{"id":"873a1ff1-7d4e-4f1b-8b3a-090395b37c87","arxiv_id":"2412.12766","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"FreeEdit performs training-free 3D object insertion, replacement, and deletion by chaining GPT-4, Shap-E, OpenMask3D, Grounding DINO, and a voxel-erosion placement locator.","lead":"FreeEdit is a text-driven 3D scene editing system that inserts, replaces, or deletes objects in room-sized meshes without per-scene training. It combines large language models, a text-to-3D generator, and an automatic placement heuristic instead of optimizing a neural radiance field.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The quantitative basis for the central 'minimal-intersection placement' claim is invalid: §4.1's Eq. 3 counts SDF>0 as penetration, contradicting §3.1 where negative SDF means inside, so Table 1 may not measure penetration at all.","rationale":"The paper's novelty claim is a training-free insertion/replacement pipeline whose key contribution is an automatic location finder that minimizes mesh intersection. The only quantitative evidence for that contribution is Table 1, computed with Eq. 3. The equation's indicator direction contradicts the paper's own SDF definition, making the headline number ambiguous. If the sign is wrong, the metric could be rewarding penetration (e.g., a deeply embedded object has fewer vertices outside the scene) or measuring something unrelated to collisions between the inserted object and existing furniture. Qualitative renders and a small user study provide some support, but they do not disentangle placement quality from object-generation quality, and the user study compares against a weak 'center-of-grounding-object' baseline. Thus the central quantitative claim is currently unverified. The reader's verdict already flags 'inconsistent definition of the penetration metric' but their weakest_assumption focuses on Eq. 2's dimensional mixing; I agree partially and elevate the metric issue because it directly invalidates the only objective comparison. The proposed synthetic test is cheap and decisive: with a known intersection percentage, the correct SDF convention is unambiguous. If the test shows Eq. 3 fails, the paper must be revised (or at least the evaluation rerun) before the 'minimizes intersection' claim can be assessed; that supports the reader's CONDITIONAL verdict but with an explicit condition.","tokens_in":12562,"tokens_out":6797,"duration_ms":65980,"concrete_test":"Create a synthetic scene with a single unit cube as an obstacle and a sphere as the primary object; place the sphere so that a known 20% of its vertices lie inside the cube. Compute the §4.1 metric as written (count dist>0) and as described in prose (count dist<0) using the paper's sign convention. A correct penetration metric must report ~20% under the intended convention; if Eq. 3 reports 0% or a value unrelated to the known overlap, the metric is mis-specified. Then recompute the 1250-case evaluation from §4.1 with the corrected sign and compare Table 1; if FreeEdit's advantage over the baseline disappears, the central placement claim loses its quantitative support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that FreeEdit's location finder places objects with minimal intersection—is supported quantitatively only by Table 1's penetration-percent metric. That metric is internally inconsistent. §3.1 defines the SDF so that negative distances lie inside a mesh and positive distances lie outside. But Eq. 3 in §4.1 defines penetration percent as the fraction of primary-object vertices with dist(vn, M) > 0, and the surrounding text first says 'count of primary object vertices having negative signed distance' and then claims intersection 'leads to positive sign.' Both cannot be true. If a primary-object vertex lies inside an existing table/chair mesh (the case the algorithm is supposed to avoid), the SDF according to the paper's own convention is negative; Eq. 3 would not count it. If instead the metric counts vertices outside the closed scene, it does not measure object-object intersection at all. Moreover, SDF sign is only well-defined for watertight meshes; ScanNet/Replica reconstructions are not guaranteed watertight, so sign may be undefined. Until the SDF convention and the indicator direction in Eq. 3 are reconciled, the reported 3.84% vs 7.45% reduction is uninterpretable as evidence for the placement claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FreeEdit, a training-free pipeline for text-guided 3D scene editing on mesh representations, supporting object insertion, replacement, deletion, translation, rotation, and iterative insertion. The pipeline combines an LLM for task classification and entity extraction, Shap-E for text-to-3D object synthesis, OpenMask3D for open-vocabulary object grounding, Grounding DINO for scale estimation, a hierarchical voxel-erosion location finder, and a rotation-based refinement step that minimizes a penetration percentage metric. The authors evaluate against Instruct-NeRF2NeRF and a traditional center-placement baseline on ScanNet and Replica scenes, reporting quantitative penetration percentages and a user study. The central claimed contribution is an automatic placement algorithm that minimizes mesh intersection with prior scene objects.","tokens_in":12839,"tokens_out":3486,"duration_ms":32177,"significance":"If the claims are substantiated, FreeEdit would be a timely and practically relevant contribution: a modular, training-free alternative to NeRF-based 3D editing that can operate on room-sized multi-object scenes without positional priors. The qualitative results and user study suggest the pipeline can produce