{"id":"b266c486-17a9-45d2-8674-364a25c69185","arxiv_id":"2607.19777","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 3D Gaussian Splatting editing pipeline that chooses camera positions using diffusion attention statistics and aligns self- and cross-attention across views to produce localized, consistent text-driven edits with fewer views.","lead":"LB-Edit picks editing camera positions by analyzing where a text-to-image diffusion model's attention is concentrated on the target object, then aligns the resulting per-view edits across views. This makes text-driven 3D scene edits more localized and consistent while using far fewer editing views than prior methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"5-step attention proxy is asserted without validation and is the load-bearing condition for ACP's distance selection; if it fails, ACP can pick the wrong d* and the efficiency/localization claims degrade.","rationale":"The reader's weakest assumption is exactly the 5-step attention proxy, and I agree it is the most load-bearing unvalidated premise. The reader's CONDITIONAL verdict with MODERATE confidence remains appropriate: the concern is real but is a request for evidence, not a demonstrated internal contradiction. I considered a few alternatives — the claimed 'by construction' view-consistency of CA alignment (Sec. 3.2) is over-stated because Eq. 8 aligns only scalar attention fields at 32×32, not the full output distribution, so the claim should be softened; and the energy-based placement (Eq. 5, Eq. 12) has heuristically chosen weights (w_vis=1.0, w_can=0.5) without sensitivity analysis. But neither is as load-bearing as the attention proxy: if d* is wrong, both ACP and MAA degrade; if the energy weights are suboptimal, the camera set is still near-optimal because d* dominates the effect. The quantitative evidence is thin (no error bars, small user study) but the paper is honest about limitations and provides enough implementation detail that the central claim is testable. The independent support is modest: qualitative ablations and a 21-participant study, but no code release or large-scale user study. Therefore I keep the verdict UNCHANGED and recommend the concrete test as the single most valuable additional experiment: a gradient of probing steps and full-schedule comparison, plus an off-axis probe sanity check. If that test shows instability, the verdict should move toward REJECT or at least REQUIRE the ACP claims to be re-scoped; if it shows stability, the central claim is substantially strengthened.","tokens_in":17570,"tokens_out":1723,"duration_ms":13944,"concrete_test":"Run ACP's distance probing with 5, 10, 25, and 50-step attention extraction (or the full 1000-step schedule) on the five scenes, and compare the resulting d* selections and S_ca+S_sa rankings; also compare probe-camera results (frontal at d*) with off-axis cameras at the same distance. If the chosen d* (or the ranking of candidate distances) changes across step counts or viewing directions, the proxy is unreliable. Additionally, recompute Table 2/Fig. 5 with d* selected by the full-schedule attention and report CLIPsim, user-study, and latency changes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central premise of ACP is that attention maps extracted from only 5 denoising steps (and from a single frontal probe camera at a few discrete distances) are a reliable oracle for full-schedule InstructPix2Pix behavior across all candidate viewpoints. Appendix A.1 states attention spatial structure stabilizes within the first few steps but provides no quantitative comparison of 5-step vs. full-schedule attention (or of probe-view at the canonical front vs. off-front views). The rest of ACP's objective (Eqs. 1–4) inherits this proxy: there is also no evidence that maximizing S_ca + S_sa at the probe view predicts edit locality/fidelity in the final 3DGS, since the final fine-tuning uses a different distance (d*), different views, and the full denoising schedule. If attention layout is not stable at 5 steps, or if it varies with viewing direction and ROI scale, the selected d* can be wrong — and both the claimed localization benefit and the MAA consistency gains would degrade. The paper's own qualitative ablations (Fig. 4a) are non-quantitative and cherry-pick one example per distance regime, so they do not settle this. This concern is not about consensus mismatch; it is about an unresolved empirical assumption that the entire system rests on.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LB-Edit, a two-stage pipeline for text-driven 3D Gaussian Splatting editing. In the first stage, Attention-Guided Editing Camera Placement (ACP) probes InstructPix2Pix self- and cross-attention at a few candidate camera distances, selecting the distance that maximizes a joint score of cross-attention ROI alignment and self-attention containment (Eqs. 1–4), then places a compact, geometrically diverse set of K editing cameras (Eq. 5, farthest-point sampling). In the second stage, Multi-View Attention Alignment (MAA) synchronizes per-view edits by sharing self-attention features through token correspondence and by replacing cross-attention outputs with a 3D-lifted, re-rendered attention field (Eqs. 6–8). The edited views are used to fine-tune the 3DGS with ROI-masked optimization. The paper reports CLIPsim and a 21-participant user study across five scenes, claiming the highest user preference and up to 7× lower latency than baselines requiring 20–60 editing views.","tokens_in":17910,"tokens_out":4833,"duration_ms":41973,"significance":"If the central