{"id":"bdef4523-3628-4efb-94e9-4e9d3c90b6c7","arxiv_id":"2506.10507","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A training-free method that propagates a 2D edit applied at any chosen viewpoint across a full 360-degree orbit by fusing anchor-view and front-view video-diffusion trajectories.","lead":"Edit360 extends 2D image edits into consistent 3D-model edits from any chosen viewing angle without retraining. For anyone making 3D content, this could turn a simple image edit into a full 360-degree asset update, removing the need to sculpt or repaint the model manually.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The circular-shift alignment in Eq. (4) assumes the anchor-conditioned V3DM trajectory follows exactly the same camera orbit as the front trajectory; this is unsupported for arbitrary edited anchor views, so SPF and CVA may be fusing geometrically misaligned views.","rationale":"The reader's conditional verdict is appropriate, and this stress-test identifies a specific, load-bearing assumption that supports the conditionality. The paper's key novelty is the dual-trajectory fusion: an unedited front-view trajectory and an edited anchor-view trajectory are aligned by Eq. (4) and merged by Eqs. (5) and (6). For that fusion to preserve both identity and the edit, the two trajectories must be in the same camera coordinate system after the circular shift. The paper does not establish this. The reader's weakest assumption focuses on the 2D editor preserving the object's geometry; the concern here is broader and remains even if the 2D editor is perfect: the V3DM, when conditioned on an edited side or back image, has no explicit pose input, so the generated anchor trajectory may not share the front trajectory's camera path. This is a concrete empirical premise, not internal inconsistency: the method could work if the V3DM happens to be sufficiently pose-consistent, but the paper provides no quantitative test in the edited-anchor regime. The quantitative NVS results (Table 2) are a useful sanity check, but they use an unedited anchor generated by the same V3DM, so they do not exercise the arbitrary-viewpoint editing scenario. The user study is qualitative and lacks edit-fidelity metrics. Given these gaps, the claim should remain conditional pending a direct test of geometric alignment with edited anchors; the verdict does not need to change from the reader's CONDITIONAL to a stronger or weaker one. Independent support exists in the clean training-free design and the improved NVS numbers, but it does not cover the most load-bearing part of the central claim.","tokens_in":12994,"tokens_out":5987,"duration_ms":75371,"concrete_test":"Take 20 synthetic 3D objects with ground-truth meshes. For each, render a front view and an anchor view from a specified azimuth (e.g., 135 degrees), apply a known local edit (e.g., a colored sphere at a known 3D coordinate) to the anchor view using a 2D editor, and run the full Edit360 pipeline. Then (1) estimate the camera pose of every generated frame (e.g., with COLMAP or a pose regressor) and compute the angular error between each circular-shifted anchor frame and the corresponding front frame; if the mean error is more than a few degrees, Eq. (4) is not a valid alignment. (2) Reconstruct the edited 3D asset and measure whether the added sphere appears near its true 3D location (e.g., chamfer distance or 3D detection within tolerance). This distinguishes geometric misalignment from 2D-editor fidelity and directly tests the any-viewpoint claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Edit360 supports edits from arbitrary viewpoints with structural coherence rests entirely on Anchor-View Editing Propagation. In SPF, Eq. (4) aligns the anchor trajectory to the front trajectory with a cyclic index shift by p, and Eq. (5) blends their latents. This is correct only if frame j of X_anchor is at the same camera pose as frame j of X_front for every j. That condition is not established. V3DMs such as SV3D and V3D take a single image as conditioning and do not accept a camera-pose label; they are trained to generate an orbit from a front-view input. When the conditioning image is an edited side or back view, the model has no way to know the azimuth of that view, so the generated orbit may start from a different orientation, elevation, or scale than the front trajectory. The circular-shift operation then aligns frames that do not correspond to the same 3D points. In that case Eq. (5) blends latents from different camera positions, producing ghosting or over-smoothing, and CVA (Eq. (6), which injects front-trajectory key/value features into the anchor trajectory) can pull the edited trajectory toward the wrong geometry, suppressing the edit itself. The quantitative evaluation does not test this regime: Table 2 uses an anchor frame generated by the