{"id":"2bae0533-f504-40ac-ba2e-552b0d15bcb2","arxiv_id":"2505.07003","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CMD performs local 3D editing and progressive 3D generation with a conditional multiview diffusion model, editing from a single image in about 20 seconds while preserving unmodified parts.","lead":"CMD is a new 3D-generation system that lets you edit a 3D model by changing just one rendered image, and it builds complex objects part by part in about 20 seconds. It combines a multi-view diffusion model with a ControlNet-style conditioner so that edited regions change while the rest of the model stays mostly intact.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Sec. 5 and Fig. 16 admit that incremental reconstruction fails to preserve mesh topology, directly undercutting the central claim of local editing 'without changing other 3D parts.'","rationale":"The reader's weakest assumption is exactly the load-bearing premise I identify: incremental reconstruction must preserve the unedited geometry and topology. The manuscript's own Sec. 5 and Fig. 16 state that this premise fails, and A.4 repeats it. Because the central contribution of the paper is local editing that leaves other 3D parts unchanged, an admitted failure of the reconstruction stage to maintain topology is a direct contradiction of the central claim, not a minor implementation detail. The method may still produce useful approximate edits, and the generation results may be valuable, but the headline claim as written is not supported. A conditional accept would be appropriate if the authors revise the claim to describe approximate preservation or add a region-locked reconstruction; as submitted, the contradiction is fundamental. The concrete test would settle the severity, but the paper's own text already concedes that the preservation guarantee fails.","tokens_in":16019,"tokens_out":3512,"duration_ms":39036,"concrete_test":"Run the released code on the 20 editing evaluation meshes. Define the unedited region as faces whose rendered projection lies outside the edited-view mask (or, for multi-view safety, outside the union of masks across the six views). After incremental reconstruction, register the output mesh to the input mesh and compute, in that unedited region: (i) per-vertex Chamfer/Hausdorff distance normalized by bounding-box diagonal, (ii) change in Euler characteristic or genus, and (iii) rendered normal-map L2 difference on the non-edited views. If the unedited-region distance exceeds a small tolerance (e.g., 0.5% of the bounding-box diagonal) or the genus changes, the 'without changing other 3D parts' claim is falsified. This test would also quantify whether the qualitative examples in Figs. 1, 6, and 7 conceal global drift.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim of the editing contribution is that the final edited mesh equals the input mesh outside the edited region: the abstract promises editing 'without changing other 3D parts', Sec. 3.3 and Fig. 3 say the generated multiview maps change 'only ... in the edited area', and Table 1 claims outputs that 'strictly follow the given image reference.' The second stage of the pipeline, incremental reconstruction (Sec. 3.2), optimizes Eq. (1) globally with differentiable rendering and continuous remeshing: it initializes from the original mesh but permits iterative face splitting and merging over the whole surface, with no mask restricting optimization to the edited area. The authors' own limitation text states that 'our incremental reconstruction could not maintain the topology of the original mesh' (Sec. 5) and 'fails to preserve the original mesh topology due to random optimization' (Fig. 16); Appendix A.4 repeats that fine-grain topology is not maintained. Since continuous remeshing is a global operation, this is not just a connectivity technicality: unedited geometry can drift. Notably, for progressive generation the appendix describes a sphere-initialization strategy that localizes topology updates ('we initialize a sphere in this region and only update the topology of the sphere'), but no analogous region-localized update is used for editing, even though edits are localized by construction. The editing evaluation also provides no quantitative geometry preservation metric (Appendix A.2: 'Due to the lack of mesh editing ground truth, we only provide qualitative comparisons about geometry manipulation'). Thus the manuscript's own admission directly conflicts with its headline claim, and the supporting evidence is insufficient to show the claim holds to any stated tolerance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CMD, a two-stage framework for image-based local 3D editing and progressive 3D generation. In the first stage, a conditional multiview diffusion model (CondMV), built on Era3D and augmented with a multiview ControlNet, takes a set of rendered color/normal views of an existing mesh plus one edited target view and generates multiview color/normal maps that are claimed to change only in the edited region. In the second stage, the edited mesh is obtained by incremental reconstruction: differentiable rendering with continuous remeshing, initialized