{"id":"8ad140de-58be-4fab-86b3-4514cca38c31","arxiv_id":"2505.04813","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A method that optimizes 3D Bezier curves to abstract a 3D shape, capturing geometry and texture through CLIP-guided losses, with keypoint-based user control and SDF-based deformation handles.","lead":"WIR3D turns 3D models into sparse 3D curve sketches that preserve both shape and textures, using CLIP to guide where strokes go. It adds interactive control over which features get abstracted and lets the curves drive shape deformation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Arbitrary-view fidelity is asserted but never evaluated outside the 0–30° elevation band used for training; with one random view per iteration and no per-view error bars, the central 'high fidelity across arbitrary views' claim is unsupported for high-elevation views.","rationale":"After reading the paper in good faith, the method's core pipeline is coherent, the ablations support the importance of stage I, SDF, and intermediate CLIP layers, and the user studies (96 and 42 respondents) are genuine evidence that WIR3D outperforms 3Doodle and ARAP on the tested views. The reader's identified weakest assumption about CLIP spatial localization is plausible and worth probing, but it is supported by a large body of prior work (CLIPasso uses intermediate spatial activations; CLIP-based image editing relies on spatial features), so I do not think it is the single most fragile point. The most load-bearing gap is the mismatch between the central claim of 'high fidelity across arbitrary views' and an evaluation that samples only 0–30° elevation, averages metrics without confidence intervals, and offers no held-out elevation or per-view analysis. Because the optimization uses one random view per iteration and no explicit consistency loss, a view-dependent failure mode would not be detected by the reported metrics. This does not invalidate the method, but it means the headline claim is broader than the evidence. The appropriate verdict remains CONDITIONAL (i.e., no change from the reader), pending the concrete test above and release of code/data.","tokens_in":15954,"tokens_out":9133,"duration_ms":95582,"concrete_test":"After optimizing the released code (or reimplementing Sec. 3) on a subset of the reported shapes, render each curve set and the target from a fixed dense grid of azimuths (0°, 90°, 180°, 270°) at elevations −30°, 0°, 30°, 60°, and 90° and compute per-view LPIPS (AlexNet) and CLIP ViT/B-32 cosine similarity. Report distributions and per-elevation means with at least 3 random seeds. Also run a small forced-choice user study (N≥20) on high-elevation views only. If per-view scores and user preferences degrade significantly for elevations outside 0–30°, the 'arbitrary views' claim should be restricted to the trained elevation band or the method needs an explicit multi-view consistency loss.","verdict_should_be":"UNCHANGED","load_bearing_attack":"WIR3D's central claim is that the abstractions 'maintain high fidelity across arbitrary views' and represent a shape 'from all possible viewing directions.' The optimization, however, samples exactly one view per iteration, all from elevations 0–30° and azimuth 0–360° (Sec. 4). There is no explicit loss enforcing multi-view consistency beyond the shared 3D curve parameters, and the CLIP and LPIPS terms are somewhat insensitive to geometric misalignment (the paper itself notes CLIP insensitivity in Sec. 4.2). Consequently, the optimizer could satisfy the average appearance over the training distribution while producing curves that look wrong from high-elevation or other untrained viewpoints. The only multi-view evidence is qualitative (Fig. 16, rotating gifs) and appears to sweep azimuth, not elevation. Quantitative Table 1 reports averages over novel views with no per-view breakdown, no confidence intervals, and no held-out elevation analysis. If high-elevation views are excluded, 'arbitrary views' is an overstatement, and the 'view-consistent 3D abstraction' contribution is weaker than claimed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces WIR3D, a method that abstracts a 3D shape into a sparse set of 3D cubic Bezier curves intended to represent both geometry and salient texture from arbitrary viewpoints. The method optimizes the curves with a two-stage procedure: stage I captures coarse geometry using CLIP semantic losses supervised by Freestyle renders, and stage II adds curves for texture using a novel localized keypoint loss that weights CLIP intermediate feature differences around projected 3D keypoints. An SDF regularization encourages curve adherence to the surface, and the authors demonstrate applications in interactive feature refinement and curve-based deformation. The evaluation includes qualitative results on textured and untextured shapes, quantitative comparisons with 3Doodle and NEF using LPIPS, CLIP image similarity, user preference, and a Chamfer-based coverage metric, plus ablation studies and two user studies.","tokens_in":16139,"tokens_out":4524,"duration_ms":50239,"significance":"If the central claims hold, WIR3D would be a useful contribution to 3D sketch abstraction, addressing a gap left by