{"id":"23fbacfb-13ff-422e-84a0-1999a9bbf9ef","arxiv_id":"2507.00429","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"DiGA3D performs text-guided 3D inpainting (removal, re-texturing, replacement) with a coarse-to-fine diffusion propagation scheme to improve multi-view appearance and geometry consistency.","lead":"This paper presents DiGA3D, a pipeline that removes, repaints, or replaces objects in 3D scenes from a text prompt, by propagating appearance and geometry across camera views with diffusion models. It targets a core problem in 3D editing: keeping the edited region consistent when viewed from new angles.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The geometric-consistency claim rests on TG-SDS warping of monocular depth (Eq. 5) that is never validated against metric or multi-view-consistent depth; Table 3 shows only 0.05 PSNR gain, so the central geometry improvement is not quantitatively supported.","rationale":"The reader identified the same weakest assumption: the accuracy and consistency of monocular-depth-based warping for TG-SDS. I agree that this is the single most load-bearing concern because the paper's claimed novelty over prior work is the explicit geometry guidance, yet the only quantitative evidence (Table 3) shows a negligible appearance gain and no geometric metric is reported. The concern is not about disagreement with the field's consensus; it is about internal support for the central claim. The proposed test is chosen to directly measure the contribution of TG-SDS to geometric accuracy and to expose whether the monocular depth prior is the bottleneck. A favorable outcome (TG-SDS improves geometry and ground-truth depth does not change results much) would strengthen the paper; an unfavorable outcome would indicate the geometry claim is not established. I do not propose changing the CONDITIONAL verdict because the paper may still have merit, but the missing validation is exactly what the condition should require. No ad hominem is intended; the critique is purely on the evidence chain.","tokens_in":14954,"tokens_out":4874,"duration_ms":55539,"concrete_test":"Run DiGA3D on a synthetic multi-view scene with known ground-truth geometry (e.g., a Blender scene rendered with depth maps). Compute a geometric metric (e.g., Chamfer distance or depth L1 error between the optimized 3D Gaussians and the ground-truth point cloud) for three configurations: full model, w/o TG-SDS, and full model with the monocular depth in Eq. 5 replaced by ground-truth depth. If the full model does not improve the geometric metric over w/o TG-SDS, or if substituting ground-truth depth substantially improves it, the geometric-consistency claim is unsupported and the TG-SDS gain is an artifact of appearance or optimization interactions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that DiGA3D achieves multi-view-consistent and geometrically plausible 3D inpainting, with the fine-stage TG-SDS loss (Eq. 6) as the explicit geometry regularizer. This loss conditions ControlNet on texture maps C'_i and depth maps D'_i obtained by DIBR warping (Eq. 5) from reference views using monocular depth estimates. The assumption is that these warped conditions are consistent enough across views to improve the underlying 3D geometry. This assumption is untested in three ways: (i) the warped depth maps are not compared against rendered depth of the optimized scene or any ground-truth geometry; (ii) the only quantitative ablation of TG-SDS (Table 3) measures appearance metrics (PSNR/SSIM/LPIPS) and shows a marginal 0.05 PSNR improvement, while the geometric benefit is shown only qualitatively in Fig. 9 via point-cloud visualization; (iii) the monocular depth estimator itself is used both in the coarse-stage depth loss (Eq. 8) and in the fine-stage warping, so the geometry supervision is circular: any systematic bias or cross-view inconsistency in the depth estimator is propagated and potentially reinforced, not corrected. The paper acknowledges in the appendix that K-means reference selection matters mainly for the fine-stage warping, but does not provide any metric to validate warping accuracy. Without a quantitative geometry metric, the claimed 'geometrically consistent' results could stem from appearance improvements or the SDS prior rather than from the geometry-specific mechanism, or could fail when monocular depth is unreliable. The load-bearing nature of this concern is high because TG-SDS is the paper's main contribution over prior multi-view inpainting work.