same V3DM from the front view, so both trajectories share the model's implicit front-view assumption; it does not validate edited anchors from arbitrary viewpoints. The user study (Table 1) is subjective and contains no edit-fidelity or geometric-consistency metric. Thus the paper's main claim is supported only by qualitative examples in the regime where the alignment assumption is most questionable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Edit360, a training-free framework for propagating a 2D edit made on an arbitrary 'anchor view' of an object into a full 360-degree multi-view sequence generated by a video 3D diffusion model (V3DM) such as SV3D or V3D. The core mechanism is Anchor-View Editing Propagation, consisting of Spatial Progressive Fusion (SPF), which circularly shifts the anchor-conditioned trajectory and blends it with the front-view trajectory using a distance-based weight, and Cross-View Alignment (CVA), which injects front-trajectory key/value features into the anchor trajectory's attention layers. The edited dense-view sequence is then used to reconstruct a 3D asset via NeuS or 3DGS. The paper reports quantitative novel-view synthesis results on GSO and OmniObject3D (Table 2), an ablation (Table 3), a 50-participant user study against Tailor3D (Table 1), and qualitative examples of local edits and style transfer.","tokens_in":13326,"tokens_out":6238,"duration_ms":64296,"significance":"If the claims are correct, Edit360 would be a practical, tuning-free way to extend 2D image editing to 3D assets with edits specified from any viewpoint. The proposed dual-trajectory fusion is simple and novel, and the quantitative NVS results (Table 2: Edit-SV3D 22.17 PSNR vs. 21.14 for SV3D) show consistent gains over strong baselines. The paper is also honest in reporting an upper-bound experiment with two ground-truth views. However, the main load-bearing claim—editing from arbitrary viewpoints with structural coherence—lacks direct quantitative validation: the user study has no inferential statistics, the NVS evaluation uses unedited anchor views generated by the base model, and the alignment assumption behind Eq. (4) is not tested.","major_comments":[{"comment":"The circular-shift alignment in Eq. (4) assumes that the anchor-conditioned V3DM trajectory follows exactly the same camera orbit as the front trajectory, so that frame i of CS(X_anchor) matches frame i of X_front in camera pose. V3DMs such as SV3D and V3D take a single image as conditioning and do not accept an explicit camera-pose label; when the conditioning image is an edited side or back view, the model has no way to know the azimuth or elevation of that view. The paper provides no experiment validating this alignment for edited anchors from arbitrary viewpoints. Table 2 uses anchor views generated by the same V3DM from the front view, so both trajectories share the model's front-view assumption; this does not test the claimed regime. Because Eq. (5) blends latents from these aligned frames, any misalignment directly propagates into the fused sequence and into the reconstructed 3D asset.","section":"Section 4.1, Eq. (4)"},{"comment":"The spatial weight alpha_i, which is central to SPF, is never defined. The text only states that it decreases with cyclic distance to the anchor view; no formula, schedule, or default value is given. Likewise, 'In later diffusion stages, SPF integrates edge and texture information' is not specified in terms of timestep ranges or integration operations. These omissions make the method impossible to reproduce exactly and undermine the ablation in Table 3, since the baseline and ablations depend on unspecified choices.","section":"Section 4.1, Eq. (5)"},{"comment":"The user study is the principal quantitative evidence for the editing claim, but the paper reports only mean Likert scores (4.52 vs. 3.02 overall) without standard deviations, confidence intervals, or any inferential test. With 50 participants and 8 examples, the reported differences could be within noise; no paired t-test, Wilcoxon signed-rank test, or inter-rater agreement is reported. The paper also does not describe how participants were recruited, whether the comparison was blinded, or whether the same 3D outputs were shown for both methods. Without this information, the 'clearly demonstrate the superiority' statement is not supported.","section":"Section 5.2, Table 1"},{"comment":"The quantitative NVS evaluation does not actually evaluate the editing pipeline. In the 'Edit-SV3D (v0&vi)' rows, the additional anchor view is generated by the base V3DM from the original front view (Sec. 5.3: 'an additional anchor view v_i for Edit360 is generated using the original single front-view input V3DM model'), so the anchor is geometrically consistent with the front view by construction. The 2GT row (26.32 PSNR) shows that much