reasonable insertions and replacements. The paper also explicitly acknowledges several limitations (flat-surface restriction, inpainting artifacts, lack of relational placement, scale-threshold issues), which is constructive. However, the quantitative evidence currently does not support the central minimal-intersection claim as cleanly as presented: the defining equation for the headline metric is inconsistent with the stated SDF convention, and the refinement step optimizes exactly the metric used for evaluation. These issues are load-bearing and must be resolved before the contribution can be assessed fairly.","major_comments":[{"comment":"The definition of penetration percent is internally inconsistent. Section 3.1 defines the SDF with negative sign for inside points and positive for outside points, but Eq. (3) counts vertices with dist(v_n, M) > 0 as penetrated, which is the opposite of the text immediately above it ('count of primary object vertices having negative signed distance'). The explanatory sentence 'intersection results in vertices penetration out of the scene leading to positive sign' contradicts the convention established in Section 3.1. Because Table 1 is the only quantitative support for the central claim that the location finder reduces intersections, this inconsistency makes the reported 3.84% vs 7.45% reduction uninterpretable. Please reconcile the sign convention, specify whether M is the union of all scene mesh objects or just the grounding object, and clarify how SDF is computed for non-watertight ScanNet/Replica meshes.","section":"§4.1, Eq. (3)"},{"comment":"The evaluation metric is the same objective used by the refinement step. The refinement step in Section 3.7 rotates the primary object to minimize penetration percent defined by Eq. (3), and Table 1 then reports penetration percent as the headline quantitative result. The improvement from 4.40% to 3.84% is therefore partly by construction and does not independently validate the placement quality. The claim that FreeEdit finds placements with minimal intersection needs support from a metric not used in the optimization, for example an independent mesh-intersection volume computed after the fact, or a comparison against random rotations/placements on the same scenes.","section":"§3.7 and §4.1"},{"comment":"The coordinate reconstruction formula appears dimensionally inconsistent and underspecified. The term (x_id + levels - 0.5) * width adds an integer number of hierarchy levels to a voxel index and multiplies by a voxel width, which mixes an ordinal count with spatial units; if 'levels' denotes the number of erosion levels, this changes the coordinate mapping in a way that is not derived. In addition, the algorithm description does not specify how ties are broken when more than one cell in the final grid has value 1, and it asserts that the selected location 'minimizes the intersection with prior objects' without describing the optimization over candidate cells. Please provide a precise pseudocode, the tie-breaking rule, and a derivation of Eq. (2).","section":"§3.6, Eq. (2)"},{"comment":"The scale of the primary object is chosen as the minimum of the scales measured over a small set of generated images, but no ablation or principled justification is given for this selection rule. Since object scale directly determines both penetration and the realism of the placement, and Table 1 aggregates results across all 1250 cases, the arbitrary minimum rule is a load-bearing free parameter. Please provide an ablation or justification for the scale selection, and report sensitivity to the number of generated images and to the choice of bounding-box dimension.","section":"§3.5 and §4.1"}],"minor_comments":[{"comment":"Throughout Section 3.6, 'slither' should be 'slide' (e.g., 'slithered over the grid').","section":"§3.6"},{"comment":"Figure 2 mentions 'Top-k views' and 'Clip' but the number k and the CLIP aggregation procedure are not defined in the text; please clarify how the top-k views are selected and used.","section":"Fig. 2"},{"comment":"The DBSCAN parameters (eps, min_samples) and the convolution threshold are not specified; please report the chosen values or state that they are fixed across all scenes.","section":"§3.6"},{"comment":"The user study reports mean scores but no number of responses per question, no variance, and no significance test; given the small sample size (35 participants), please add these details.","section":"§4.1"},{"comment":"The related work section would benefit from explicitly discussing recent mesh-based or training-free editing methods, if any exist, to sharpen the 'first training-free approach' claim.","section":"§2"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely problem and the modular pipeline is attractive, but the quantitative evaluation needs to be redone with a consistent metric and, ideally, an independent one. The most serious problem is the sign inconsistency in Eq. (3), which is fixable but undermines Table 1 as published. I would also encourage the authors to report per-scene breakdowns and failure cases. The claim to be the 'first' training-free approach would benefit from a more careful positioning relative to concurrent or unpublished work; this does not affect my recommendation. Overall, the central idea is defensible, but the current evidence base is not sufficient for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know this paper is a promising engineering integration, but its headline quantitative claim doesn't survive contact with its own equation. The penetration percent metric used in Table 1 is defined inconsistently: Section 3.1 says SDF is negative inside, positive outside, but Eq. 3 counts dist(vn, M) > 0 as penetration. If a primary object vertex is truly inside the scene mesh, the SDF would be negative, so Eq. 3 would not count it. The text tries to explain this away by saying penetration results in 'vertices penetration out of the scene leading to positive sign,' which doesn't make sense. As written, the 3.84% vs 7.45% reduction is uninterpretable.