{"id":"6973e288-6f3e-4ec1-9e84-e60501577d30","arxiv_id":"2412.13552","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DragScene propagates a single-view drag edit across multiple views by reconstructing a coarse point cloud with the edit's latent features, then reconstructing the edited 3D scene.","lead":"DragScene is a new method that lets users edit a 3D scene by dragging points on one photo of the scene, while keeping the edit consistent across all other viewpoints. It combines an existing 2D drag editor, a 3D point-cloud reconstruction step, and Gaussian Splatting to rebuild the edited scene.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Multi-view consistency depends on DUSt3R point clouds reconstructed from out-of-distribution drag-edited images; the paper never quantifies this failure mode, and its only quantitative metric (FID to the reference view) does not measure cross-view consistency.","rationale":"The reader's weakest assumption identifies exactly the DUSt3R reliability issue in the edited region; I agree with that identification. The reader's verdict of CONDITIONAL is appropriate because the mechanism could work, but the paper does not demonstrate it quantitatively. My stress-test sharpens the concern in one respect: the paper's only quantitative metric, FID between novel views and the edited reference view, is a poor proxy for multi-view consistency, since it can reward novel views that simply copy the reference appearance rather than reflecting a coherent 3D edit. The proposed test would settle the matter by measuring cross-view agreement against ground truth. I considered whether a more fundamental issue exists, such as the latent-space rendering R_{v_i}(P_Z) being ill-defined, but that is secondary: even a perfect renderer inherits the point cloud's errors. The paper's independent support is limited to qualitative figures and a single qualitative ablation; no code or data are released. These issues are addressable, so UNCHANGED (conditional) remains the right verdict if the authors commit to the proposed test; if they cannot provide such evidence, the claim should be downgraded to unverified.","tokens_in":11481,"tokens_out":3759,"duration_ms":35926,"concrete_test":"On a synthetic scene with ground-truth geometry (e.g., a 3DGS sofa), render a reference view and apply a known drag that deforms the sofa (scale height by 1.2). Render all views of the true deformed scene as ground truth. Run DragScene on the reference view only. Then (1) compute average LPIPS between DragScene's novel-view outputs and the ground-truth deformed novel views; (2) reconstruct the edited 3DGS from DragScene outputs and compute Chamfer distance to the ground-truth deformed mesh in the masked region; (3) directly probe DUSt3R by feeding the edited reference image plus original views and comparing the masked-region pointmap depth to the known deformed depth. If LPIPS > 0.15 or Chamfer distance exceeds the original object scale by >5%, the DUSt3R assumption fails and the consistency claim collapses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that DragScene propagates single-view drag edits across views while maintaining multi-view consistency—rests on the point-cloud reconstruction in Section 3.2. Equation (3) aligns the masked region of every novel view to the edited reference's pointmap X^{0,m}_i, but that pointmap is itself produced by DUSt3R from a drag-edited image, which is far outside the natural-image distribution DUSt3R was trained on. If DUSt3R predicts incorrect geometry in the edited region (e.g., the 'make the sofa taller' edit should produce a taller sofa, but the pointmap has no 3D evidence for the extrapolated shape), then R_{v_i}(P_Z) renders wrong latent targets and masks. Subsequent views are pulled toward latents that do not correspond to any consistent 3D surface, and the final 3DGS reconstruction exhibits ghosting or floaters. The paper's only quantitative evaluation (Fig. 9) computes FID between novel-view outputs and the edited reference image; this measures deviation from the editing target, not multi-view consistency, and is computed on a single scene with no error bars. The ablation in Fig. 7 shows one qualitative example; no failure-rate statistics are given, and the Limitation section concedes failures for large view ranges without delimiting them. Thus the paper's central claim is logically plausible but empirically unsecured.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DragScene is a method for editing a reconstructed 3D scene by dragging points on a single reference view. Given a reference image rendered from the scene, a user-specified mask and handle/target point pairs, the authors first apply DragDiffusion to obtain a 2D edit and a DDIM-inverted latent for the edited view. They then reconstruct a coarse point cloud with DUSt3R from the edited reference and several original views, using the regression loss in Eq. (3) to keep unedited regions aligned with the original views while forcing edited regions to agree with the reference edit. The edited view's latent is attached to the point cloud, and rendered latent maps and masks are used to optimize DDIM-inverted latents of other views via Eqs. (4)-(7). Finally, the edited images are used to reconstruct a new 3D Gaussian Splatting scene. The paper claims this is the first drag-style 3D editing method applicable to real-world scenes and that it maintains multi-view consistency.","tokens_in":11735,"tokens_out":6679,"duration_ms":61647,"significance":"The problem addressed is real: existing text-driven 3D editing lacks localized geometric