claims hold, LB-Edit would be the first diffusion-based 3DGS editor to jointly address editing-aware camera placement, self-attention alignment, and cross-attention alignment, with a substantial efficiency advantage. The method is clearly motivated: fixed COLMAP cameras are indeed optimized for reconstruction rather than for editing a user-selected ROI. The formulation is detailed and the appendices provide concrete implementation choices. However, the empirical support is thin: CLIPsim in Table 2 has no error bars or significance tests, the user study in Table 3 is small (21 participants, 14 tasks) with no inter-rater analysis, and the ablations in Fig. 4 are qualitative. The paper's most load-bearing assumption—that attention maps extracted from only 5 denoising steps and a single frontal probe camera are a reliable oracle for full-schedule behavior—is asserted in Appendix A.1 but not quantitatively validated. The contribution is potentially valuable, but the evidence as presented is insufficient for acceptance.","major_comments":[{"comment":"The ACP distance probing runs InstructPix2Pix for only 5 denoising steps, justified by the claim that 'attention spatial structure stabilizes within the first few steps' (Appendix A.1). This is the load-bearing premise of Eq. (4): if 5-step attention does not predict full-schedule editing behavior across viewpoints and distances, the selected d* can be wrong, and both the localization benefit and the MAA consistency gains degrade. The paper provides no quantitative comparison of 5-step versus full-schedule attention maps, nor a check that the frontal probe view is representative of off-front candidate cameras. Please add such a comparison (e.g., correlation of S_ca and S_sa between 5-step and full schedule across multiple scenes and viewpoints) or substantially soften the claim.","section":"§3.1, Appendix A.1"},{"comment":"The quantitative evidence for the central 'highest user preference' claim is thin. Table 2 reports single CLIPsim point estimates with no standard deviations, repetitions, or significance tests; the single non-winning scene (bear) is rationalized post hoc as a CLIPsim pathology. Table 3 reports 21 participants over 14 tasks with no inter-rater agreement or confidence intervals; several margins are small (e.g., editing locality 32.9% vs 30.6%). Without error bars or paired significance testing, the headline results are not established. Please report per-task and per-participant distributions, paired tests or confidence intervals, and multi-seed CLIPsim values.","section":"§4.3, Tables 2 and 3"},{"comment":"The attention-optimal distance is selected by maximizing S_ca + S_sa, justified by the claim that both scores have 'comparable empirical ranges.' No empirical ranges are actually reported, and the stability claim is only stated for 5 scenes. If the ranges are not comparable, the unweighted sum is an arbitrary objective, and the whole ACP selection inherits that arbitrariness. Additionally, the candidate multipliers m_i (Appendix A.2) and the energy weights in Eq. (5) are free parameters with no sensitivity analysis. Please report the distributions of S_ca and S_sa across distances and scenes, and a sensitivity study over these weights and multipliers.","section":"Eq. (4), Appendix A.5"},{"comment":"The ablations for both ACP and MAA are qualitative, with one example per distance regime and one qualitative panel per alignment component. The textual claims—'Closer-than-Optimal views are filtered out by high SA leakage' and 'Further-than-Optimal rejected due to low CA-to-mask alignment'—should be supported by quantitative metrics (e.g., mean and standard deviation of S_sa, S_ca, and a downstream locality metric such as mask IoU of the edited region) over all scenes and prompts. Similarly, the value of +SA and +CA is shown only as images; provide quantitative consistency and locality numbers. Without this, the central contribution of ACP and MAA is not convincingly validated.","section":"Fig. 4a/4b"}],"minor_comments":[{"comment":"In the Limitations paragraph, 'extending AGEVS to handle multiple disjoint regions' should presumably be 'extending ACP' or a similarly named component; as written it appears to be a typo.","section":"Section 6"},{"comment":"Appendix F repeats the heading 'Evaluation Prompt Details' and the same introductory sentence as Appendix E, and Table 5 appears only once. This looks like a duplication error; remove one appendix.","section":"Appendix F"},{"comment":"The caption mentions '†CLIPsim outlier caused by global color bleeding (see text)', but the dagger is not placed on the DGE(5) row in the table body, and the outlier discussion appears in §4.3 without a clear pointer. Please make the marker and the cross-reference explicit.","section":"Table 4"},{"comment":"The phrase 'we analyze this case Ablation' appears to be missing a reference; it is likely intended to point to Sec. D.2 or Fig. 9. Please correct the citation.","section":"§4.3"},{"comment":"Equation (18) duplicates Eq. (8) verbatim in Appendix C; consider referring to the earlier equation to avoid redundancy.","section":"Eq. (8) vs Eq. (18)"},{"comment":"The candidate distance multipliers are said to be 'adjusted per scene' but no adjustment procedure is described. If this is a free parameter that affects results, state how it was set (e.g., by scene size or prior trials) for reproducibility.","section":"Appendix A.2"},{"comment":"There are two GaussianEditor entries in the references (Chen et al. 2024a and Wang et al. 2024a) and the intro cites both as 'GaussianEditor [Chen et al. 2024a; Wang et al. 2024a]'. Please verify that the citations map to the intended papers and avoid citation ambiguity.