occlusion-contour methods and by dataset-dependent reconstruction approaches. The two-stage optimization, the localized keypoint weighting, and the SDF-based surface anchoring are reasonable and potentially transferable ideas. The paper's strengths include the breadth of qualitative results across varied shapes, the inclusion of two user studies that provide partially independent evidence, and the use of established baselines (3Doodle, NEF). The automatic keypoint detection via back-projection and clustering is also a sensible mechanism. However, the quantitative evidence is weaker than the narrative suggests: the headline metrics overlap in family with the optimization losses, no variance information is reported for stochastic optimization, and the 'arbitrary views' claim is not tested outside the training elevation band. These issues need to be addressed before the contribution can be fully assessed.","major_comments":[{"comment":"The quantitative comparison is partly circular. The optimization losses in Eq. (2) and Eq. (4) use cosine similarity on global CLIP embeddings plus L2 distances on intermediate CLIP features, with ResNet CLIP variants (RN101 and RN50x16), and the LPIPS term uses the VGG variant (supplemental Sec. C). The headline evaluation in Table 1 uses CLIP ViT/B-32 image similarity and AlexNet LPIPS. The architecture split reduces but does not eliminate the concern: Table 1 still measures the same family of perceptual losses that the method is explicitly trained against, so the reported superiority over 3Doodle on LPIPS and CLIP image similarity is partly an artifact of optimizing objectives of the same type. I do not regard this as fatal because the user study (N=96) is independent and is the strongest evidence in the paper, but the claim that WIR3D 'outperforms' baselines on the first two metrics is not independent confirmation. Please report per-shape results, metrics obtained from unrelated evaluators (e.g., human judgment already available), or at least discuss this limitation quantitatively.","section":"Sec. 3.2 and Sec. 4.2, Table 1"},{"comment":"No variance or confidence information is reported for any quantitative result. The optimization is stochastic: one random view is sampled per iteration, random initializations are used, and the CLIP supervision is augmented (supplemental Sec. G). Tables 1 and 2 report only point estimates. Consequently, small differences such as LPIPS 0.227 vs 0.229 (No SDF) or CLIP image similarity 0.909 vs 0.904 cannot be interpreted as meaningful, and even the larger apparent gains over 3Doodle lack error bars. Please report means with standard deviations or confidence intervals across multiple optimization runs and across views, and state the number of runs used.","section":"Sec. 4, paragraphs 1-2; Table 1 and Table 2"},{"comment":"The central claim that abstractions 'maintain high fidelity across arbitrary views' is not supported by the evaluation. All optimization views are sampled from elevations 0-30 degrees and azimuth 0-360 degrees, and no held-out or high-elevation evaluation is reported. Table 1 averages over novel views without a per-view breakdown, and the qualitative multi-view evidence (Fig. 16 and the supplemental) appears to demonstrate azimuthal rotation rather than a full sphere of viewpoints. Since there is no explicit multi-view consistency loss beyond the shared 3D curve parameters, it remains possible that the curves satisfy the training view distribution while misrepresenting the shape at high elevations. Please either add quantitative evaluation on held-out elevation bands (e.g., 30-60 and 60-90 degrees) and per-view error statistics, or soften the 'arbitrary views' claim to match the evaluated range.","section":"Sec. 4, first paragraph; Sec. 4.1; Fig. 16"},{"comment":"The localized keypoint loss rests on the assumption that CLIP intermediate activations carry spatial information that can be meaningfully localized through a Gaussian weight map. The paper cites Shomron and Weiser [51] for this, but that reference studies spatial correlation in CNN activations generally for value prediction, not CLIP feature localization specifically, and no experiment in the paper directly validates that CLIP layers 3 and 4 provide the required spatial resolution for the proposed weighting. This is load-bearing because the texture stage and the user-control application both depend on the localization. The 'No Local' ablation and the noisy-keypoint experiment in Fig. 9 show robustness and overall utility, but they do not establish that the keypoint projections correlate with the features being emphasized. Please provide a direct spatial-localization check, for example by visualizing CLIP activation maps at keypoint locations or by comparing optimization outcomes with keypoints placed on different features against ground-truth feature regions.","section":"Sec. 3.2, Eq. (4) and Sec. 3.3"}],"minor_comments":[{"comment":"Typo: 'progressvely' should be 'progressively'.","section":"Fig. 6 caption"},{"comment":"The paper uses inconsistent capitalization: 'Wir3D' appears in the user-study passages and in Fig. 24, while the method name is 'WIR3D' elsewhere. Please standardize.","section":"Sec. 4.2 and Sec. H"},{"comment":"The