{"id":"1cc00c35-2a6f-456e-8f7e-c4f28f09f721","arxiv_id":"2602.18314","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Diff2DGS uses diffusion video inpainting plus 2D Gaussian Splatting to reconstruct occluded deformable surgical scenes, but its geometric superiority claim rests on a circular RAFT-depth evaluation.","lead":"The paper removes surgical instruments from video with a diffusion model, then reconstructs deformable tissue with 2D Gaussian Splatting, reporting better images and geometry on three surgical datasets. The geometry claim is weakened because the same RAFT depth estimator is used both as training supervision and as the evaluation reference.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SCARED geometry evaluation is contaminated: the inpainting module is fine-tuned on SCARED and then evaluated on SCARED with no split described, so the claimed masked-region gains may reflect memorization rather than reconstruction.","rationale":"The reader's verdict of REJECT is well-supported, but their weakest_assumption focused on the circularity of using RAFT depth for both supervision and evaluation. That is a valid concern, but the paper attempts to address it by validating RAFT on SCARED (Table V) and by reporting GT-depth results on SCARED (Table III). The more serious and less defensible flaw is that the SCARED GT evaluation itself is contaminated: the inpainting module is fine-tuned on SCARED, and no split is described. Since SCARED is the only dataset with true 3D ground truth, any leakage there destroys the paper's strongest evidence for the geometry claim. The ablation table (Table IV) only reports PSNR/SSIM on StereoMIS, not geometry, so the adaptive depth loss's geometric benefit is never isolated. These issues together justify rejection: the central claim is not reliably established, and the primary evaluation protocol is compromised. My agreement is 'partial' because I identify a different root cause than the reader, though the conclusions align.","tokens_in":11247,"tokens_out":2974,"duration_ms":28421,"concrete_test":"Require the authors to disclose and enforce a strict sequence-level split: fine-tune the inpainting module on four of the five SCARED sequences and evaluate on the held-out fifth sequence (leave-one-sequence-out, repeated five times). Recompute Table III masked-region PSNR/RMSE and full-image RMSE under this protocol. If the masked-region PSNR drops by more than a few dB relative to the reported 30.53 dB, or if the relative ordering against Deform3DGS/SurgicalGS changes, the claimed geometry advantage is largely attributable to training/test leakage. Also report the same comparison without any SCARED fine-tuning.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim is superiority in both appearance and geometry. The only dataset with true 3D ground truth is SCARED, and Table III reports large masked-region improvements (PSNR 30.53 vs. 21–24 for prior methods). However, Section III-A states: 'The model is fine-tuned on surgical sequences of SCARED dataset.' No train/test split is reported. SCARED contains only five porcine cadaver sequences; if the same sequences are used for fine-tuning and evaluation, the inpainting module has seen the clean, unmasked tissue during training. The SCARED evaluation then measures the model's ability to reconstruct images it has memorized, not its ability to inpaint unseen occluded tissue. This would inflate both the masked-region PSNR/RMSE and the qualitative geometry results, undermining the only non-circular geometry evidence in the paper. The RAFT-based RMSE on EndoNeRF/StereoMIS is a separate circularity concern, but the SCARED contamination is more directly load-bearing because it affects the one dataset with independent depth ground truth.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Diff2DGS, a two-stage pipeline for dynamic surgical scene reconstruction. First, a diffusion-based video inpainting module removes surgical instruments and fills occluded tissue with temporally consistent content. Second, 2D Gaussian Splatting (2DGS) is combined with a Learnable Deformation Model (LDM) to model tissue deformation, and an adaptive depth loss is used to improve geometric fidelity. The method is evaluated on EndoNeRF, StereoMIS, and SCARED. The authors claim state-of-the-art performance in both appearance and geometry, citing 38.02 dB PSNR on EndoNeRF and 34.40 dB on StereoMIS, and they introduce a SCARED-based quantitative depth evaluation protocol.","tokens_in":11595,"tokens_out":5444,"duration_ms":50872,"significance":"If the central