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's core idea is compelling and the pipeline is well-specified, but the empirical validation is not yet at the level expected for the claimed state of the art. The most important issue is the unvalidated 5-step attention proxy that underpins ACP; without a quantitative comparison of 5-step versus full-schedule attention, the main technical contribution rests on an unverified assumption. The user study and CLIPsim results also lack the statistical rigor needed to support the headline claims. I recommend major revision rather than rejection, as the central approach appears sound and the missing evidence could plausibly be supplied by the authors."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper has a genuinely new idea: instead of taking the fixed COLMAP cameras as given, it probes the diffusion editor's self- and cross-attention at a few candidate distances and places a compact set of editing cameras where attention is well-contained in the target ROI. The method section is detailed and the design choices are explained with actual numbers (e.g., the occupancy normalization for self-attention leakage). The Table 1 mapping of prior work is useful. MAA's joint alignment of self- and cross-attention in a single forward pass is also new as a combination, and the reference-target partitioning is a pragmatic way to keep it tractable.\n\nBut the empirical support is thinner than the claims. The load-bearing assumption is that 5 denoising steps are enough to estimate attention layout; Appendix A.1 asserts this with no quantitative comparison to the full schedule. If attention at step 5 doesn't predict the layout at the steps that actually form the edit, the selected distance d* can be wrong, and with it the localization and consistency advantages. This needs a direct validation: correlate 5-step scores with full-schedule scores, or show that the selected d* yields better edits than nearby distances in a quantitative way.\n\nThe other gaps are more standard: CLIPsim in Table 2 has no error bars or significance tests (some margins are small: 19.78 vs 19.46 on room); the user study is 21 participants over 14 tasks with no significance test; and both ablations—ACP and MAA—are qualitative images only. For a paper whose two contributions are ACP and MAA, you need quantitative ablations of each. Code and data release would also be expected for a systems paper like this; without it, reproducibility is limited. The limitations section acknowledges the single-ROI assumption, which is honest but doesn't affect the main claims.\n\nNone of this is fatal: the idea is sound and the method description is clear. But as it stands, the paper overstates its evidence. I'd send it to reviewers, but the revision expectations would be substantial: validate the 5-step proxy, add error bars or significance testing, quantify the ablations, and release code.\n\nBottom line: worth engaging with, worth a serious referee, but as-is I'd want major revision before I'd trust the specific numbers.","headline":"Genuinely new attention-guided camera placement, but the 5-step attention proxy is unvalidated and the evidence is too thin to back the claims; worth a major-revision review, not unconditional acceptance.","tokens_in":18359,"tokens_out":7980,"would_cite":false,"duration_ms":68684,"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":"Probing diffusion attention finds the camera distance that makes 3D edits local and consistent.","keywords":["3D Gaussian Splatting","text-driven 3D editing","attention-guided camera placement","multi-view consistency","self-attention alignment","cross-attention alignment","diffusion editing","view selection"],"falsifier":"Render two camera sets at the same scene: one at the distance d* selected by the 5-step attention proxy and one at a distance with a clearly lower S_ca+S_sa. Run the full editing pipeline (including MAA and fine-tuning) on each. If the low-score set matches or beats the high-score set in user-rated instruction fidelity and locality, the attention oracle isn't determining the outcome. A more direct check: compute the rank correlation between 5-step and full-schedule attention maps for a set of scenes, prompts, and distances; a weak correlation would invalidate the proxy.","tokens_in":17449,"feed_emoji":"🎯","tokens_out":6932,"duration_ms":62729,"temperature":0.7,"pith_summary":"LB-Edit argues that the two classic failures of text-driven 3D scene editing—edits that miss or bleed outside the target object, and per-view edits that disagree after fusion—share a root cause: the editing views come from reconstruction cameras that were never chosen with the edit in mind. The paper proposes to ask the diffusion editor itself, via its self- and cross-attention maps, at which camera distance the edit would stay contained in the region of interest (ROI). Attention-Guided Editing Camera Placement (ACP) probes a few candidate distances, scores them by cross-attention concentration (F1 against the projected ROI mask minus background activation) plus self-attention containment (one minus background-to-ROI leakage), and places a diverse, compact camera set at the optimal distance. Then Multi-View Attention Alignment (MAA) shares self-attention features across views through token-level correspondence and re-renders a shared 3D cross-attention field into each view, so appearance and spatial location agree in a single editing pass. The result is that the same scene