term 'Janusing artifacts' should likely be 'Janus artifacts', and reference [54] on generative AI is an unusual citation for this graphics-specific phenomenon; please use a more standard reference.","section":"Supplemental Sec. A"},{"comment":"The metric named 'Coverage' is a one-direction Chamfer distance from surface samples to curves. This measures proximity rather than coverage in the sense of how much of the surface is represented; please either rename the metric or add an explicit justification that this quantity captures coverage.","section":"Sec. 4.2, Table 1"},{"comment":"The paper states that code will be released 'in the near future' but no code is provided. Given that the method is an optimization with several tuned hyperparameters, releasing the code would substantially aid reproducibility; please state in the final version where and when the code will be available.","section":"Sec. 1 and Data Availability"}],"recommendation":"major_revision","confidential_remarks":"The paper presents an interesting and well-motivated method, and the user studies give some independent support for the core claims. The main weaknesses are the partially circular quantitative evaluation, missing variance information, and the unsupported 'arbitrary views' generalization. These are fixable within the scope of a revision, but they are load-bearing for the paper's strongest claims. I recommend major revision with careful attention to the evaluation protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: WIR3D is a solid methods paper with real contributions, but its headline claim about arbitrary-view fidelity is not backed by the experiments, which sample only 0–30° elevation. Everything else is in decent shape.\n\nWhat's new: the localized keypoint loss, the two-stage geometry/texture split, and the SDF-regularized curves as deformation handles are all real additions over 3Doodle and CLIPasso. The qualitative results are strong across diverse shapes, and the N=96 user study gives independent evidence that WIR3D beats 3Doodle on representing the target shape.\n\nThe main soft spot is the 'arbitrary views' claim. Optimization uses one view per iteration, all from 0–30° elevation and 0–360° azimuth. Table 1 averages over novel views but gives no per-elevation breakdown, and the rotating gifs sweep azimuth, not elevation. So 'maintain high fidelity across arbitrary views' is unsupported for high-elevation viewpoints. It's an overstatement, not a fatal flaw. View-consistency is built in by 3D curves, but that doesn't guarantee a good 2D sketch from a top-down angle.\n\nSmaller issues: no code released despite the stated plan; Tables 1 and 2 have no variance across stochastic runs; and the LPIPS/CLIP evaluation partially overlaps with the optimization losses, though they split architectures (VGG vs AlexNet, ResNet vs ViT-B/32) and the user study is independent. The localization assumption cites a 2019 CNN paper, which is weak, but the ablations (No Local, random keypoints) suggest the weighting does what it claims.\n\nThe paper is honest about its limitations, and the supplementary is thorough. None of the soft spots are load-bearing; the central method holds up.\n\nWho this is for: people working on sketch-based modeling, non-photorealistic rendering, or interpretable 3D abstraction. It deserves a serious referee. I'd send it to review with a request for high-elevation evaluation or a toned-down claim, error bars, and code.","headline":"WIR3D is a solid methods paper with real contributions, but its arbitrary-view fidelity claim outruns the 0–30° elevation evaluation.","tokens_in":16705,"tokens_out":4272,"would_cite":true,"duration_ms":38444,"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":"WIR3D turns 3D shapes into sparse, meaningful 3D curves that stay faithful from every viewing angle.","keywords":["3D shape abstraction","Bezier curves","sketch abstraction","CLIP supervision","localized keypoint loss","differentiable rendering","neural SDF","deformation handles"],"falsifier":"Run the stage-II optimization on the same shape twice, once with keypoints on a salient feature and once with the same number of keypoints on a flat, featureless region, keeping all other settings fixed; if the two abstractions are indistinguishable by user ranking or by L2 distance between curve sets, then the localized loss is not doing spatial work. A stronger quantitative version is to take ground-truth keypoint positions, project them into views, and check whether the argmax of the CLIP intermediate activation difference, or the recovered center of the weight map, falls within a few pixels of the projected keypoint; chance-level localization would falsify the paper's central mechanism.","tokens_in":15703,"feed_emoji":"✏️","tokens_out":5694,"duration_ms":52547,"temperature":0.7,"pith_summary":"WIR3D claims that a 3D shape can be abstracted into a sparse set of 3D Bezier curves that stay faithful to both the geometry and the salient visual features, including texture, from every viewing angle. The authors argue that this cannot be done with surface analysis alone, because occluding contours are view-dependent and ignore texture, so they supervise the curves with the intermediate