{"id":"2a108e9c-7a6a-4216-bd3f-6882ca9eecfd","arxiv_id":"2505.10827","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"NeuSEditor edits neural implicit 3D surfaces from multi-view images using a text prompt, preserving scene identity via a source-conditioned target renderer and a Phong-shaded distillation loss.","lead":"This paper introduces NeuSEditor, a method that edits 3D scenes captured from multiple photos using text commands while keeping the original scene's identity intact. It combines a three-part neural network (foreground, background, and edited target) with a new geometry-aware training loss to improve output quality.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Phong-shaded input to Stable Diffusion is unvalidated and may inject a hidden bias; Table 1 contradicts the claim that LPE improves fidelity.","rationale":"The reader's weakest_assumption is exactly the assumption I would flag: an out-of-distribution repeated-grayscale Phong image fed into Stable Diffusion is the key technical step enabling the claimed geometry-aware distillation. The paper gives no validation of this input modality, and the internal evidence is mixed. The ablation table contradicts the textual claim that LPE improves fidelity, since lpips (Alex) worsens from 0.685 to 0.700 while clip improves only slightly, and only the multi-view consistency metric improves. That makes the central quantitative claim about the loss's benefit shaky. The strongest part of the paper is the architectural contribution: the three-renderer separation with source conditioning is well motivated, and the user study (though lacking significance testing) shows consistent preference for NeuSEditor on most scenes. Those strengths are independent of the LPE term. The right verdict is CONDITIONAL: the architecture and the PDS-based results are plausible, but the core novelty, LPE, needs either a corrected and consistent ablation or a direct validation experiment. If the authors supply a clear ablation isolating the Phong input and showing that it beats alternatives (normal maps, depth maps) on both fidelity and multi-view consistency, the concern would be resolved. The guidance-scale hand-tuning and the baseline finetuning also weaken comparisons slightly, but they are secondary. I agree with the reader's overall assessment and their choice of the weakest assumption.","tokens_in":17911,"tokens_out":1859,"duration_ms":16781,"concrete_test":"Run an ablation that keeps the full architecture and LPDS fixed, and compares three variants of the geometry term: (a) the proposed repeated-grayscale Phong input; (b) a normal-map rendering that is also replicated to three channels; and (c) an albedo/depth-based shading input. For each variant, recompute the table 1 metrics on the same DTU scan24 'church' edit and also measure the distribution distance (e.g., FID or KLD over latents) between the encoder features of the Phong images and real images from the training set. If the repeated-grayscale Phong variant gives no better or worse metrics than the two control shading inputs, and if the encoder treats the Phong images as strongly out-of-distribution, then LPE's claimed geometric signal is not established. If instead the Phong variant clearly improves geometry over the controls, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central novelty is the geometry-aware distillation term LPE (Eq. 15), which feeds a repeated-grayscale Phong-shaded target rendering into Stable Diffusion's encoder. The reader identifies this as the weakest assumption, and I agree. Section 4.3 states the Phong image is converted to three channels by repeating it; this is an out-of-distribution input for SD, which was trained on natural images. If the encoder produces unreliable or degenerate latents for such inputs, LPE would not provide meaningful geometric signal. More specifically, the gradient in Eq. 18 uses (z_hat_tgt - z_src) as the error, but if the Phong latent is systematically biased by the grayscale pattern, the optimization could be pulled toward a spurious direction rather than toward better geometry. There is no quantitative experiment in the paper that isolates the effect of the Phong channel conversion, e.g., by comparing against a normal-map shading or an albedo-only rendering, or by measuring whether the latents from repeated-grayscale Phong images are even close to the distribution of natural-image latents. Additionally, the ablation in Table 1 is internally inconsistent: on LPIPS (Alex) the baseline is 0.721, Arch+PDS is 0.685, but Ours is 0.700, meaning adding LPE worsens the perceptual distance while the text claims LPE improves fidelity. The only metrics that improve with LPE are multi-view consistency (lpips MV). This does not disprove the utility of LPE for reducing floaters, but it does contradict the stated claim that 'phong-aware distillation enhances fidelity of fine details' in Section 5.2. Thus the central quantitative support for the main contribution is missing or contradictory.