claims were fully supported, the paper would be a useful contribution to real-time intraoperative 3D reconstruction: explicitly inpainting occluded tissue before reconstruction, adapting 2DGS to deformable scenes, and pushing geometry evaluation beyond image-quality metrics on a dataset with true depth ground truth. The manuscript also provides code, ablations, and a unified SCARED evaluation protocol. However, as written, the evidence for the geometric superiority claim is invalid: the RMSE tables contradict the claim, and the SCARED evaluation is contaminated by fine-tuning/evaluation overlap. The appearance-level contribution appears sound but needs to be separated from the unsupported geometric claims.","major_comments":[{"comment":"The central claim that Diff2DGS 'outperforms state-of-the-art methods in both appearance and geometry' is contradicted by the paper's own RMSE results. On every scene split in Tables I and II, SurgicalGS has lower RMSE than Diff2DGS: EndoNeRF-Cutting 1.779 vs 2.443, EndoNeRF-Pulling 2.193 vs 2.574, StereoMIS-P27 2.747 vs 3.874, StereoMIS-P3 2.536 vs 3.717. Since RMSE is the only geometry metric reported on these datasets, the paper's geometry superiority claim is not supported by its own data.","section":"Section III-B, Tables I and II"},{"comment":"The depth evaluation on EndoNeRF and StereoMIS is circular. The method uses RAFT stereo depth for depth supervision (the depth loss in Eq. (14) is trained against RAFT estimates) and then reports RMSE against those same RAFT estimates. Table V validates RAFT on SCARED keyframes, not on instrument-occluded or inpainted regions, so accuracy in the masked regions is untested. Thus the RMSE numbers in Tables I and II measure consistency with a training reference, not true anatomical accuracy. Geometry claims on these datasets should be removed or replaced by evaluation against independent ground truth.","section":"Section III-A, Section III-D, Eq. (14)"},{"comment":"The SCARED geometry evaluation is contaminated by training/evaluation overlap. Section III-A states that the inpainting module is 'fine-tuned on surgical sequences of SCARED dataset,' and Table III evaluates on SCARED with synthetic instrument masks, using the original unmasked images as ground truth. No train/test split is reported. Since SCARED contains only five porcine cadaver sequences, the model may have memorized the clean tissue appearance during fine-tuning, which would inflate the masked-region PSNR of 30.53 dB and the RMSE of 8.21 mm. The authors must specify a split that excludes the evaluation sequences from fine-tuning and report per-sequence results.","section":"Section III-A, Table III"},{"comment":"The paper's conclusion and abstract claim 'superior performance' and 'reliable reconstruction' largely on the basis of geometric accuracy. Given that the only geometry evidence with true ground truth (SCARED) is contaminated, and the other geometry evidence is circular, the current manuscript does not substantiate its main selling point. A revised version should either provide a valid held-out SCARED evaluation or substantially weaken the geometry claims.","section":"Section IV and abstract"}],"minor_comments":[{"comment":"The abstract reports 34.40 dB PSNR on StereoMIS, while the introduction reports 33.90 dB. This inconsistency should be corrected.","section":"Section I vs abstract"},{"comment":"The LPIPS value for SurgicalGS on StereoMIS-P3 is listed as 0.925, which is likely a typo (probably 0.0925). Please verify.","section":"Table II"},{"comment":"The TCS discussion is confusing: the 'Inpaint result' row in Table III has TCS 0.003, which is better than Diff2DGS's 0.027, yet the text says Diff2DGS attains 'suboptimal performance—yet remains superior to all other compared methods.' Clarify what baseline is being compared and whether the inpainting result is considered a competing method.","section":"Section III-B, Eq. (19)"},{"comment":"No error bars, confidence intervals, or multiple-run statistics are provided. For a paper claiming state-of-the-art performance, at least a small number of runs with variance should be reported.","section":"Tables I-III"},{"comment":"There is a typo: 'lantent space' should be 'latent space.'","section":"Section II-C, Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The appearance-level contribution and the idea of using SCARED for geometry evaluation are worth pursuing, but the current geometric claims are not supported by the evidence. The paper would need a proper held-out SCARED split, a re-evaluation that avoids circular RAFT-based depth metrics, and a revision of the title/abstract claims before it could be considered for publication. I do not see this as a reject-and-discard case, but the required changes are substantial."