activations of CLIP, a pretrained vision-language model. The optimization is split into a geometry stage and a texture stage, with the texture stage driven by a newly introduced localized keypoint loss that lets specific features be emphasized. A neural SDF loss keeps curves on the surface, which is what allows the curves to double as intuitive deformation handles and enables interactive feature refinement. A sympathetic reading of the results is that WIR3D establishes a new capability: semantically meaningful, user-controllable, view-consistent 3D sketch abstraction for arbitrary meshes, including noisy photo-reconstructed ones.","feed_headline":"Sparse 3D curves capture shape and texture from every view","feed_subtitle":"CLIP-guided strokes preserve geometry and salient features, then double as deformation handles.","key_machinery":"The load-bearing mechanism is the localized keypoint loss. Given a set of 3D keypoints, either user-selected or automatically detected by back-projecting CLIP features and clustering, each keypoint is projected into the current view and converted into a Gaussian weight map $I_{\\mathrm{weight}}$; this map is downsampled and multiplied into the per-pixel difference between CLIP intermediate activation maps of the curve render and the target render, so that optimization effort concentrates around the keypoint. The loss also includes a mean-pooled weight applied to the global CLIP cosine distance and an LPIPS term. Supporting it are a two-stage optimization, with CLIP RN101 for geometry on Freestyle contour renders and CLIP RN50x64 for texture on surface renders, a neural SDF loss $L_{\\mathrm{SDF}}$ that penalizes curve samples away from the zero level set, an NDC regularization keeping curves inside the frame, and differentiable rasterization of projected 3D Bezier control points.","core_discovery":"The central claim is that a set of cubic Bezier curves optimized against spatially weighted CLIP activations yields a 3D abstraction that is simultaneously sparse, view-consistent, and semantically faithful. In the paper's own framing, WIR3D 'can abstract a myriad of shapes from different domains with various visual concepts, geometric structures, and textures,' and the abstractions 'maintain high fidelity across arbitrary views.' The novel component is the localized keypoint loss: 3D keypoints are projected into each sampled view, a Gaussian weight map is constructed around them, and this map multiplies the difference between CLIP intermediate activations of the rendered curves and of the target shape, focusing optimization on the feature at the keypoint. The method also anchors curves to the surface with a neural SDF loss, so the optimized strokes are not just decorative but can serve as deformation handles. The authors demonstrate the claim qualitatively across textured and untextured shapes, on messy in-the-wild reconstructions, and with user studies favoring WIR3D over the 3Doodle baseline in 88% of pairwise comparisons.","pith_inferences":["A direct test of the localization assumption would be to measure whether CLIP intermediate activation maps predict the projected positions of known keypoints; the paper only tests this indirectly through the random-keypoint ablation, so the spatial-precision claim is the least-supported link.","The same spatially weighted CLIP supervision could be applied to abstractions in other modalities, such as video frames or multi-object scenes, where the keypoints would enforce temporal or spatial consistency.","Because the curves are bound to the surface through the SDF loss, they could serve as a general rigging primitive; skinning weights derived from curve distance might extend beyond the proof-of-concept deformation shown here.","The automatic keypoint detector inherits the biases of the CLIP features used for back-projection, so the method's notion of 'salient feature' may silently be CLIP's notion of salience rather than a universal one."],"forward_implications":["Because the strokes are defined as 3D Bezier curves, the abstraction is view-consistent by construction and avoids the flickering that plagues per-view occluding contours.","The level of abstraction is controlled simply by the number of curves; adding strokes automatically adds finer detail, as shown in the paper's Fig. 6.","Users can add or refine detail interactively by selecting keypoints, with refinement completing in roughly a minute.","The curves' adherence to the surface lets them act as deformation handles; a user study found WIR3D-based deformations preferable to ARAP-based deformation 80% of the time.","The method needs no clean input mesh and produces meaningful abstractions even from photo-reconstructed models with boundary and non-manifold edges."],"supporting_citations":[{"why":"The 3Doodle baseline that WIR3D extends; it supplies the 3D curve optimization setting and the theorem equating projected 3D cubic Bezier curves with 2D rational Beziers.","marker":"[7]"},{"why":"CLIPasso supplies the base semantic loss comparing CLIP global and intermediate activations, which WIR3D adapts into its localized loss.","marker":"[61]"},{"why":"CLIP is the pretrained foundation model whose intermediate activations provide the visual-semantic