{"id":"db03dadf-032f-4821-a6b4-1b8e3a201c8f","arxiv_id":"2507.21872","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A dual-branch diffusion framework jointly edits images and LiDAR point clouds in driving scenes using 3D Gaussian Splatting object priors, improving fidelity and boosting detection of rare vehicle classes.","lead":"MultiEditor edits camera images and LiDAR point clouds together in driving scenes, using a 3D Gaussian Splatting model of the target vehicle as a shared source of appearance and geometry. If it works as reported, self-driving teams could generate realistic training data for rare vehicle types that real datasets under-represent.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quantitative evaluation is reconstruction of the same object conditioned on itself; claimed editing superiority may not generalize to novel-object insertion.","rationale":"I read the paper in good faith. The method is well-engineered and plausibly works. However, the quantitative evidence for the central claim is weaker than it appears. The 'editing' experiments are actually a reconstruction benchmark built by masking an object and giving the model the same object as condition, so the model is being tested on copying rather than novel insertion. This is particularly problematic because the paper emphasizes controllability and rare-category editing; the only tests of actual novel insertion are qualitative (atypical vehicles) or downstream (van augmentation), and the latter lacks a data-quantity control. A simple control experiment—comparing against a baseline that adds an equal number of real van instances—would settle whether the detection gains come from the method's cross-modality consistency or merely from more training samples. My concern does not reject the approach; it says the current evidence is insufficient to support the strength of the claims. This aligns with the reader's conditional verdict, though we identify a different weakest point (evaluation protocol rather than 3DGS availability).","tokens_in":601,"tokens_out":5614,"duration_ms":113353,"concrete_test":"Construct a held-out editing benchmark: insert a 3DRealCar target object that is different from the occluded object into a KITTI scene at predefined poses, and compare MultiEditor, AnyDoor, and AnyDoor+RangeLDM on the same metrics (FID, LPIPS, CLIP-I, CD, FPD, DAS), computing CD/FPD against a 3DGS-rendered ground truth of the inserted vehicle. Also run a detection control: augment the training set with an equal number of real van instances (or AnyDoor+RangeLDM outputs) and compare AP@0.7; if the gains match MultiEditor, the downstream benefit is not evidence for the multimodal editing mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1, the main quantitative evidence for 'superior performance', is generated on a reconstruction benchmark, not an editing benchmark. The supplement states: 'we simulate editing scenarios by deliberately occluding the same target object in image and point cloud modalities' and the model is trained/evaluated by reconstructing those exact occluded objects, with the same object's appearance pasted into the condition (Eq. 7) and its 3DGS depth fed to both branches. The model is therefore tested on recovering objects it has already seen, conditioned on those very objects. Baselines such as AnyDoor and RangeLDM receive a reference image or mask but do not get the same object-specific 3DGS depth paste, so the comparison does not isolate editing ability. For atypical vehicles, the paper provides only qualitative visuals (Figures 5 and 8) with no quantitative metrics, so the claim of flexible editing of atypical vehicles is unsupported by the numbers. Finally, the downstream detection result (Table 3) only compares training with generated samples versus no added samples; it does not control for the increase in positive samples or compare against a simpler augmentation, so the 'substantially enhances' claim may reflect data quantity rather than MultiEditor's joint consistency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"MultiEditor is a dual-branch latent diffusion framework for jointly editing image and LiDAR range-view data in driving scenes, using a 3D Gaussian Splatting (3DGS) model of the target object to render both RGB and depth priors. The image branch combines pasted 3DGS renders, CLIP embeddings, and a reconstruction/refinement dual-branch objective, while the range branch uses a custom VAE-based latent diffusion model with pasted 3DGS depth. A depth-guided deformable cross-modality attention module exchanges intermediate features between the two branches. Experiments on KITTI report image FID/LPIPS/CLIP-I, point-cloud CD/FPD, and DAS, claiming superiority over SD, PbE, AnyDoor, RangeLDM, and AnyDoor+RangeLDM, together with ablations and a downstream detection experiment on rare van-class