can be edited with as few as 5 views and up to 7× less latency than methods using 20–60 fixed views, while improving user-rated instruction fidelity, consistency, and locality.","feed_headline":"Five attention-picked views edit 3D scenes 7x faster","feed_subtitle":"New method reads the diffusion editor's own attention to localize edits to one object and keep them consistent across viewpoints.","key_machinery":"The core mechanism is the attention-concentration score that ACP maximizes: S_ca = F1(thresholded cross-attention, projected ROI mask) − (1 − 90th-percentile background cross-attention) and S_sa = 1 − (occupancy-normalized background-to-ROI self-attention leakage). Together they select the editing distance d*. Camera placement then scores candidate directions on a Fibonacci lattice by visibility and canonical alignment, and picks K cameras by farthest-point sampling. For consistency, MAA computes a reference pass that caches extended cross-view self-attention and builds a per-Gaussian 3D attention field via inverse splatting; target views replace their self-attention output with the matched","core_discovery":"The central claim is that the diffusion editor's attention maps are a reliable oracle for camera placement: the distance that maximizes cross-attention ROI alignment and self-attention containment yields edits that stay confined to the target object, and the same attention maps, lifted into 3D, can synchronize edits across views. This turns camera placement from a hand-crafted geometric problem into an attention-driven decision, and turns multi-view consistency into a joint self- and cross-attention alignment problem solved within one U-Net pass. If right, it removes the dependence on many fixed views, making localized 3D editing practical for interactive use.","pith_inferences":["The paper's speed gain is partly an artifact of the proxy (5-step attention as an oracle); a stronger test would compare the selected d* against an exhaustive sweep of distances using full-schedule edits, to establish that the proxy isn't cherry-picking.","The count-averaged inverse-splatting aggregation (rather than transmittance-weighted) may be suboptimal when the ROI is severely occluded from the reference views; an occlusion-aware weighting is an obvious extension.","Treating camera placement as an attention-driven decision could transfer to video editing, where per-frame attention maps could place a temporally consistent camera path.","The single-dominant-ROI assumption suggests the framework would need extension for instructions like 'change both the bear and the vase'; per-object distance probing might be the natural next step."],"forward_implications":["Editing a 3D scene becomes nearly as fast as editing a handful of 2D views: 5 ACP-chosen cameras can replace 20–60 fixed views with equal or better text-image alignment and user preference.","Users gain the freedom to target a specific object in a cluttered scene by natural language, without manually selecting key views or worrying that the object is cropped or too small in training views.","Because self- and cross-attention are aligned in one pass, appearance and spatial drift are suppressed jointly, which should resolve the duplicated- or missing-geometry artifacts that plague instance-level edits like adding glasses or changing a nose.","The same attention-probing principle could be reused for other per-object 3D operations—deletion, recoloring, material changes—wherever the 2D editor's attention can be probed cheaply."],"fun_headline_variants":["Attention-guided cameras make 3D editing 7x faster","One attention pass picks cameras for 3D editing, 7x faster","LB-Edit: Attention picks views to edit 3D scenes 7x faster","Attention maps decide camera placement, cut editing latency 7x","Attention localizes and syncs 3D edits in one pass"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper assumes that attention maps from a single frontal probe camera and only five denoising steps reliably predict the editor's behavior across all viewpoints and the full denoising schedule; if that proxy misranks distances, the chosen camera set would no longer be optimal and the reported localization and consistency gains would degrade.","fun_headline_variants_meta":{"raw":{"variants":["Attention-guided cameras make 3D editing 7x faster","One attention pass picks cameras for 3D editing, 7x faster","LB-Edit: Attention picks views to edit 3D scenes 7x faster","Attention maps decide camera placement, cut editing latency 7x","Attention localizes and syncs 3D edits in one pass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002031,"raw_usage":{"total_tokens":7770,"prompt_tokens":779,"completion_tokens":6991,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":6896}},"tokens_in":523,"tokens_out":6991,"duration_ms":46183,"temperature":1.0,"reasoning_tokens":6896,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T11:43:51.542003+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render two camera sets at the same scene: one at the distance d* selected by the 5-step attention proxy and one at a distance with a clearly lower S_ca+S_sa. Run the full editing pipeline (including MAA and fine-tuning) on each. If the low-score set matches or beats the high-score set in user-rated instruction fidelity and locality, the attention oracle isn't determining the outcome. A more direct check: compute the rank correlation between 5-step and full-schedule attention maps for a set of scenes, prompts, and distances; a weak correlation would invalidate the proxy.","supporting_citations":[],"review_version":1}