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"NeuSEditor proposes a two-stage pipeline for text-guided editing of neural implicit surfaces reconstructed from multi-view images. In the first stage, a source SDF-based renderer learns the identity of the input scene, including a separately modeled background. In the second stage, a target renderer, initialized from and conditioned on the source renderer, is optimized with a posterior distillation loss (PDS) augmented by a geometry-aware loss (LPE) that aligns the stochastic latents of a Phong-shaded target rendering with the source latents. The paper reports quantitative metrics (CLIP, LPIPS), a 41-participant user study over 34 editing experiments on DTU, Blender, and IN2N data, and qualitative comparisons against IN2N, PDS-NeRF, and PDS-Splat, concluding that the method outperforms prior work in identity preservation and geometric consistency.","tokens_in":18202,"tokens_out":2563,"duration_ms":27170,"significance":"If the claims hold, the contribution is practically valuable: it offers single-prompt text-guided editing of neural implicit surfaces without iterative dataset updates, with an architecture that explicitly separates foreground identity, background, and target edits. The user study is substantial (41 participants, 34 experiments, randomized order), and the evaluation spans three benchmark datasets with several scenes each. The method also ships a useful set of ablations, including a comparison of analytical versus numerical SDF gradients. The central novelty, the Phong-enhanced distillation loss, is clearly stated and is potentially interesting to the 3D editing community, but its validation is incomplete, and one ablation result is internally inconsistent.","major_comments":[{"comment":"The ablation table directly contradicts the prose claim that the Phong-aware loss improves fidelity of fine details. On single-view LPIPS, the full method ('Ours') is worse than 'Arch + PDS' for both backbones (Alex: 0.700 vs 0.685; VGG: 0.683 vs 0.676), while only the multi-view consistency metric improves. Since the central claim of the paper is that LPE improves rendering and geometric quality, this inconsistency is load-bearing and needs to be resolved, either by reporting per-scene results, by explaining why single-view perceptual similarity may degrade while the user study still favors the full method, or by softening the claim in §5.2.","section":"§5.2, Table 1"},{"comment":"The geometry-aware term LPE feeds a Phong-shaded grayscale image, repeated across three channels, into the Stable Diffusion encoder. This input is out-of-distribution for a model trained on natural images, and the paper provides no experiment isolating the effect of this channel conversion, no comparison against alternative shading representations (e.g., normal maps or albedo-only renderings), and no measurement that the resulting latents carry geometric information rather than systematic grayscale-pattern bias. If the latents are biased, the gradient in Eq. (18) could pull the optimization in a spurious direction. The authors should add a dedicated validation: for example, measure latent-space distances between Phong-shaded inputs and natural images, or ablate LPE with a different shading input to show the geometric signal comes from the shading itself.","section":"§4.3, Eqs. (13)–(18)"},{"comment":"The identity-preservation metric is the LPIPS distance between the edited rendering and a reference rendering of the unedited scene produced by a separately trained NeRF. This conflates the intended magnitude of the edit with catastrophic forgetting: a method that successfully makes a drastic edit will necessarily have a large LPIPS distance even if no scene content is forgotten. The claim that 'our method preserves the original scene significantly better than competing approaches' therefore needs a more controlled evaluation, such as masking the foreground/background or measuring identity-related features in regions that should remain unchanged.","section":"§5.3.1, Table 2"}],"minor_comments":[{"comment":"There is a typo: 'editting' should be 'editing'.","section":"§5.3.1"},{"comment":"The labels are inconsistent: Table 2 uses 'PDSGS', while the text and Tables 3–4 use 'PDSSplat' for the same method. Please unify the notation.","section":"Table 2 and §5.3"},{"comment":"The user study reports mean scores but no statistical significance tests, confidence intervals, or inter-rater variability. Given the strong claims, at least a pairwise significance test would be appropriate.","section":"§5.3.2, Tables 3–4"},{"comment":"The conditioning input to the target geometry network is written as MLPtgt(Gsrc(x), Htgt(x)); the notation