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DiGA3D, a coarse-to-fine text-guided 3D inpainting pipeline built on 3D Gaussian Splatting. In the coarse stage, it selects multiple reference views via K-means clustering over camera centers, applies DDIM inversion to obtain deterministic latents, and uses an Attention Feature Propagation (AFP) module to transfer attention features from reference views to other views for multi-view appearance consistency. In the fine stage, it introduces a Texture-Geometry Score Distillation Sampling (TG-SDS) loss that conditions a ControlNet on DIBR-warped texture and depth maps from reference views, aiming to improve geometric consistency. The method is evaluated on object removal, re-texturing, and replacement against several baselines, with quantitative tables and qualitative comparisons.","tokens_in":15339,"tokens_out":4892,"duration_ms":51512,"significance":"If the claims are validated, DiGA3D would be a useful unified framework for versatile 3D inpainting, combining a coherent coarse-to-fine design with explicit texture and geometry conditioning. The paper includes extensive qualitative comparisons, ablations of key components, a user study for re-texturing, and additional no-reference metrics, which are valuable. The main contributions, AFP and TG-SDS, are clearly described and the overall idea is plausible. However, the empirical support for the central geometric-consistency claim is currently weak: the quantitative ablation of TG-SDS is small, no error bars are reported, and the evaluation of re-texturing/replacement relies on a metric that is aligned with the text-guided objective being optimized. The framework is worth further development, but the current evidence does not fully support the strength of the abstract's claims.","major_comments":[{"comment":"The central claim that TG-SDS improves geometric consistency is not quantitatively supported. Table 3 shows that removing TG-SDS changes PSNR from 20.71 to 20.66 (+0.05), SSIM from 0.58 to 0.57, and LPIPS from 0.28 to 0.29, with no error bars or significance test. The qualitative point-cloud comparison in Fig. 9 is suggestive, but the paper needs a quantitative geometry metric (e.g., depth error against multi-view stereo, Chamfer distance to a reconstructed surface, or multi-view depth consistency) to substantiate the 'geometrically consistent' claim. Without such a metric, the improvement could be driven by appearance optimization rather than geometry.","section":"§3.4, Table 3, Fig. 9"},{"comment":"The geometry supervision is circular with respect to the monocular depth estimator. The same estimator [31] produces the depth maps used in the coarse-stage depth loss (Eq. 8, with only scale/shift alignment) and the warped depth maps D'_i that condition ControlNet in TG-SDS (Eq. 5). If the monocular estimates have systematic bias or cross-view inconsistency, the warped conditions will reinforce that error. The paper does not validate warping accuracy (e.g., reprojection error, consistency of D'_i with rendered depth, or against metric ground truth). I request a direct quantitative check of DIBR warping quality and an ablation using multi-view-consistent depth if available.","section":"§3.4, Eq. (5), Eq. (8)"},{"comment":"The CLIP directional similarity metric used for re-texturing and replacement is aligned with the text-guided diffusion objective that DiGA3D optimizes, so the large margins in Table 2 (0.1751 vs 0.0742/0.0702) may partly measure objective alignment rather than perceptual quality. The table reports no per-scene variance or error bars, and the user study covers only re-texturing (18 participants, 108 votes). Please add per-scene CLIPdir scores, background-preservation metrics, multi-view consistency metrics, and a user study for replacement, or soften the claim that the method 'aligns more closely with text prompts'.","section":"§4.1, Table 2"},{"comment":"All quantitative tables report single-run means without error bars or significance tests. The differences against the strongest baseline in Table 1 are small (PSNR 20.71 vs 20.49; LPIPS tie; masked LPIPS 0.56 vs 0.54, i.e., worse), and Table 4 evaluates K on a single scene with near-identical PSNR values (19.87 vs 19.94). Reporting multiple seeds or bootstrapped confidence intervals is necessary to establish that the observed improvements are not noise.","section":"§4.4, Tables 1-4"}],"minor_comments":[{"comment":"Learning rates, number of optimization iterations, loss weights λrgb, λdepth, λTG-SDS, and timestep ranges are not given; without these the experiments are difficult to reproduce.","section":"§4.1 Implementation Details"},{"comment":"The index i is reused for the current view and for the