of the gain comes from simply conditioning on an extra view, not from propagating an edit. The paper therefore does not provide a quantitative test of the claim that edits from arbitrary viewpoints are propagated with structural coherence; a direct evaluation would require, e.g., editing the anchor with a known ground-truth 3D edit and measuring the correspondence of the edited region across views.","section":"Section 5.3, Table 2"},{"comment":"The pipeline assumes that an off-the-shelf 2D editor (ControlNet/InstantStyle) produces an anchor image that is geometrically faithful to the original 3D object. The text states 'After editing the anchor view' without any investigation of whether the editor preserves pose, proportions, lighting, and camera intrinsics. Diffusion-based editors often alter these attributes; if they do, the fusion in Eq. (5) blends misaligned geometry and CVA may suppress the edit itself. The manuscript provides no failure analysis or quantitative measure of edit fidelity (e.g., CLIP directional score), so the identity-preservation promise is not demonstrated.","section":"Section 3.2"}],"minor_comments":[{"comment":"The caption cites 'Tailor3D [48]', but the Tailor3D reference is [34] in the reference list; [48] is SV3D.","section":"Figure 7 caption"},{"comment":"The notation 'v0&vi' in Table 2 is not defined in the table; the text defines it only later in Sec. 5.3. Please move the definition to the table caption or introduce it before the table.","section":"Section 5.3, Table 2"},{"comment":"The description 'filtering out 300 objects with significant shape differences' is vague; please specify the filtering criterion (e.g., human inspection, a similarity metric, or a fixed procedure) so the evaluation can be replicated.","section":"Section 5.1"},{"comment":"The term 'Dual-Stream Diffusion Network' appears only in the Figure 3 caption and not in the body text. Please define it in the text or remove it from the caption.","section":"Section 3.2"},{"comment":"The phrase 'from any angle' is stronger than what is evaluated, since all experiments use 0-degree elevation orbits. Please qualify the claim to horizontal orbits or add experiments with varying elevations.","section":"Abstract and Section 1"},{"comment":"The user study description does not say whether participants saw the two methods' outputs side-by-side or sequentially, or whether the order was randomized. Please provide these details.","section":"Section 5.2"},{"comment":"Reference [35] is cited for DALL·E in Figure 5, but the title 'Text to Image Generator with Latent Diffusion Models' does not match DALL·E. Please clarify the actual model used.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable system paper, but the evaluation of the central editing claim is substantially weaker than the NVS evaluation. If the editors value the editing contribution, I would ask for the additional experiments described in the major comments. Also, the paper does not release code or data, which limits reproducibility. The novelty is modest given the concurrent Tailor3D, but the arbitrary-viewpoint capability is a differentiator."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe dual-trajectory fusion idea is real and the ablations are clean, but the headline claim—edits from arbitrary viewpoints—is only backed by a small user study. The quantitative evaluation never exercises the edited-anchor regime.\n\nThe method runs two denoising trajectories through a video 3D diffusion model: one conditioned on the original front view, one on a user-edited anchor view. They're aligned with a circular shift, blended with proximity weights, and aligned via cross-view attention. That's a genuine extension over Tailor3D's fixed front/back views, and the paper is clearly written. SPF and CVA both help in the ablation study, and the NVS numbers in Table 2 are modest but consistent gains over SV3D/V3D.\n\nThe problem is the evidence for the core claim. In Table 2, the extra anchor view is generated by the same V3DM from the front view, so both trajectories share the model's implicit front-view assumption. The circular shift in Eq. (4) assumes frame j of the anchor trajectory is at the same camera pose as frame j of the front trajectory. That's true by construction when the anchor is generated, but not when it's an off-the-shelf 2D edit of a back or side view. V3DMs don't accept camera-pose labels; an edited back view gets treated as a front view, the orbit starts from the model's canonical orientation, and the shift will likely misalign corresponding frames. The stress-test note hits the mark. The only direct evidence for the arbitrary-view claim is a 50-participant user study with no statistical tests, no error bars, and no edit-fidelity or geometric-consistency metrics. That's thin.