vehicles. The paper is clearly written and the module design is coherent, but the quantitative evaluation protocol substantially undermines the paper's central claims.","tokens_in":16034,"tokens_out":4887,"duration_ms":64357,"significance":"If the evaluation were unbiased, the paper would make a useful contribution: it proposes a novel architectural use of 3DGS priors for joint image-and-LiDAR object editing, with a plausible mechanism for cross-modality feature exchange. The training pipeline is detailed, and the downstream detection experiment addresses a practically important long-tail problem. The main limitation is that the quantitative results in Table 1 are generated on a reconstruction benchmark in which the model receives the very object it is asked to reproduce, rather than on a genuine editing benchmark with novel objects. This confound affects the headline claims of visual and geometric superiority, and the atypical-vehicle and downstream claims are currently supported only by qualitative examples and an uncontrolled augmentation experiment. The method itself is not inherently unsound, so the issues are addressable with additional experiments, but they are load-bearing for the paper's conclusions.","major_comments":[{"comment":"The quantitative evaluation is a reconstruction benchmark, not an editing benchmark. The supplement states that editing scenarios are simulated by 'deliberately occluding the same target object' and training the model to reconstruct it. In Eq. (7), the pixel-level condition hp_c is built by pasting the target object's image into the ROI, and in Eq. (12) hp_r is built from the 3DGS-rendered depth of the same object. The model is therefore conditioned on the appearance and geometry of the very object it is asked to generate. Low CD and FPD in Table 1 largely measure how faithfully the output copies the injected condition, rather than measuring editing ability for novel objects or novel placements. This circularity is load-bearing for the claim of 'superior performance' in geometric fidelity and suggests the numbers in Table 1 should not be read as evidence of editing superiority.","section":"Experiments—Dataset Construction; Eqs. (7) and (12)"},{"comment":"The comparison in Table 1 does not give the baselines the same object-specific 3DGS prior. AnyDoor receives a reference image but not the 3DGS depth paste, and RangeLDM is evaluated without the 3DGS structural condition. Since the 3DGS render is a major part of the proposed method's advantage, the comparison conflates the effect of the object prior with the effect of the editing framework. A fair comparison would either provide all methods with an equivalent 3DGS-rendered condition or evaluate on a protocol where the target object is not the same as the pasted condition, e.g., inserting a 3DRealCar asset into a held-out KITTI scene. Without such a comparison, the claimed superiority over prior editing methods is not established.","section":"Table 1 and Baselines"},{"comment":"The claim that MultiEditor is 'the first framework capable of flexibly editing atypical vehicles' is supported only by qualitative visuals. No quantitative metrics are reported for the 3DRealCar insertion experiments, so it is impossible to assess fidelity, geometric accuracy, or cross-modality consistency for novel object types. The paper should report metrics comparable to Table 1 on a held-out set of atypical-vehicle insertions, including different viewpoints and distances, and ideally include a baseline that receives the same 3DGS prior.","section":"Editing of Atypical Vehicles; Figures 5 and 8"},{"comment":"The downstream detection experiment does not control for the increase in positive training samples. The Real+Gen condition adds 1,192 far-side van samples to the training set; any augmentation method that adds labeled van instances would likely improve AP on a rare class. The paper should compare against adding the same number of real van instances, or against simpler augmentation such as copy-paste of real van points and images, and should report recall and per-class AP in addition to AP@0.7 and mAP. Without these controls, the 'substantially enhances detection accuracy' claim cannot be attributed to MultiEditor's joint consistency rather than to sample count.","section":"Downstream Task Benefits; Table 3"},{"comment":"The cross-modality DAS metric is partially circular. The 3DGS-rendered depth of the target object is injected into both the image and the range-image branches as a condition, and DAS then compares the generated point cloud against depth estimated from the generated image. Both outputs therefore share the same 3DGS depth prior, so a low DAS may reflect consistency with the injected prior rather than consistency with independently verified geometry. The paper should also report a cross-modality metric that does not