from the original mesh, minimizing a normal, alpha, and smoothness loss. For progressive generation, an input image is decomposed by SAM/CLIP/KMeans segmentation and parts are generated sequentially with a global condition. Experiments compare against SDS-based and multiview generation baselines, with CLIP similarity and a user study for editing and reconstruction metrics for generation.","tokens_in":16314,"tokens_out":3100,"duration_ms":32808,"significance":"If the central preservation claim held, CMD would be a practically significant contribution: it offers image-based, 3D-guidance-free local editing of textured meshes in about 20 seconds, and it demonstrates a progressive generation pipeline for complex multi-component assets. The paper has clear strengths: it fine-tunes both the backbone and the ControlNet rather than only the control branch, it constructs a large self-augmented training set with part-level manipulation and object composition, it provides ablations for MVControlNet and the global condition, and it reports a user study and efficiency benchmarks. The progressive generation results in Table 4 and Fig. 10 are encouraging. However, the editing contribution's load-bearing claim, that unedited 3D parts remain unchanged, is explicitly contradicted by the paper's own limitation statements, and the editing evaluation lacks quantitative geometry-preservation metrics.","major_comments":[{"comment":"The abstract and Sec. 3.3 claim that CMD edits a 3D model \"without changing other 3D parts,\" and Fig. 3 states that the generated multiview outputs remain unchanged outside the edited area. This claim is load-bearing for the editing contribution, but the paper's own limitation text contradicts it. Sec. 5 states that \"our incremental reconstruction could not maintain the topology of the original mesh,\" Fig. 16 says the reconstruction \"fails to preserve the original mesh topology due to random optimization,\" and Appendix A.4 repeats that fine-grain topology is not maintained. Because the reconstruction optimizes the global objective in Eq. (1) with continuous remeshing over the whole surface, without any mask restricting face splitting or merging to the edited region, this is not merely a topological technicality: unedited geometry can drift. The paper should either restrict the incremental optimization to the edited region (as is already done for progressive generation in Appendix A.2, where a sphere is initialized and only its topology is updated), or substantially soften the preservation claim and provide quantitative evidence of the drift.","section":"Sec. 3.3 and Fig. 3 vs. Sec. 5, Fig. 16, and Appendix A.4"},{"comment":"The geometry-preservation claim is evaluated only qualitatively. Appendix A.2 explicitly states that \"due to the lack of mesh editing ground truth, we only provide qualitative comparisons about geometry manipulation.\" This is avoidable: since the input mesh is available, the paper could measure preservation by comparing the edited mesh against the input mesh in the unedited region, e.g., Chamfer distance or Hausdorff distance restricted to faces outside the edited area, or changes in vertex/face counts and local surface normals. Without such metrics, the central claim that unedited parts remain unchanged is not quantitatively supported. The CLIP similarity in Table 2 and the user study in Table 5 measure prompt alignment and overall preference, not preservation of unedited geometry.","section":"Sec. 4.2 and Appendix A.2 (Evaluation Metrics)"}],"minor_comments":[{"comment":"\"exiting MVControlNet based methods\" should be \"existing MVControlNet based methods.\"","section":"Sec. 2"},{"comment":"The text alternates between \"Unet\" and \"UNet\"; please use one consistent spelling.","section":"Appendix A.2"},{"comment":"The CLIP similarity scores are reported without error bars or significance tests; given the small evaluation set of 20 models, please report variance or per-case statistics.","section":"Table 2"},{"comment":"The statement \"CMD achieves an 8-times speedup than state-of-the-art mesh editing methods\" is not directly supported by Table 1, where the cited baselines range from 4 minutes to 120 minutes; please clarify which baseline the speedup is computed against.","section":"Sec. 4.2"},{"comment":"The six predefined viewpoints {0, 45, 90, 180, 270, 315} are not evenly distributed; please justify this choice, especially since the edited view is always the 0-degree view.","section":"Sec. 3.3"},{"comment":"The reference to \"OpenArt 2023\" in Sec. 3.3 is minimal; please provide a more complete citation or a description of the editing tool used to generate target images.","section":"Appendix A.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a well-engineered paper with a genuinely useful capability — image-based local 3D editing in ~20 seconds using a multiview ControlNet — and a progressive generation scheme that improves on Era3D. The central claim, however, is overstated: the paper says editing happens \"without changing other 3D parts,\" but its own Sec. 5 and Fig. 16 admit the incremental reconstruction does not preserve mesh topology, and the editing evaluation has no geometry metric. That gap is real, not manufactured.