\n\nOther soft spots: no code release, and implementation details are missing—the alpha_i schedule, timesteps for the late-stage edge/texture integration, and how CVA is applied across attention layers. These are fixable, not fatal.\n\nWho's this for? Researchers working on 3D generation and editing with video diffusion models. It deserves a serious referee: the idea is novel, the writing is honest, and the main failure mode is a missing experiment rather than a contradiction. I'd send it to review, but ask for code, implementation details, and a quantitative evaluation that actually uses edited anchors at non-front angles, ideally against ground-truth multi-view data so you can measure edit propagation and ghosting.\n\nVerdict: conditional, leaning accept once the arbitrary-view evidence is added.","headline":"New propagation mechanism, but the arbitrary-view editing claim is under-tested; the paper needs code and a proper evaluation with edited anchors.","tokens_in":13906,"tokens_out":6650,"would_cite":true,"duration_ms":60691,"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":"Edit360 claims that one 2D edit from any chosen viewpoint can be propagated, without training, into a multi-view-consistent 3D asset that preserves the object's identity.","keywords":["3D asset editing","video diffusion models","multi-view consistency","tuning-free framework","anchor view editing propagation","spatial progressive fusion","cross-view alignment","novel view synthesis"],"falsifier":"Take an edit that deliberately shifts an object's position or changes its aspect ratio in the anchor view, run Edit360, and render the reconstructed asset from several angles: ghosting, duplicated edges, or identity drift beyond the edit itself would show that the alignment assumption fails.","tokens_in":12775,"feed_emoji":"🎨","tokens_out":7957,"duration_ms":84718,"temperature":0.7,"pith_summary":"Edit360 is trying to show that a user can edit a 3D asset by making a single 2D change in one image, taken from whichever angle best reveals the change, and get back a full 3D asset in which that change is visible and consistent from every viewpoint. The proposed framework is tuning-free: it reuses a pre-trained video 3D diffusion model and needs no fine-tuning. The paper argues that existing approaches either force edits to be made on the front view, which misses details visible only from behind, or lose identity information when another edited view replaces the input. If the claims hold, creators can customize 3D content for games, animation, and virtual reality with the same ease as editing a photo.","feed_headline":"A single edited image becomes a coherent 3D asset from all angles","feed_subtitle":"No training required: edit the view that best shows the change, and every other view updates consistently.","key_machinery":"The load-bearing component is Anchor-View Editing Propagation, a dual-stream denoising mechanism built on a video 3D diffusion model. It consists of Spatial Progressive Fusion (SPF) and Cross-View Alignment (CVA): SPF circular-shifts the anchor-view trajectory into the coordinate frame of the front-view trajectory and blends the two streams with a proximity-based weight, so edits dominate near the anchor and fade smoothly away; CVA concatenates key/value features from the front-view denoising stream into the anchor-view self-attention, forcing shared structure and preventing ghosting where the two streams disagree. Together they propagate the 2D edit while holding the object's identity in place.","core_discovery":"The central claim is that a single edited anchor view can drive multi-view-consistent editing across a full 360-degree orbit. A second camera trajectory starts at the anchor view, where the user's 2D edit has been applied, while the original front-view trajectory is kept unedited as an identity reference. At every denoising step the two trajectories are spatially aligned by a circular shift, fused with weights that decay as views move away from the anchor, and then reconciled by injecting the front-view stream's attention keys and values into the anchor-view stream. The result, the paper argues, is a dense sequence of edited views that reconstructs a high-quality 3D asset while preserving identity and geometric structure, enabling local element edits, global style transfer, and multi-view conditional generation.","pith_inferences":["Editorial extension: if the alignment assumption holds, the same dual-stream propagation could chain several anchor edits in one pass, replacing multiple sequential edits with a single fused generation.","Editorial extension: the circular-shift alignment ties the method to a fixed-elevation orbit; extending it to arbitrary camera paths or changing elevation would require a more general spatial alignment than cyclic shift.","Editorial extension: the paper reports a two-ground-truth-view upper bound, so a natural next