\n\nWhat's actually new: the combination of GPT-4 for task parsing, Shap-E for object synthesis, OpenMask3D for open-vocabulary grounding, Grounding DINO for scale estimation, and a voxel-erosion placement heuristic into a single training-free pipeline for insertion, replacement, deletion on room-scale meshes. That integration hasn't appeared in the cited literature, and the qualitative figures suggest it works on some real scans. The interactive translation/rotation/iterative insertion stuff is a nice extra. I give them credit for a working system and for stating limitations clearly.\n\nSoft spots, in order of severity. The metric is the big one. Second, the placement algorithm is underspecified: Eq. 2 mixes 'levels' with voxel width in a way that looks dimensionally off, and tie-breaking among candidate cells isn't described. Third, the comparison baselines are weak—they use a 'traditional baseline' that is their own pipeline minus the location finder, and they don't compare against existing insertion methods like InseRF or Control-NeRF. Fourth, the refinement step optimizes the same penetration metric used for evaluation, so part of the reported gain is by construction. Fifth, SDF sign is only reliable for watertight meshes; ScanNet and Replica reconstructions are not guaranteed watertight, so the metric may be undefined.\n\nOn balance, I think this deserves peer review. The central idea—training-free, text-driven editing of mesh scenes via foundation models—is timely and the system appears to work qualitatively. The flaws are correctable in revision: fix the metric, clarify the placement math, and add a stronger baseline comparison. I'd be skeptical about the numbers until then, but I wouldn't desk reject it.\n\nRecommendation: send to review, with a strong request for a corrected metric and a reproducibility note.","headline":"A promising training-free 3D editing system, but the main quantitative claim is undercut by an inconsistent penetration metric; warrants review with major revision.","tokens_in":13389,"tokens_out":2398,"would_cite":false,"duration_ms":22062,"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":"The paper claims that FreeEdit, a modular mesh-based pipeline, can insert, replace, or delete objects in large 3D scenes from a text prompt alone, without any per-edit training, by locating a collision-free placement automatically.","keywords":["3D scene editing","training-free","text-guided editing","mesh representation","object insertion","object replacement","object deletion","collision-free placement"],"falsifier":"Take a planar slab with two identical flat regions symmetric under 90-degree rotation. If the algorithm has no tie-breaking rule, rotating the entire scene mesh by 90 degrees should change the chosen placement cell, which would show the claimed 'optimal location' is not well-defined and would invalidate the collision-minimization claim as stated.","tokens_in":12377,"feed_emoji":"🏠","tokens_out":9763,"duration_ms":72795,"temperature":0.7,"pith_summary":"FreeEdit aims to make text-driven editing of 3D scenes practical without retraining a neural radiance field for each edit. It uses mesh representations of room-sized scenes and treats insertion, replacement, and deletion as basic building blocks that can be composed for iterative edits. Given only a scene and a prompt, it infers what object to edit and where, then places a generated object at a spot chosen to minimize mesh intersection. The paper claims this is the first training-free approach for multi-object 3D scene editing, offering near real-time interaction compared to NeRF-based methods that must re-optimize per edit.","feed_headline":"No retraining needed: text prompt edits 3D scenes","feed_subtitle":"Mesh-based pipeline inserts, replaces, deletes objects and avoids collisions for near real-time scene edits.","key_machinery":"The key mechanism is the hierarchical voxel-erosion location finder. On the grounding object, vertices whose normals point upward (roughly parallel to the global Z axis) are selected; these are clustered, and a voxel grid is built with cell size equal to a fraction of the scaled object's width. A small all-ones filter is convolved over the grid, and cells whose average exceeds a threshold are kept, producing a coarser grid; this erosion repeats until no cell survives or the grid is smaller than the filter. The surviving cell in the last grid is interpreted as the optimal placement, and its coordinates are reconstructed by Eq. 2. The procedure encodes the paper's claim that flat, spacious, unoccupied regions are the right place to put an object, and it is the part of the pipeline that replaces user-provided bounding boxes or masks.","core_discovery":"The central discovery is that 3D scene editing can be decomposed into a sequence of frozen, pretrained components plus one lightweight geometric search, eliminating per-edit optimization. The pipeline classifies the prompt into insertion, replacement, or deletion, extracts the primary object and the grounding object, synthesizes a 3D mesh for the object, grounds the target object in the scene, and scales the object using relative bounding-box widths from generated 2D images. The distinctive step is the location finder: it filters the grounding