\n\nWhat's new: the MVControlNet extension, which conditions a multiview diffusion model on rendered color and normal maps rather than just text or a single image, is a natural but effective idea. The mix-dropping strategy for classifier-free guidance with multiple conditions is sensible. The progressive generation with a global conditioning image is a nice touch, and the quantitative gains over Era3D/InstantMesh/Unique3D in Table 4 look credible (Chamfer distance, volume IoU, LPIPS all improving). The 20s runtime is a genuine practical advantage over SDS-based editors. The paper also deserves credit for being transparent: the topology limitation is spelled out in the limitations section, not hidden.\n\nSoft spots, in order of importance. First, the abstract and Table 1 promise \"strictly follows the given image reference\" and \"without changing other 3D parts.\" The reconstruction stage optimizes Eq. (1) globally with continuous remeshing, so even if the generated maps are cleanly edited, the mesh can drift or retopologize outside the edit region. The authors acknowledge this but do not fix it for editing, while for progressive generation they do use a localized sphere update. That asymmetry is telling: the editing pipeline could adopt the same strategy but doesn't. Second, the editing quantitative evaluation is thin: CLIP similarity on 20 examples, no error bars, no geometry preservation measure. The user study (Table 5) helps, but it is preference, not fidelity. Third, the reliance on an external image editor is acknowledged but means the whole pipeline inherits its failure modes.\n\nIs the paper worth a serious referee? Yes. The core diffusion idea is solid, the experiments on generation are reasonably convincing, and the limitation is at least stated. The right fix is to either recalibrate the claims or, better, add a mask or a localized reconstruction step and measure geometry preservation. As submitted, it's a conditional accept or borderline; the contribution is real but the headline does not match the evidence.","headline":"Solid multiview-diffusion editing/generation system with a genuine 20-second capability, but the headline claim of local editing without changing other 3D parts is undercut by the paper's own admission that incremental reconstruction fails to preserve mesh topology.","tokens_in":16905,"tokens_out":2436,"would_cite":true,"duration_ms":23783,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CMD edits a 3D mesh from a single edited view in about 20 seconds.","keywords":["conditional multiview diffusion","3D editing","progressive 3D generation","image-based editing","multiview ControlNet","incremental reconstruction","textured mesh","differentiable rendering"],"falsifier":"Render the original and edited meshes from the same fixed viewpoints and compare face normals, vertex positions, and Euler characteristic outside the edited region; if any face outside the edited area moves or the mesh's topology changes measurably, the \"unchanged parts\" claim is false.","tokens_in":15830,"feed_emoji":"🎨","tokens_out":5375,"duration_ms":50013,"temperature":0.7,"pith_summary":"CMD is a generative pipeline that claims 3D editing can be reduced to editing one rendered image and then lifting that edit back onto a mesh. The paper proposes a conditional multiview diffusion model, CondMV, that takes multiview color and normal maps of the original mesh together with the edited view and produces new multiview color and normal maps in which only the edited region changes. Those maps feed an incremental differentiable-rendering reconstruction initialized with the original mesh, yielding an edited textured mesh in about 20 seconds. For complex assets, the same mechanism generates objects part by part, using a global condition to keep each step spatially consistent. The practical payoff is an editing loop that avoids full regeneration and does not require a user to specify 3D regions or hand-crafted proxies.","feed_headline":"One edited image edits a 3D model in ~20 seconds","feed_subtitle":"Swap one rendered view and the pipeline rewrites the textured mesh instead of regenerating the whole model.","key_machinery":"CondMV is a cross-modality multiview diffusion model with a multiview ControlNet: a tiny convolutional network encodes the six-view color and normal conditions, zero-convolution layers inject them into the denoising UNet, and row-wise multiview attention keeps the generated views consistent. A global-condition variant concatenates the final target image as an extra latent channel to guide progressive generation. The other load-bearing piece is incremental reconstruction, which combines differentiable rasterization with continuous remeshing, initializes from the original mesh, and is supervised by normal, alpha-mask, and smoothness losses.","core_discovery":"The paper's central claim is that local 3D editing and progressive 3D generation can both be carried by a single conditional multiview diffusion model. Given a textured mesh, CMD renders color and normal maps from six fixed viewpoints as conditions. A user edits one view with an external