control, and drag-style interaction is a natural alternative. DragScene's design is elegant and representation-agnostic in principle; using DUSt3R pointmaps as 3D clues to transfer a latent edit across views is a sensible way to avoid per-scene optimization. The ablations in Figs. 7-9 and the comparisons in Fig. 6 show plausible qualitative results. If the multi-view consistency claim can be substantiated quantitatively, this would be a useful practical contribution to 3D scene editing. However, as presented, the evidence for the central consistency claim is mostly qualitative, with only a single-scene, target-deviation FID and NIQE curve, so the contribution is not yet established to the standard of a top venue. The paper also does not release code or data, which limits reproducibility.","major_comments":[{"comment":"The central claim that DragScene achieves multi-view consistency is supported only by qualitative examples. There is no user study and no quantitative cross-view consistency metric. The only quantitative evaluation, Fig. 9, reports FID between novel-view edited images and the edited reference image; this measures deviation from the target view, not consistency among views, and the text does not state the number of scenes or edits over which the scores are averaged. Please report quantitative consistency (e.g., re-projection error on the reconstructed point cloud, pairwise LPIPS/PSNR between matched views, or a user study) across multiple scenes and edits.","section":"Section 4.3 and Fig. 6"},{"comment":"The coarse point cloud that carries the edit is reconstructed from an edited reference image that is out of distribution for DUSt3R. If the masked-region pointmap is wrong (e.g., for the 'make the sofa taller' edit), the rendered latent targets in Section 3.3 will pull other views toward inconsistent geometry. The only evidence that this does not happen is the single qualitative ablation in Fig. 7; the paper itself concedes failures for large view ranges but does not delimit them. Please provide a quantitative evaluation of the edited-region geometry, for example by comparing the reconstructed point cloud with a ground-truth edited mesh in a synthetic scene or by measuring re-projection consistency of the edited region across held-out views.","section":"Section 3.2, Eq. (3), and Limitation"},{"comment":"The mask M is a 2D user annotation, but Eq. (2) transforms it with the same camera-pose mapping used for 3D pointmaps, h(M) being the homogeneous embedding of the mask. The depth or 3D surface used to lift the 2D mask into 3D is not specified. Since this mask is used in Eq. (3) and later rendered as R_vi(P_M) in Eqs. (4)-(6), the ambiguity directly affects the method's reproducibility. Please state explicitly how the 2D mask is associated with the reconstructed point cloud and how Eq. (2) is evaluated.","section":"Section 3.2, Eq. (2)"},{"comment":"The comparison with GaussianEditor and DragDiffusion-3D is entirely qualitative (Fig. 6). Given that the paper claims superiority in both editing quality and controllability, a quantitative comparison (e.g., user preference rates, or a protocol running the same set of edit intents across multiple scenes) is needed to rule out selection effects in the displayed examples.","section":"Section 4.2"}],"minor_comments":[{"comment":"The sentence 'As shown in Fig. 7, the optimal results are achieved when eta is around 0.4' appears to reference the wrong figure; Fig. 7 is the point-cloud reconstruction ablation, while Fig. 9 contains the NIQE and FID plots.","section":"Section 4.5"},{"comment":"The implementation text says 'using the AdamW optimizer[11]' but reference [11] is the original Adam paper; please cite AdamW or correct the reference.","section":"Section 4.1"},{"comment":"The model name is written inconsistently as 'DUSt3R' in the main text and 'Dust3R' in Section 4.4; please use one spelling throughout.","section":"Section 3.2"},{"comment":"The description of LoRA fine-tuning does not specify which images are used for the 80 fine-tuning iterations; please clarify whether only the reference image or a set of views is used.","section":"Section 4.1"},{"comment":"The regression loss is described as an optimization over pointmaps, but the text does not state which variables are updated or how the optimization is initialized and run; please add a sentence describing the optimization procedure.","section":"Section 3.2, Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":"The main reservation is the validation burden, not the technical soundness of the core idea. The authors should be pushed to supply quantitative multi-view consistency evidence and a user study, and to clarify the mask lifting in Eq. (2). The paper would also be strengthened by a clear statement on code/data release, since the central mechanism is currently underspecified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version. DragScene is the first thing I've seen that takes drag-style editing to real-world 3D scenes rather than single objects, and the core trick—reconstruct a coarse point cloud from the edited reference plus original views, assign the edited latent to that cloud, render latent maps to supervise other views—is a real idea worth thinking about. The losses are concrete, and the ablation showing that omitting either the edited image or the original images breaks the result tells you the mechanism is doing something.