rely on the 3DGS depth used during generation, for example by projecting the generated point cloud onto a held-out LiDAR scan or by using an independent depth sensor.","section":"Evaluation Metrics—DAS"}],"minor_comments":[{"comment":"The text introducing the two denoising models says 'we introduce two denoising models, ϵθr and ϵθr', but the second should clearly be ϵθc; please fix this typo.","section":"Eq. (3)"},{"comment":"The references for the VGG-19 network and the ImageNet pretraining are missing; the text shows '(?)' in both places.","section":"Eq. (10)"},{"comment":"The normalization factor d in the projection equation is not defined; please specify that it is the depth of the transformed point and how it is computed from the range value.","section":"Eq. (14)"},{"comment":"The 'feature addition condition' ablation row is not described in the method section; please either define it in the main text or move the definition to the supplement with a clear reference.","section":"Table 2"},{"comment":"The paper compares against combined AnyDoor+RangeLDM but does not compare with multi-modal editing or synthesis baselines such as X-Drive, MObI, or GenMM; a brief explanation of why these are not included in Table 1 would help the reader calibrate the claims.","section":"Related Work"}],"recommendation":"major_revision","confidential_remarks":"The core architectural idea is timely and plausible, and the paper has useful engineering detail. My main concern is that the quantitative evaluation is currently a reconstruction benchmark with the target object injected as the condition, which makes the headline numbers in Table 1 difficult to interpret as evidence of editing superiority. This is not an unfixable flaw: the authors could add a genuine novel-object editing benchmark with fair baselines, quantify atypical-vehicle results, and control the downstream detection experiment for sample-count effects. I therefore recommend major revision rather than rejection. I would also gently encourage the authors to make the evaluation code and dataset splits public, as the current manuscript promises code only upon acceptance and the dataset is manually screened without a release statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zhi, this one is worth a look, but the headline numbers don't support the claims as written. The genuinely new ingredient is using a 3D Gaussian Splatting model as a shared appearance-and-depth prior for both the image and LiDAR branches. That is a sensible and useful idea for object-level editing in driving scenes, and the roller insertion at 45° intervals in Fig. 8 is a nice demonstration of controllability. The dual-branch latent diffusion architecture and the depth-guided deformable cross-modality module are reasonable engineering; the ablations show each component moves at least one metric.\n\nThe soft spot is the evaluation. The authors train and test by occluding a real object and asking the model to recover it, while giving the model the very object's 3DGS-rendered RGB and depth pasted into the mask. That turns the benchmark into reconstruction with the answer in the input, so the low Chamfer distance and FPD partly measure copying, not editing. AnyDoor and RangeLDM do not get the same object-specific depth paste, so the comparison in Table 1 does not isolate editing ability. The cross-modality module's ablation gain is also small: removing it changes DAS from 3.16 to 3.20, and FPD from 97.5 to 98.8 — real but hardly 'significantly enhancing'. On atypical vehicles the paper gives only qualitative results, and the downstream detection experiment adds 1,192 synthetic samples to 1,297 real ones, roughly doubling positive count, without controlling for sample quantity or comparing a simpler augmentation. That makes the 'substantially enhances' claim underdetermined.\n\nThere is also the structural precondition: the method needs a clean per-object 3DGS model. If that's unavailable or coarse, both branches degrade. The supplementary acknowledges code will be released only after acceptance, so nothing is independently checkable right now.\n\nOverall, the core idea is worth pursuing, and a serious reader can get value from the architecture and the qualitative demos. But the quantitative evidence needs a proper editing benchmark where the object is new to the scene or at least held out, and baselines need the same conditioning information. I'd send it to peer review, but I'd expect a major revision before acceptance.","headline":"A plausible 3DGS-prior framework for joint image/LiDAR editing, but the main evaluation is reconstruction-with-the-answer, so the headline numbers overstate editing performance.","tokens_in":16581,"tokens_out":2584,"would_cite":false,"duration_ms":28769,"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":"A dual-branch diffusion framework with 3D