supervision for both optimization stages.","marker":"[48]"},{"why":"DiffVG is the differentiable rasterizer that renders the projected Bezier curves into images for loss computation.","marker":"[28]"},{"why":"Backto3D provides the 2D-to-3D feature back-projection method used for automatic keypoint detection when the user does not supply keypoints.","marker":"[66]"},{"why":"Cited as the basis for the spatial-correlation assumption between intermediate activations and input image locations, on which the localized keypoint loss rests.","marker":"[51]"},{"why":"LPIPS provides the perceptual loss term that is sensitive to geometric layout, used both inside the localized loss and as an evaluation metric.","marker":"[70]"},{"why":"NEF is one of the two main baselines; it represents the implicit edge-field approach for 3D curve reconstruction that WIR3D compares against.","marker":"[69]"}],"fun_headline_variants":["CLIP-guided curves sketch 3D shapes from any view","Sparse curves that hold shape and texture with CLIP","Geometry-aware 3D curves via CLIP and keypoint focus","WIR3D: Sparse CLIP strokes double as deformation handles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The texture stage and feature-control application rest on the assumption that the internal feature maps of CLIP are spatially correlated with the input image, so that a weight map built around projected keypoints actually concentrates the optimization on the intended feature; if CLIP's activations do not localize that way, the localized loss collapses to the unweighted loss and keypoint control stops working.","fun_headline_variants_meta":{"raw":{"variants":["CLIP-guided curves sketch 3D shapes from any view","Sparse curves that hold shape and texture with CLIP","Geometry-aware 3D curves via CLIP and keypoint focus","WIR3D: Sparse CLIP strokes double as deformation handles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000384,"raw_usage":{"total_tokens":2029,"prompt_tokens":938,"completion_tokens":1091,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":1018}},"tokens_in":554,"tokens_out":1091,"duration_ms":8479,"temperature":1.0,"reasoning_tokens":1018,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:20:35.782784+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the stage-II optimization on the same shape twice, once with keypoints on a salient feature and once with the same number of keypoints on a flat, featureless region, keeping all other settings fixed; if the two abstractions are indistinguishable by user ranking or by L2 distance between curve sets, then the localized loss is not doing spatial work. A stronger quantitative version is to take ground-truth keypoint positions, project them into views, and check whether the argmax of the CLIP intermediate activation difference, or the recovered center of the weight map, falls within a few pixels of the projected keypoint; chance-level localization would falsify the paper's central mechanism.","supporting_citations":[{"cited_title":"3Doodle: Compact Abstraction of Objects with 3D Strokes.ACM Transactions on Graphics (TOG), 43(4):1–13, 2024","cited_arxiv_id":null,"evidence_quote":"The 3Doodle baseline that WIR3D extends; it supplies the 3D curve optimization setting and the theorem equating projected 3D cubic Bezier curves with 2D rational Beziers."},{"cited_title":"CLIPasso: Semantically-Aware Object Sketching","cited_arxiv_id":null,"evidence_quote":"CLIPasso supplies the base semantic loss comparing CLIP global and intermediate activations, which WIR3D adapts into its localized loss."},{"cited_title":"Learning Transferable Visual Models from Nat- ural Language supervision","cited_arxiv_id":null,"evidence_quote":"CLIP is the pretrained foundation model whose intermediate activations provide the visual-semantic supervision for both optimization stages."},{"cited_title":"Differentiable vector graph- ics rasterization for editing and learning.ACM Trans","cited_arxiv_id":null,"evidence_quote":"DiffVG is the differentiable rasterizer that renders the projected Bezier curves into images for loss computation."},{"cited_title":"Back to 3d: Few-shot 3d keypoint detection with back-projected 2d features","cited_arxiv_id":null,"evidence_quote":"Backto3D provides the 2D-to-3D feature back-projection method used for automatic keypoint detection when the user does not supply keypoints."},{"cited_title":"Spatial correlation and value prediction in convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Cited as the basis for the spatial-correlation assumption between intermediate activations and input image locations, on which the localized keypoint loss rests."},{"cited_title":"The unreasonable ef- fectiveness of deep features as a perceptual metric","cited_arxiv_id":null,"evidence_quote":"LPIPS provides the perceptual loss term that is sensitive to geometric layout, used both inside the localized loss and as an evaluation metric."},{"cited_title":"NEF: Neural Edge Fields for 3D Parametric Curve Reconstruction from Multi-view Images","cited_arxiv_id":null,"evidence_quote":"NEF is one of the two main baselines; it represents the implicit edge-field approach for 3D curve reconstruction that WIR3D compares against."}],"review_version":1}