\n\nThe soft spots are all in the evidence. The evaluation is essentially qualitative: a handful of scenes, comparison by eye, no user study. The only numbers in the paper are Figure 9, and they don't measure what the paper claims. FID between novel views and the edited reference view measures deviation from the target, not multi-view consistency. On a single scene with no error bars, that's not support for the central claim. Also, the stress-test concern holds up: the mechanism depends on DUSt3R producing reliable geometry in the masked region from a drag-edited image, and that's out of distribution. The limitation section admits failures at large view range but gives no scope or rate. No code or data, and some hyperparameters (sigma, m, lambda) are left unspecified.\n\nNone of that kills the core idea. The failure mode is real, but it's an empirical question, not a logical one. The paper is worth a serious referee, because the method is novel, well-motivated, and clearly described, and the identified weaknesses are addressable with a stronger evaluation: user study, a real cross-view consistency metric, failure-rate counts, and code/data release. I'd send it to review, expecting heavy revision. I wouldn't cite it in its current form, but I'd keep an eye on the follow-up.","headline":"A genuinely new pipeline for drag-editing real 3D scenes, with the point-cloud latent propagation as the key idea; the evaluation, however, doesn't yet back the multi-view consistency claim.","tokens_in":12332,"tokens_out":1771,"would_cite":false,"duration_ms":16761,"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":"DragScene claims that a single drag on one rendered view can edit the whole 3D scene consistently.","keywords":["3D scene editing","drag-style editing","multi-view consistency","latent diffusion models","point cloud reconstruction","3D Gaussian Splatting","single-view interaction","view-consistent editing"],"falsifier":"Render a synthetic 3D scene with known ground-truth geometry, apply a drag that changes the shape, such as lengthening an object, reconstruct the point cloud from the edited reference view alone, and compare depth in the masked region to ground truth; if the depth error in the edited region is large or grows with drag distance, the multi-view-consistency mechanism will fail exactly where it is needed.","tokens_in":11240,"feed_emoji":"🖱️","tokens_out":5708,"duration_ms":48357,"temperature":0.7,"pith_summary":"DragScene claims that a user can edit a real-world 3D scene by drawing a mask and dragging a point on a single rendered view, with the edit propagated coherently to every other view. The paper proposes to do this by running a 2D latent-diffusion drag edit on one reference image, reconstructing a coarse point cloud that fuses the edited image with a few original views, and then using latent maps rendered from that point cloud to supervise the drag edit of all other views. If this works, it turns a precise 2D interaction into a general 3D editing tool that is not tied to one 3D representation and does not require text prompts or per-object retraining. The paper reports qualitative and quantitative evidence on real scenes from common 3D editing benchmarks, arguing that earlier deformation-based and prompt-based methods cannot achieve this level of localized control.","feed_headline":"One drag on a single view edits the whole 3D scene","feed_subtitle":"DragScene carries the edit to every other viewpoint through 3D latent maps, so the result stays coherent from all angles.","key_machinery":"The load-bearing object is the 3D latent map: a coarse point cloud reconstructed by DUSt3R from the edited reference view and a small set of original views, with every point carrying the latent code of the edited reference image and the user's mask. Rendered into a new viewpoint, this map supplies both the target latent for the edited region and the mask that protects unchanged regions. Its role is to translate a single 2D edit into view-consistent latent supervision; the paper's ablation shows that omitting either the edited-image supervision or the original-image supervision degrades the point cloud and breaks feature correspondence, which in turn produces mismatched edited views.","core_discovery":"The central discovery is a mechanism for keeping drag edits multi-view consistent. Instead of running a 2D drag editor independently on each view, which produces inconsistent features, DragScene first edits one reference view and obtains its diffusion latent. It then reconstructs a coarse point cloud from the edited reference image plus sparse original views, using a regression loss that keeps unmasked regions aligned with the original scene and masked regions aligned with the edit. The reference latent is assigned to that point cloud, and latent maps plus masks are rendered into all other views, where they drive a latent optimization with a reconstruction term pulling the edited region toward the 3D latent map and a mask term keeping unchanged regions near their original latent. The edited multi-view images are finally used to reconstruct the target 3D scene, demonstrated with 3D Gaussian Splatting, and the paper positions DragScene as the first drag-style 3D editing method applicable to real-world scenes.","pith_inferences":["A direct extension is to replace the chosen 2D drag editor with any newer one; the paper itself notes that editing quality is bounded by the 2D drag model, so the framework's ceiling will track progress in 2D drag editing.","The single-reference-view design implies that very large viewpoint changes or heavy occlusions could break the point-cloud bridge, a failure mode the paper acknowledges as unsatisfactory results for large view ranges.","A testable prediction is that multi-view