loss to preserve high-frequency texture details. On the Tanks and Temples dataset with WikiArt styles, the method reportedly outperforms both baselines on content fidelity, style similarity, and consistency metrics.","pith_inferences":["An extension beyond the paper: the NV-attention principle should transfer to other diffusion-based 3D editing tasks such as text-guided editing or relighting, since it only swaps self-attention layers.","An extension beyond the paper: the view-grouping criterion based on camera-center distance could be made content-aware, for example by grouping views with high visual overlap or similar depth maps, which may improve robustness on large-baseline captures.","An extension beyond the paper: separating content and style controls could combine with other ControlNet variants to add user interaction such as object removal or addition, which the paper lists as a limitation.","A testable extension beyond the paper: measuring local consistency with pixel-level correspondences or optical flow across neighboring stylized views would provide a more direct check than the CLIP-DC metric reported."],"forward_implications":["Stylization of an existing 3DGS scene takes about 20 minutes total, with no extra pre-training of new networks.","Content and style are controlled through separate cross-attention inputs, reducing content leakage from the style image and style leakage from the content images.","Neighboring views are denoised jointly rather than independently, improving local color and texture consistency.","Quantitative evaluation over eight scenes shows improvements on CFSD, CSD, and CLIP-DC relative to both baselines, and a user study favors ArtNVG on content fidelity, style alignment, and visual quality.","The number of neighboring views N trades off cluster fragmentation against inter-cluster continuity, with N = 15 yielding the best consistency in their experiments."],"supporting_citations":[{"why":"Introduces 3D Gaussian Splatting, the 3D representation that ArtNVG stylizes and finetunes.","marker":"[19]"},{"why":"CSGO supplies the content and style projection modules used to decouple content and style controls.","marker":"[47]"},{"why":"The Tile ControlNet, built on this control mechanism, serves as the content control module conditioned on style.","marker":"[56]"},{"why":"ARF's NNFM loss is used during finetuning to align textures and details with the style image.","marker":"[55]"},{"why":"StyleGaussian is the feedforward VGG-based baseline that ArtNVG compares against.","marker":"[23]"},{"why":"InstantStyleGaussian is the diffusion-based baseline using inversion-based content preservation that ArtNVG compares against.","marker":"[52]"},{"why":"StyleNeRF2NeRF shows that attention-based sharing improves multi-view consistency, a precedent for NV attention.","marker":"[11]"},{"why":"GaussCtrl uses attention-based latent alignment for multi-view consistency, another predecessor of NV attention.","marker":"[46]"},{"why":"CFSD is the metric used to evaluate content fidelity while suppressing style influence.","marker":"[7]"},{"why":"CSD is the metric used to evaluate style similarity, and it is described as performing well on WikiArt.","marker":"[37]"}],"fun_headline_variants":["ArtNVG decouples content and style for 3D stylization","3D stylization with separated content and style control","Neighboring-view attention boosts 3D style consistency","ArtNVG splits content, style, aligns views for 3D art","Gaussian stylization via decoupled controls and view alignment"],"cache_read_input_tokens":17408,"weakest_assumption_plain":"The method assumes that the self-attention key and value features from camera-center-close views are mutually consistent in the diffusion latent space, so that concatenating them reinforces shared details instead of injecting conflicting content.","fun_headline_variants_meta":{"raw":{"variants":["ArtNVG decouples content and style for 3D stylization","3D stylization with separated content and style control","Neighboring-view attention boosts 3D style consistency","ArtNVG splits content, style, aligns views for 3D art","Gaussian stylization via decoupled controls and view alignment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000123,"raw_usage":{"total_tokens":1089,"prompt_tokens":922,"completion_tokens":167,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":79}},"tokens_in":538,"tokens_out":167,"duration_ms":2302,"temperature":1.0,"reasoning_tokens":79,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:28:07.774968+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on a scene captured with wide-baseline cameras, so neighboring camera centers are far apart and overlapping content is small, then measure CFSD and CLIP-DC against the no-NV-attention variant; if consistency scores do not improve, or if visible seams appear where clusters meet, the key-value consistency assumption fails.","supporting_citations":[{"cited_title":"splatfacto","cited_arxiv_id":null,"evidence_quote":"The Tile ControlNet, built on this control mechanism, serves as the content control module conditioned on style."},{"cited_title":"Arf: Artistic radiance fields","cited_arxiv_id":null,"evidence_quote":"ARF's NNFM loss is used during finetuning to align textures and details with the style image."},{"cited_title":"Style-NeRF2NeRF: 3D Style Transfer From Style-Aligned Multi-View Images","cited_arxiv_id":"2406.13393","evidence_quote":"StyleNeRF2NeRF shows that attention-based sharing improves multi-view consistency, a precedent for NV attention."},{"cited_title":"Style in- jection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer","cited_arxiv_id":null,"evidence_quote":"CFSD is the metric used to evaluate content fidelity while suppressing style influence."}],"review_version":1}