Gaussian Splatting priors jointly edits images and LiDAR point clouds in driving scenes, and the generated rare-vehicle data improves downstream detection on underrepresented classes.","keywords":["multimodal object editing","3D Gaussian Splatting","latent diffusion models","LiDAR point clouds","driving scene generation","cross-modality consistency","long-tail data augmentation","rare vehicle classes"],"falsifier":"Take a scene and a target object for which a 3DGS model exists, then corrupt the 3DGS-rendered depth map by adding realistic noise (e.g., 0.5 m error) while leaving the rendered appearance identical; if the Chamfer Distance and Depth Alignment Score stay near the uncorrupted values, the geometric-prior claim is not load-bearing. Alternatively, attempt to insert a vehicle category with no prebuilt 3DGS template; if cross-modality consistency collapses to the naive-baseline level, the claim that 3DGS priors are what enable joint editing is falsified.","tokens_in":15616,"feed_emoji":"🚗","tokens_out":6770,"duration_ms":70086,"temperature":0.7,"pith_summary":"MultiEditor aims to show that a single dual-branch diffusion model can edit an RGB image and a LiDAR point cloud of the same driving scene at once, inserting one object into both modalities so that its appearance in the image and its geometry in the point cloud agree. The paper's core move is to make a 3D Gaussian Splatting (3DGS) model of the target object the shared prior: 3DGS renders both the object's appearance and its depth from any viewpoint, and those renderings condition both editing branches. The authors claim this joint approach outperforms separately combined single-modality editors on visual and geometric fidelity, controllability, and cross-modality consistency, and that synthetic rare-vehicle samples improve downstream 2D and 3D detection accuracy on underrepresented classes. A reader should care because long-tailed rare vehicles are safety-critical and scarce in real data, so a method that manufactures consistent multimodal training samples could directly improve perception robustness.","feed_headline":"Joint image–LiDAR editing with 3DGS priors lifts rare-class detection","feed_subtitle":"One model edits camera and LiDAR together, keeping geometry consistent and boosting rare-class detection.","key_machinery":"The load-bearing object is the depth-guided deformable cross-modality condition module (Eqs. 13-18). It first converts a range-image cell (φ, θ) with range r into 3D Cartesian coordinates, projects that point into the image plane using the camera intrinsics and LiDAR-to-camera extrinsics, and then uses deformable cross-attention to sample local features around the projected pixel from the other modality's latent. A zero-initialized tanh gate lets the module start as identity and gradually learn how strongly to inject cross-modal features. This module is what turns the coarse geometric correspondence from 3DGS-rendered depth into adaptive mutual guidance, and the ablation without it (w/o cross-modality, DAS 3.20 vs 3.16; FPD 98.80 vs 97.49) shows it is the component the consistency claim rests on.","core_discovery":"On the paper's own terms, the central discovery is that joint editing of images and point clouds is feasible and beneficial when the target object is represented by a 3D Gaussian Splatting model that supplies both appearance and geometric depth. The dual-branch denoising formulation (Eqs. 3-5) makes each modality's noise prediction conditional on features extracted from the other modality's latent, anchored by 3DGS-rendered depth through the depth-guided deformable cross-modality condition module. The paper reports that this design yields a Chamfer Distance of 1.65 for point cloud editing versus 33.23 for the range-only baseline, and a Depth Alignment Score of 3.16 versus 11.34 for the naive combination of separate editors, and that augmenting van-class training data with generated samples improves YOLOv5 mAP and PointPillars AP on that rare class.","pith_inferences":["A direct consequence the paper does not explore: the method's benefit should be largest exactly where 3DGS templates are hardest to build—extreme long-tail classes with few real views—so the practical ceiling is set by template availability, not by the diffusion model.","One could test a natural extension by replacing the 3DGS depth prior with a learned monocular depth estimate; if cross-modality consistency degrades sharply, that would confirm the specific role of geometric priors versus learned correspondence.","The DAS metric relies on a pretrained depth estimator; our reading is that gains on DAS might partly reflect the estimator's sensitivity to synthetic textures, so a direct LiDAR-image alignment metric (e.g., projected-point overlap) would be a sturdier check of the consistency claim."],"forward_implications":["If MultiEditor is