consistency degrades as edit magnitude grows, because larger edits move the edited image further from the reconstruction model's training distribution; measuring masked-region reprojection error as a function of drag distance would expose this.","The latent-bridge mechanism does not inherently depend on scene staticity, so extending it to dynamic scenes or to text-guided drags, both listed as future work in the paper, is a plausible path."],"forward_implications":["If DragScene is correct, users can edit captured real scenes, including faces, furniture, and outdoor scenery, by dragging in one view without writing text prompts or training per object.","Because the method operates on rendered images and a point-cloud latent bridge, it can be retargeted to newer 3D representations beyond 3D Gaussian Splatting whenever multi-view images can be rendered and reconstructed.","The framework converts a capable 2D drag editor into a 3D editor, so future improvements in 2D drag editing would directly improve the quality and variety of 3D edits.","Multi-view consistency is enforced at the latent level before final reconstruction, which should reduce the artifacts and discontinuities seen when drag editing is applied naively to each view separately.","The approach supports localized and creative edits, such as widening a blade or lengthening a face, which the paper argues are poorly handled by deformation-based and prompt-driven 3D editing baselines."],"supporting_citations":[{"why":"introduces point-based drag editing that DragScene extends from images to 3D scenes.","marker":"[34]"},{"why":"supplies the latent-diffusion drag editing and optimization procedure used on the reference and novel views.","marker":"[38]"},{"why":"reconstructs the coarse point cloud from the edited reference image and original views that becomes the 3D latent map.","marker":"[43]"},{"why":"provides the 3D Gaussian Splatting representation used to render multi-view images and reconstruct the edited scene.","marker":"[23]"},{"why":"provides the DDIM inversion that produces the latent representations optimized during editing.","marker":"[39]"},{"why":"supplies one of the real-world 3D editing benchmark datasets used for experiments.","marker":"[17]"},{"why":"supplies the other real-world benchmark dataset used to evaluate editing on diverse scenes.","marker":"[1]"},{"why":"serves as the prompt-based 3D editing baseline that DragScene is compared against.","marker":"[42]"}],"fun_headline_variants":["Drag one view, edit the entire 3D scene","Single-view drag gives multi-view 3D consistency","Drag a single image to edit a full 3D scene","One drag redefines the whole 3D scene","Drag on one view, see it everywhere in 3D"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that the point-cloud reconstruction of the edited reference view gets the edited region's geometry right, even though the edited image is something the reconstruction model was not trained on; if that geometry is wrong, the latent maps mislead other views and consistency breaks.","fun_headline_variants_meta":{"raw":{"variants":["Drag one view, edit the entire 3D scene","Single-view drag gives multi-view 3D consistency","Drag a single image to edit a full 3D scene","One drag redefines the whole 3D scene","Drag on one view, see it everywhere in 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1328,"prompt_tokens":938,"completion_tokens":390,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":308}},"tokens_in":554,"tokens_out":390,"duration_ms":3925,"temperature":1.0,"reasoning_tokens":308,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:00:45.306028+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a synthetic 3D scene with known ground-truth geometry, apply a drag that changes the shape, such as lengthening an object, reconstruct the point cloud from the edited reference view alone, and compare depth in the masked region to ground truth; if the depth error in the edited region is large or grows with drag distance, the multi-view-consistency mechanism will fail exactly where it is needed.","supporting_citations":[{"cited_title":"Drag your gan: Interactive point-based manipulation on the generative image manifold","cited_arxiv_id":null,"evidence_quote":"introduces point-based drag editing that DragScene extends from images to 3D scenes."},{"cited_title":"Dragdiffusion: Harnessing diffusion models for interactive point-based image editing","cited_arxiv_id":null,"evidence_quote":"supplies the latent-diffusion drag editing and optimization procedure used on the reference and novel views."},{"cited_title":"Dust3r: Geometric 3d vision made easy","cited_arxiv_id":null,"evidence_quote":"reconstructs the coarse point cloud from the edited reference image and original views that becomes the 3D latent map."},{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"provides the 3D Gaussian Splatting representation used to render multi-view images and reconstruct the edited scene."},{"cited_title":"Instruct-nerf2nerf: Editing 3d scenes with instructions","cited_arxiv_id":null,"evidence_quote":"supplies one of the real-world 3D editing benchmark datasets used for experiments."},{"cited_title":"Mip-nerf 360: Unbounded anti-aliased neural radiance fields","cited_arxiv_id":null,"evidence_quote":"supplies the other real-world benchmark dataset used to evaluate editing on diverse scenes."},{"cited_title":"Gaussianeditor: Editing 3d gaussians delicately with text instructions","cited_arxiv_id":null,"evidence_quote":"serves as the prompt-based 3D editing baseline that DragScene is compared against."}],"review_version":1}