object's upward-facing vertices, clusters them, voxelizes the surface at a resolution tied to the object width, and repeatedly applies an averaging convolution over the grid to shrink to a stable candidate cell. That cell is used to compute the placement coordinates, aiming to minimize intersection with the rest of the scene. For replacement, the grounded object is removed, the cavity is filled, and the replacement mesh is scaled to fit. The paper reports the location finder cuts penetration-percentage roughly in half relative to a center-placement baseline, both before and after a rotation-refinement step.","pith_inferences":["A natural extension would be to replace the upward-normal filter with a local-supportness criterion so the placement search handles non-flat surfaces, which the paper explicitly lists as a current limitation.","The minimum-width-ratio scaling rule could be swapped for a learned object-size predictor, since the paper notes that large scaling values produce unrealistic placements.","The same voxel-erosion search could be reused as a generic surface-placement cost map for robotic manipulators or virtual staging tools, because it only needs a mesh and a footprint size.","The paper measures quality primarily by geometric intersection; an alternative evaluation could compare semantic plausibility of the chosen locations against human judgments, since placement appropriateness is not fully captured by vertex penetration."],"forward_implications":["If FreeEdit works as claimed, 3D scene editing no longer requires training a NeRF for each edit, enabling near real-time, iterative customization of room-sized scenes.","Insertion, replacement, deletion, translation, rotation, and iterative multi-object insertion all run through the same frozen-component pipeline, so a stacked prompt can add several objects one after another without retraining.","Because the pipeline operates on meshes, it can be applied to both reconstructed real-world scans and synthetic room-scale scenes, as demonstrated in the paper.","The location finder's automatic placement removes the need for positional priors such as bounding boxes or masks, reducing user interaction to the text prompt itself.","The penetration-percent metric drops by roughly half compared with a center-placement baseline, and a rotation-refinement step further reduces intersections."],"supporting_citations":[{"why":"Supplies the large language model that classifies the edit task and extracts the primary and grounding entities from the text prompt.","marker":"[1]"},{"why":"Generates the 3D mesh of the object to insert or replace, conditioned on the extracted entity text.","marker":"[18]"},{"why":"Performs open-vocabulary instance segmentation in the 3D scene to ground the target object named in the prompt.","marker":"[38]"},{"why":"Detects bounding boxes in 2D images, providing the widths used to scale the primary object relative to the grounding object.","marker":"[24]"},{"why":"Generates 2D images from the text prompt that are used by the detector for the relative-scaling computation.","marker":"[32]"},{"why":"Serves as the main NeRF-based baseline that requires per-edit training and that FreeEdit compares against qualitatively and quantitatively.","marker":"[15]"}],"fun_headline_variants":["Text-driven 3D edits without any training","Real-time 3D scene editing, training-free","Insert, replace, delete in 3D with a text prompt","Mesh-based 3D editing: no per-edit training","FreeEdit: training-free 3D scene editing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method depends on the assumption that the multi-level voxel-erosion search always settles on a single best cell and that Eq. 2 turns that cell into correct coordinates, but the paper does not state how ties among candidate cells are broken, and the formula's mixing of hierarchy level with voxel width is left unexplained.","fun_headline_variants_meta":{"raw":{"variants":["Text-driven 3D edits without any training","Real-time 3D scene editing, training-free","Insert, replace, delete in 3D with a text prompt","Mesh-based 3D editing: no per-edit training","FreeEdit: training-free 3D scene editing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000254,"raw_usage":{"total_tokens":1602,"prompt_tokens":1016,"completion_tokens":586,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":506}},"tokens_in":632,"tokens_out":586,"duration_ms":4959,"temperature":1.0,"reasoning_tokens":506,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:45:20.646811+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a planar slab with two identical flat regions symmetric under 90-degree rotation. If the algorithm has no tie-breaking rule, rotating the entire scene mesh by 90 degrees should change the chosen placement cell, which would show the claimed 'optimal location' is not well-defined and would invalidate the collision-minimization claim as stated.","supporting_citations":[{"cited_title":"Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann","cited_arxiv_id":null,"evidence_quote":"Performs open-vocabulary instance segmentation in the 3D scene to ground the target object named in the prompt."},{"cited_title":"High-resolution image synthesis with latent diffusion models","cited_arxiv_id":null,"evidence_quote":"Generates 2D images from the text prompt that are used by the detector for the relative-scaling computation."},{"cited_title":"Instruct-nerf2nerf: Edit- ing 3d scenes with instructions","cited_arxiv_id":null,"evidence_quote":"Serves as the main NeRF-based baseline that requires per-edit training and that FreeEdit compares against qualitatively and quantitatively."}],"review_version":1}