test is whether automatically selecting non-front anchors, rather than random ones, closes more of that gap."],"forward_implications":["Users can specify the angle that best shows the change, or let a language model pick it, and the edit carries across all views of the reconstructed asset.","Edits that are only visible from the back or side, such as wings, capes, or rear decorations, no longer require the front view to show them.","Because propagation happens during denoising rather than through training, the mechanism can be dropped onto different video 3D diffusion models without retraining.","Reconstruction from the edited multi-view sequence works with either neural-surface or Gaussian-splatting backends, so the output can feed standard 3D pipelines.","Supplying additional ground-truth views as anchors raises reconstruction quality, indicating that the framework scales with the number of anchor views."],"supporting_citations":[{"why":"It supplies the base video 3D diffusion model that Edit360 extends and the strongest single-view baseline it compares against.","marker":"[48]"},{"why":"It provides a second video 3D diffusion base model, showing that the propagation mechanism is not tied to one generator.","marker":"[6]"},{"why":"It is the concurrent 3D editing method restricted to front and back views, and the user-study comparison against it grounds the arbitrary-viewpoint claim.","marker":"[34]"},{"why":"It supplies the off-the-shelf style-preserving 2D editor used to modify the anchor view.","marker":"[49]"},{"why":"It supplies the off-the-shelf conditional 2D editor used for controlled anchor-view edits.","marker":"[60]"},{"why":"It provides the neural-surface reconstruction that turns the edited multi-view sequence into the output 3D asset.","marker":"[50]"},{"why":"It provides the Gaussian-splatting reconstruction used as the alternative 3D output representation.","marker":"[19]"},{"why":"It is the household-object dataset on which the quantitative novel-view synthesis numbers are measured.","marker":"[10]"}],"fun_headline_variants":["Edit one view, get full 360° 3D consistency","One 2D edit, coherent 3D object from all angles","Single-view edit propagates to all angles in 3D","From one anchor view to a full 360° 3D asset","Anchor-view edit yields multi-view consistent 3D"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the 2D editor changes only the edited region and leaves the object's pose, scale, and framing in the anchor image exactly as they were, so the two view sequences line up point-by-point; if the edit warps or moves the object, the fusion will blend mismatched views.","fun_headline_variants_meta":{"raw":{"variants":["Edit one view, get full 360° 3D consistency","One 2D edit, coherent 3D object from all angles","Single-view edit propagates to all angles in 3D","From one anchor view to a full 360° 3D asset","Anchor-view edit yields multi-view consistent 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000627,"raw_usage":{"total_tokens":2857,"prompt_tokens":856,"completion_tokens":2001,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":472,"completion_tokens_details":{"reasoning_tokens":1912}},"tokens_in":472,"tokens_out":2001,"duration_ms":14770,"temperature":1.0,"reasoning_tokens":1912,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:24:22.882428+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an edit that deliberately shifts an object's position or changes its aspect ratio in the anchor view, run Edit360, and render the reconstructed asset from several angles: ghosting, duplicated edges, or identity drift beyond the edit itself would show that the alignment assumption fails.","supporting_citations":[{"cited_title":"Sv3d: Novel multi-view syn- thesis and 3d generation from a single image using latent video diffusion","cited_arxiv_id":null,"evidence_quote":"It supplies the base video 3D diffusion model that Edit360 extends and the strongest single-view baseline it compares against."},{"cited_title":"Adding Conditional Control to Text-to-Image Diffusion Models","cited_arxiv_id":null,"evidence_quote":"It supplies the off-the-shelf conditional 2D editor used for controlled anchor-view edits."},{"cited_title":"NeuS: Learning Neural Im- plicit Surfaces by V olume Rendering for Multi-view Recon- struction","cited_arxiv_id":null,"evidence_quote":"It provides the neural-surface reconstruction that turns the edited multi-view sequence into the output 3D asset."},{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"It provides the Gaussian-splatting reconstruction used as the alternative 3D output representation."},{"cited_title":"Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items","cited_arxiv_id":null,"evidence_quote":"It is the household-object dataset on which the quantitative novel-view synthesis numbers are measured."}],"review_version":1}