summation over reference views; using j for the sum would improve clarity.","section":"§3.3, Eq. (4)"},{"comment":"The masking of the TG-SDS gradient ('only backpropagate the gradient for the masked pixels') is stated only textually; please specify how the mask is applied in latent space and whether the mask is eroded or dilated.","section":"§3.4"},{"comment":"The compared regions are small and the prompts differ across methods; zoomed crops of the edited region would help the reader assess the claimed consistency.","section":"Figures 5-7"},{"comment":"The placeholder 'HERE' on the project page should be replaced, and the SPIn-NeRF dataset description should clarify how the 60/40 train/test split is used for re-texturing and replacement tasks, which lack ground truth.","section":"§2, §4"}],"recommendation":"major_revision","confidential_remarks":"The paper describes a coherent and plausible framework, and the qualitative results are competitive. The main weakness is that the load-bearing claim of geometric consistency via TG-SDS rests on a tiny quantitative ablation and a qualitative point-cloud figure, while the re-texturing/replacement evaluation uses a metric aligned with the method's objective. I would encourage the authors to add a dedicated geometry evaluation and uncertainty estimates before acceptance; these are fixable within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things worth knowing. The paper is a competent integration of existing ideas—multi-reference attention feature propagation from video editing plus a ControlNet-conditioned SDS loss using warped depth/texture—applied to unified 3D inpainting (removal, re-texturing, replacement). That combination is new, and the qualitative results look plausible. But the evidence for the central geometry claim is thin: TG-SDS, the explicit geometry regularizer, moves PSNR by 0.05 on the removal task, and its geometric benefit is shown only as point-cloud pictures, not numbers. The large CLIP dir gains in Table 2 are partly circular, because the SDS objective and the metric both measure text-image alignment. Those two issues are real and worth flagging.\n\nWhat the paper does well: the AFP adaptation is clean, the K-means reference selection is a sensible, low-cost fix for the single-reference problem, and the ablation does show that AFP gives a meaningful gain (0.21 PSNR) over the 2D-inpainter baseline. The writing is honest about failure cases (the Janus problem, a coffee-cup replacement failure). The math is standard and the method is coherent.\n\nThe soft spots are mostly on the empirical side. On object removal, DiGA3D beats GScream by 0.22 PSNR and ties on SSIM and LPIPS, with a worse masked LPIPS. No error bars anywhere, so we cannot tell if 20.71 vs 20.49 is noise. The re-texturing/replacement evaluation uses CLIP dir, which measures the same text-image alignment that the SDS loss optimizes, so the large margins (0.175 vs 0.074) are not independent evidence. The user study is 18 participants, 108 votes, and is reported without analysis. On geometry: the warping uses monocular depth (DPT) that is not metric and is not validated for cross-view consistency; the same depth estimator also feeds the coarse depth loss, so any systematic bias is propagated, not corrected. The stress-test note is right that this is the load-bearing but unsubstantiated part. No code release makes it harder to verify details.\n\nNone of this is fatal—the method is plausible and the integration is genuinely new—but the abstract's \"extensive experiments demonstrate effectiveness\" overstates what the numbers support.\n\nWho it's for: people working on 3D scene editing with Gaussian Splatting who want a training-free unified baseline. It deserves a serious referee, but I'd advise the authors be asked for error bars, an independent geometry metric (e.g., depth consistency or chamfer distance against multi-view-stereo depth), and a TG-SDS ablation that does not use a self-aligned metric. With those, the paper could be solid.","headline":"Plausible integration of known components for unified 3D inpainting, but the geometry claim is under-supported and the CLIP dir numbers are partly self-aligned.","tokens_in":15922,"tokens_out":3567,"would_cite":true,"duration_ms":39754,"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":"DiGA3D claims one text-guided pipeline can remove, re-texture, or replace objects in a 3D scene while keeping appearance and geometry consistent across all views.","keywords":["3D inpainting","3D Gaussian Splatting","diffusion models","multi-view consistency","score distillation sampling","attention feature