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a reasonable engineering paper that tackles a real problem—instrument occlusion in deformable surgical reconstruction—but the central claim of superiority in both appearance and geometry is not supported by the evidence in the manuscript. The appearance numbers are plausible; the geometry numbers are not, and the only independent ground-truth evaluation on SCARED is compromised.\n\nWhat's actually new: combining video diffusion inpainting with 2D Gaussian Splatting in a deformable setting, plus an adaptive depth loss weight. Each component exists, but the combination is new, and the ablation suggests each piece contributes. The observation that PSNR/SSIM don't track geometric accuracy is worth making, and the attempt to bring SCARED depth ground truth into the evaluation is the right direction.\n\nWhere it falls down. First, their own RMSE tables show SurgicalGS with lower RMSE on both EndoNeRF and StereoMIS, so 'outperforms in geometry' is false unless you redefine the metric. Second, the depth evaluation on those two datasets is circular: RAFT depth is used to initialize the point cloud and to supervise the depth loss, then treated as ground truth for RMSE. The Table V validation of RAFT on SCARED keyframes doesn't cover the inpainted, occluded regions where errors would matter most. Third, and most load-bearing, the only non-circular geometry evidence is SCARED, and the inpainting module is fine-tuned on SCARED with no train/test split described. With only five porcine sequences, that could easily inflate the masked-region PSNR and depth numbers. There are smaller issues too: an abstract/conclusion inconsistency (34.40 vs 33.90 dB on StereoMIS), no error bars, and several hyperparameters unreported.\n\nThe right fix is fairly clear: report a clean SCARED split, measure depth in the inpainted regions against the structured-light ground truth rather than RAFT, and be honest about what the RMSE tables actually show. The appearance story may well survive; the geometry story needs real work.\n\nWho this is for: researchers building surgical reconstruction pipelines that include inpainting. It deserves a serious referee—the problem is genuine and the system is nontrivial—but it needs major revision before its claims can be accepted. I wouldn't cite the geometric results as they stand.","headline":"Useful engineering paper, but the geometry claims don't survive contact with their own tables, and the independent SCARED evaluation is contaminated by fine-tuning on the same data.","tokens_in":12053,"tokens_out":2944,"would_cite":false,"duration_ms":26098,"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":"Diff2DGS reconstructs surgical scenes by first inpainting instruments with a diffusion video model, then fitting 2D Gaussians with learned deformations and adaptive depth weighting.","keywords":["surgical scene reconstruction","2D Gaussian splatting","diffusion inpainting","deformable tissue","adaptive depth loss","EndoNeRF","StereoMIS","SCARED"],"falsifier":"Take a surgical video with known ground-truth depth from a structured-light sensor (like SCARED) or CT, artificially occlude tissue with instrument masks, run Diff2DGS, and compare the recovered geometry in the masked regions against the true anatomy. If the depth RMSE in those regions is no better than Deform3DGS or SurgicalGS, or if appearance improves while geometry diverges, the central claim fails.","tokens_in":11166,"feed_emoji":"🏥","tokens_out":3204,"duration_ms":27323,"temperature":0.7,"pith_summary":"The paper claims that reliable 3D reconstruction of deformable surgical scenes becomes tractable if you first remove instruments from the video by temporally consistent diffusion inpainting, then represent the tissue with 2D Gaussian Splatting driven by a lightweight learnable deformation model. It further claims that once occluded regions are inpainted, an adaptive depth loss is needed to prevent appearance-optimized splatting from producing geometrically wrong surfaces. The method reports state-of-the-art appearance metrics on EndoNeRF and StereoMIS and better depth error on SCARED, where real ground truth exists. The deeper message is that image-quality metrics alone can mask poor geometry, so evaluation should include depth accuracy.","feed_headline":"Inpaint tools first: 38 dB PSNR on occluded surgical