correct, autonomous-driving datasets can be rebalanced by inserting rare vehicle categories into both images and point clouds in a geometrically consistent way, without manual annotation of the synthetic samples.","Training 2D and 3D detectors on real plus generated data should improve accuracy on underrepresented classes; the paper reports +4.5% AP@0.5 for YOLOv5 and +5.34% AP Easy for PointPillars on the van class.","The same joint-editing pipeline could serve as a controllable test-bed for perception models, varying object pose, distance, and orientation at 45-degree intervals as shown for atypical vehicles.","Because both branches are trained jointly with a shared 3DGS prior, the framework is a step toward a unified generative model over camera and LiDAR rather than independent per-modality generators."],"supporting_citations":[{"why":"Supplies the 3D Gaussian Splatting representation used to render RGB images and depth maps that serve as the shared appearance and structure prior for target objects.","marker":"(Kerbl et al. 2023)"},{"why":"Provides the latent diffusion backbone and variational autoencoder used in both branches, as well as the inpainting model used during dataset construction.","marker":"(Rombach et al. 2022)"},{"why":"Paint-by-Example initializes both diffusion branches and serves as a single-modality image-editing baseline that MultiEditor is compared against.","marker":"(Yang et al. 2023)"},{"why":"AnyDoor is an image-editing baseline and one half of the combined multi-modality baseline used for cross-modality comparison.","marker":"(Chen et al. 2024)"},{"why":"RangeLDM is the point-cloud editing baseline and the other half of the combined baseline, and it supplies the range-image diffusion approach that the LiDAR branch adapts.","marker":"(Hu, Zhang, and Hu 2024)"},{"why":"X-Drive provides the cross-modality generation protocol and the Depth Alignment Score used to evaluate consistency between generated images and point clouds.","marker":"(Xie et al. 2024)"},{"why":"3DRealCar supplies the RGB-D car data used to build 3DGS models of atypical vehicles such as road rollers.","marker":"(Du et al. 2024)"},{"why":"Deformable attention is the mechanism the cross-modality module adapts for adaptive local feature sampling around projected correspondences.","marker":"(Xia et al. 2022)"}],"fun_headline_variants":["3DGS priors enable joint image-LiDAR editing for rare-class lift","Controllable multimodal editing with 3DGS boosts rare-class accuracy","Rare-class detection improves via 3DGS-guided joint image-LiDAR editing","One model edits camera and LiDAR with 3DGS, lifting rare-class detection","3DGS prior makes image-LiDAR editing consistent, helping rare-class detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes a good 3D Gaussian Splatting model of each object to be inserted already exists, and that it renders accurate images and depth from the viewpoints needed; if that model is absent, coarse, or inaccurate, the pixel pastes and depth-guided cross-modal alignment lose their anchor and editing fidelity degrades.","fun_headline_variants_meta":{"raw":{"variants":["3DGS priors enable joint image-LiDAR editing for rare-class lift","Controllable multimodal editing with 3DGS boosts rare-class accuracy","Rare-class detection improves via 3DGS-guided joint image-LiDAR editing","One model edits camera and LiDAR with 3DGS, lifting rare-class detection","3DGS prior makes image-LiDAR editing consistent, helping rare-class detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001115,"raw_usage":{"total_tokens":4639,"prompt_tokens":935,"completion_tokens":3704,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":3598}},"tokens_in":551,"tokens_out":3704,"duration_ms":29896,"temperature":1.0,"reasoning_tokens":3598,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:16:40.119936+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a scene and a target object for which a 3DGS model exists, then corrupt the 3DGS-rendered depth map by adding realistic noise (e.g., 0.5 m error) while leaving the rendered appearance identical; if the Chamfer Distance and Depth Alignment Score stay near the uncorrupted values, the geometric-prior claim is not load-bearing. Alternatively, attempt to insert a vehicle category with no prebuilt 3DGS template; if cross-modality consistency collapses to the naive-baseline level, the claim that 3DGS priors are what enable joint editing is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Paint-by-Example initializes both diffusion branches and serves as a single-modality image-editing baseline that MultiEditor is compared against."},{"cited_title":"E.; and Huang, G","cited_arxiv_id":null,"evidence_quote":"Deformable attention is the mechanism the cross-modality module adapts for adaptive local feature sampling around projected correspondences."}],"review_version":1}