propagation","object removal","text-guided editing"],"falsifier":"Run DiGA3D's object replacement on a scene where monocular depth is known to disagree across nearby views (for example, textureless or repetitive regions), with and without the TG-SDS loss, and measure the edited region against ground-truth geometry such as a laser scan or multi-view-stereo depth, using Chamfer distance or depth error. If the TG-SDS variant does not reduce geometric error, the claimed geometry regularization is falsified; a simpler color-only check on the paper's own numbers would not settle it, since the reported masked-PSNR difference is only 0.05.","tokens_in":14728,"feed_emoji":"🎨","tokens_out":8336,"duration_ms":84252,"temperature":0.7,"pith_summary":"DiGA3D proposes a single text-guided pipeline for three 3D inpainting tasks—object removal, re-texturing, and replacement—built on 3D Gaussian Splatting. The paper argues that multi-view consistency can be achieved by first propagating 2D diffusion attention features from a few reference views to all other views, then refining the scene with a score distillation loss conditioned on warped texture and depth maps. This coarse-to-fine combination is meant to solve two failure modes of earlier methods: appearance drift for views far from a single reference, and geometry artifacts when the inpainted region changes shape substantially. A reader should care because the target is a practical unified editing tool for 3D scenes rather than a separate method for each task.","feed_headline":"One pipeline removes, re-textures and replaces objects in 3D","feed_subtitle":"Diffusion attention sharing plus geometry-guided SDS keeps multi-view edits consistent","key_machinery":"Two mechanisms carry the argument. Attention Feature Propagation (AFP) uses DDIM-inverted latents from the chosen reference views, then replaces part of each current view's self-attention output with cross-attention against the reference keys and values, so all views share appearance information through the diffusion U-Net. Texture-Geometry SDS (TG-SDS) uses depth-image-based rendering to warp each reference view's image and monocular depth into a target view within the same cluster, converts the warps into Canny texture maps and estimated depth maps, feeds them into ControlNet as conditions, and applies the SDS gradient to the rendered masked pixels. A monocular depth loss with least-squares scale-and-shift alignment anchors geometry during the coarse stage.","core_discovery":"On the paper's own terms, the central discovery is that multi-view 3D inpainting can be decomposed into an appearance-propagation stage and a geometry/texture-regularization stage without retraining the underlying 2D diffusion models. In the coarse stage, K-means clustering on camera centers selects several reference views; DDIM inversion turns each rendered view into a deterministic latent; and an Attention Feature Propagation module mixes reference-view self-attention features into every other view's attention while injecting a CLIP image embedding into the U-Net residual blocks. In the fine stage, a Texture-Geometry SDS loss feeds warped Canny texture maps and monocular depth maps into a multi-condition ControlNet and backpropagates gradients only on masked pixels. The paper reports that this yields improved PSNR and masked PSNR on the SPIn-NeRF object-removal benchmark, higher CLIP directional similarity for re-texturing and replacement, and cleaner point clouds in ablations.","pith_inferences":["A direct testable extension would replace monocular depth with metric depth from multi-view stereo or LiDAR; if TG-SDS's geometry gain is real, metric depth should enlarge it, and if the gain is mostly an artifact of scale/shift alignment, it should shrink.","The AFP attention-sharing schedule could transfer to video inpainting or novel-view synthesis, where temporal or cross-view consistency is the same bottleneck.","Because the paper's ablation shows TG-SDS adds only 0.05 PSNR on removal while the point-cloud improvement is qualitative, a geometric metric such as Chamfer distance against a scanned mesh would settle the claimed geometry gain.","The K-means reference selection is motivated by warping stability in the fine stage, which suggests consistency may degrade for sparse camera trajectories or large clusters; stress-testing those regimes would reveal the method's actual coverage limits."],"forward_implications":["If the central claim holds, one pipeline can serve object removal, re-texturing, and replacement, so users no longer need task-specific 3D inpainting systems.","Appearance consistency