image editor, and CondMV, a multiview diffusion backbone augmented with a multiview ControlNet, generates synchronized novel-view color and normal maps. The reconstruction stage then uses continuous remeshing with differentiable rendering, initialized from the original mesh, to produce the edited textured mesh. The paper argues that only the edited area changes in the generated multiviews, that the process needs no explicit 3D guidance, and that the same formulation enables progressive generation when the target is segmented into components and each step is conditioned on the previously generated parts plus a global layout image.","pith_inferences":["If the reported topology drift is addressed by restricting remeshing to the edited faces, the method's \"unchanged parts\" promise would align with its actual output; the paper itself identifies this as future work.","The same conditioning scheme could in principle be applied to other reference modalities, such as multiple edited views, sketches, or depth maps, as long as they can be rasterized into the multiview condition tensor.","The global-condition trick suggests a natural extension to user-guided part-by-part modeling, where each progressive step is steered by a coarse 2D layout rather than a full target image.","Because the pipeline is fully image-based, its fidelity is bounded by the image editor's output; cleaning or constraining the edited view before diffusion would likely translate directly into cleaner 3D edits."],"forward_implications":["A user can perform local geometry and texture edits by editing one rendered image, with the edited textured mesh produced in roughly 20 seconds rather than tens of minutes.","Sequential edits are possible: the output of one edit can serve as the input mesh for the next, supporting iterative refinement.","Complex, multi-part objects can be generated progressively by segmenting a target image and conditioning each step on previous components plus a global layout, improving fidelity over one-shot generation.","Because color and normal maps are generated jointly, geometry and appearance are edited simultaneously rather than in separate passes.","The method needs no manually specified 3D editing regions and no score-distillation optimization, unlike prior text-driven mesh editors."],"supporting_citations":[{"why":"Supplies the base multiview diffusion model whose backbone and row-wise attention CMD extends.","marker":"[Li et al. 2024a]"},{"why":"Establishes the cross-domain color and normal multiview generation convention that CondMV conditions on.","marker":"[Long et al. 2024]"},{"why":"Provides the ControlNet zero-convolution conditioning mechanism that MVControlNet generalizes to multiview inputs.","marker":"[Zhang et al. 2023]"},{"why":"Supplies the continuous remeshing primitive that the incremental reconstruction stage builds on.","marker":"[Palfinger 2022]"},{"why":"Provides the differentiable rasterizer used to compute the reconstruction loss against generated normals and masks.","marker":"[Laine et al. 2020]"},{"why":"Defines the diffusion training objective and sampling procedure used by CondMV.","marker":"[Ho et al. 2020]"},{"why":"Provides the Objaverse LVIS training set, augmented with part manipulation and object composition to train CMD.","marker":"[Deitke et al. 2023]"},{"why":"Supplies the segmentation model used to decompose complex target images into progressive generation steps.","marker":"[Kirillov et al. 2023]"}],"fun_headline_variants":["Edit one image, not the whole 3D model","Single-view edit revises only the edited region in 3D","Conditional multiview diffusion edits 3D from one rendered view","Progressive 3D generation with per-part control via CMD","Local 3D edits from a single image, no full regeneration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that incremental reconstruction, initialized with the original mesh and optimized by differentiable rendering, preserves the unedited geometry and topology while updating only the edited region; the paper states in its limitations that this premise can fail because random optimization does not maintain the original mesh topology.","fun_headline_variants_meta":{"raw":{"variants":["Edit one image, not the whole 3D model","Single-view edit revises only the edited region in 3D","Conditional multiview diffusion edits 3D from one rendered view","Progressive 3D generation with per-part control via CMD","Local 3D edits from a single image, no full regeneration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1339,"prompt_tokens":934,"completion_tokens":405,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":316}},"tokens_in":550,"tokens_out":405,"duration_ms":4206,"temperature":1.0,"reasoning_tokens":316,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:26:40.424037+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render the original and edited meshes from the same fixed viewpoints and compare face normals, vertex positions, and Euler characteristic outside the edited region; if any face outside the edited area moves or the mesh's topology changes measurably, the \"unchanged parts\" claim is false.","supporting_citations":[],"review_version":1}