scenes","feed_subtitle":"Diffusion inpainting removes instruments; 2D Gaussian splatting then models tissue motion and geometry in real time.","key_machinery":"The key mechanism is the combination of a diffusion video inpainting model (built on Stable Diffusion v1.5 with a causal temporal attention mask) that restores instrument-occluded tissue, and a 2D Gaussian Splatting renderer where each planar Gaussian is deformed over time by a Learnable Deformation Model (LDM) that predicts scale, rotation, and position changes through Gaussian basis functions of time. An adaptive depth loss weight, driven by the ratio of RGB to depth loss, steers training toward geometric fidelity. The depth is also used to initialize the point cloud and as a supervision signal.","core_discovery":"The central claim is that occlusion by surgical instruments, the main failure mode of prior deformable Gaussian splatting methods, can be eliminated before reconstruction rather than masked during optimization. Diff2DGS shows that a diffusion-based video inpainting module with temporal attention produces spatiotemporally consistent tissue appearance under surgical tools, and that plugging those inpainted frames into a 2D Gaussian Splatting pipeline — extended with a learnable deformation model and an adaptive depth loss weight — yields reconstructions with higher PSNR/SSIM and lower depth error than prior methods. The paper also demonstrates the corollary that optimizing purely for image qua","pith_inferences":["Inference: The inpainting-before-reconstruction design suggests a general principle — for any transient occlusion (blood, smoke, tools), a temporally consistent 2D inpainting preprocessor can turn a dynamic reconstruction problem into a cleaner static-with-deformation problem.","Inference: Because the adaptive depth loss decouples geometry from appearance, similar mechanisms could improve other medical or robotic domains where appearance objectives dominate training.","Inference: The reliance on RAFT depth as both initialization and evaluation creates a coupling; a testable implication is that on a dataset with independent ground truth (CT or structured light), the reported depth advantage could shrink or grow depending on whether RAFT errors correlate with the masked regions."],"forward_implications":["If correct, instrument-occluded regions would reconstruct with fewer artifacts, removing the need for per-pixel masking during optimization.","Depth estimates would be closer to true anatomy, making intraoperative navigation and autonomous assistance more reliable.","The finding that appearance metrics can hide geometric error would push future surgical scene benchmarks to include depth ground truth.","The method's rendering speed (hundreds of times faster than NeRF-based baselines) makes real-time intraoperative use feasible."],"fun_headline_variants":["Diff2DGS: diffusion inpainting then 2DGS for occluded surgery","Inpaint tools first: better 3D reconstruction in occluded surgery","Diff2DGS: 38 dB PSNR by inpainting tools before splatting","Image quality isn't geometry: Diff2DGS adds depth weighting"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method assumes that RAFT stereo depth is a trustworthy geometric reference for instrument-occluded and inpainted tissue regions; if RAFT errors correlate with the masked areas, both initialization and depth evaluation would be biased in the same direction.","fun_headline_variants_meta":{"raw":{"variants":["Diff2DGS: diffusion inpainting then 2DGS for occluded surgery","Inpaint tools first: better 3D reconstruction in occluded surgery","Diff2DGS: 38 dB PSNR by inpainting tools before splatting","Image quality isn't geometry: Diff2DGS adds depth weighting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000966,"raw_usage":{"total_tokens":3968,"prompt_tokens":785,"completion_tokens":3183,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":3106}},"tokens_in":529,"tokens_out":3183,"duration_ms":21110,"temperature":1.0,"reasoning_tokens":3106,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T21:56:10.423942+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a surgical video with known ground-truth depth from a structured-light sensor (like SCARED) or CT, artificially occlude tissue with instrument masks, run Diff2DGS, and compare the recovered geometry in the masked regions against the true anatomy. If the depth RMSE in those regions is no better than Deform3DGS or SurgicalGS, or if appearance improves while geometry diverges, the central claim fails.","supporting_citations":[],"review_version":1}