across views can be obtained without a joint 3D latent: sharing attention features during 2D inpainting is enough for a coarse consistent result.","Explicit warped texture and depth conditions are sufficient to regularize 3D geometry, reducing multi-facet artifacts even when monocular depth is only known up to scale and shift.","The reported gain in masked PSNR (17.22 versus 15.84 for the closest removal baseline) indicates the main improvement is inside the inpainted region, not just the unchanged background.","Training-free pretrained diffusion models can power versatile 3D inpainting; no fine-tuning of the 2D inpainter is required."],"supporting_citations":[{"why":"Provides the text-prompt-based 2D inpainting backbone that hosts DDIM inversion and AFP in the coarse stage.","marker":"[48]"},{"why":"Supplies the multi-condition ControlNet that consumes warped texture and depth maps in the TG-SDS loss.","marker":"[43]"},{"why":"DDIM inversion generates the deterministic latents whose attention features AFP propagates across views.","marker":"[34]"},{"why":"Monocular depth estimates drive the depth loss, the DIBR warping, and the warped depth conditions.","marker":"[31]"},{"why":"Depth-image-based rendering performs the reference-to-target warping of Eq. (5).","marker":"[10]"},{"why":"Score Distillation Sampling is the base loss that TG-SDS extends with texture and geometry conditions.","marker":"[28]"},{"why":"Supplies the SPIn-NeRF dataset and the object-removal baselines used for quantitative comparison.","marker":"[26]"},{"why":"3D Gaussian Splatting is the differentiable representation optimized by the photometric, depth, and TG-SDS losses.","marker":"[14]"},{"why":"Lang SAM converts mask prompts into the 2D masks used for inpainting and masked gradient backpropagation.","marker":"[15]"}],"fun_headline_variants":["Diffusion propagation unifies 3D remove, re-texture, swap","Attention sharing keeps multi-view 3D inpaints consistent","Coarse-to-fine diffusion for geometry-aware 3D editing","One diffusion pipeline for versatile 3D inpainting","Multi-view 3D inpainting via attention propagation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that monocular depth maps, after least-squares scale-and-shift alignment, agree well enough across views that DIBR warping produces trustworthy geometry conditions; if the depth estimator is inconsistent across views, the TG-SDS loss would reinforce the wrong shape.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion propagation unifies 3D remove, re-texture, swap","Attention sharing keeps multi-view 3D inpaints consistent","Coarse-to-fine diffusion for geometry-aware 3D editing","One diffusion pipeline for versatile 3D inpainting","Multi-view 3D inpainting via attention propagation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001306,"raw_usage":{"total_tokens":5356,"prompt_tokens":1007,"completion_tokens":4349,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":4262}},"tokens_in":623,"tokens_out":4349,"duration_ms":35172,"temperature":1.0,"reasoning_tokens":4262,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:15:46.592307+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DiGA3D's object replacement on a scene where monocular depth is known to disagree across nearby views (for example, textureless or repetitive regions), with and without the TG-SDS loss, and measure the edited region against ground-truth geometry such as a laser scan or multi-view-stereo depth, using Chamfer distance or depth error. If the TG-SDS variant does not reduce geometric error, the claimed geometry regularization is falsified; a simpler color-only check on the paper's own numbers would not settle it, since the reported masked-PSNR difference is only 0.05.","supporting_citations":[{"cited_title":"Adding conditional control to text-to-image diffusion models","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-condition ControlNet that consumes warped texture and depth maps in the TG-SDS loss."},{"cited_title":"Vi- sion transformers for dense prediction","cited_arxiv_id":null,"evidence_quote":"Monocular depth estimates drive the depth loss, the DIBR warping, and the warped depth conditions."},{"cited_title":"Depth-image-based rendering (dibr), com- pression, and transmission for a new approach on 3d-tv","cited_arxiv_id":null,"evidence_quote":"Depth-image-based rendering performs the reference-to-target warping of Eq. (5)."},{"cited_title":"Segment any- thing","cited_arxiv_id":null,"evidence_quote":"Lang SAM converts mask prompts into the 2D masks used for inpainting and masked gradient backpropagation."}],"review_version":1}