is clear, but the figure and text would benefit from stating explicitly whether Gsrc(x) refers to the source SDF value, the concatenated feature vector, or both.","section":"§10, Eq. (22)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript does not mention release of code or data. Given the sensitivity of the results to hyperparameters (e.g., the CFG guidance scale of 350) and the details of the user-study setup, a public implementation and the full set of edited meshes/renders would substantially strengthen reproducibility. The inconsistency in Table 1 and the unvalidated Phong-channel input are the main technical obstacles to acceptance; I believe both are addressable in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a genuinely new combination: a three-renderer architecture (background, source, target) with additive learning for text-guided edits of implicit surfaces, plus a Phong-shaded distillation term that pushes the diffusion prior to attend to geometry. Second, the paper's own ablation (Table 1) does not back the prose claim that the Phong loss 'enhances fidelity of fine details': on LPIPS the full method is worse than the architecture with plain PDS (0.700 vs 0.685 on Alex, 0.683 vs 0.676 on VGG). Only the multi-view consistency metric improves. That is not fatal, but it is an internal inconsistency the authors need to resolve.\n\nThe architecture itself looks sound. Separating background and foreground, conditioning the target renderer on the source, and avoiding dataset updates are all sensible and are shown to help. The user study (41 participants, 34 tasks) is a real step up from anecdotal comparisons, and the supplementary is honest: it describes mode collapse, gradient computation choices, and limitations (high CFG weights, oversaturation). The baselines were finetuned with best-of-selection, which is not ideal but at least disclosed.\n\nThe soft spots, in rough order of severity:\n\n- The Phong-shaded input to Stable Diffusion is repeated grayscale, which is out-of-distribution. The stress-test worry is legitimate: if the VAE encoder maps these images to biased latents, the gradient in Eq. 18 could pull geometry in a spurious direction. The authors provide no experiment that isolates this (e.g., albedo-only shading or normal-map shading, or a distributional check of the latents). The claim that this is 'geometry-aware' rests entirely on that assumption.\n\n- Table 1's inconsistency is a genuine flaw. The authors should either re-analyze the data, soften the claim, or explain why LPIPS to the original scene is the wrong measure for 'fine detail fidelity' while still using it elsewhere as an identity-preservation metric.\n\n- The user study has no error bars or significance tests. The per-scene numbers in Table 4 are plausible but we can't tell how much noise is in them.\n\n- Guidance scale is hand-tuned per prompt (350 generally, lower for minor edits). That's a hyperparameter, not a fatal flaw, but it weakens the 'prompt-free' simplicity claim.\n\nCitation pattern looks fine; the PDS/DDS lineage is properly acknowledged. No code release is a minor negative.\n\nBottom line: this is a reasonable empirical contribution that likely improves on prior art, and the architecture idea is worth publishing. But the main novel loss term needs either rigorous validation or a much more careful claim. I'd send it to review—it deserves referee time—but I'd expect a revised version that addresses the Phong-channel issue and fixes the ablation narrative.","headline":"Solid incremental extension of PDS with a new architecture and a Phong-shaded distillation loss, but the ablation table undercuts the paper's headline claim for that loss.","tokens_in":18820,"tokens_out":3255,"would_cite":false,"duration_ms":33461,"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":"NeuSEditor shows that text-guided editing of neural implicit surfaces preserves identity and geometry when the network separates background, source, and target renderers and adds a Phong-shaded posterior distillation loss.","keywords":["text-guided 3D editing","neural implicit surfaces","signed distance functions","posterior distillation sampling","score distillation sampling","identity preservation","multi-view reconstruction","Phong shading"],"falsifier":"Replace the Phong-shaded target render in $\\mathcal{L}_{\\mathrm{PE}}$ with a constant gray image of the same shape and rerun the ablation; if the multi-view LPIPS and floater counts stay the same, the geometric signal in the loss is not doing the claimed work. A complementary check is to render a set of shapes with Phong shading, push them through the diffusion encoder, and see whether latent-space distances between renderings correlate with surface-normal difference rather than with albedo difference.","tokens_in":17647,"feed_emoji":"✂️","tokens_out":11198,"duration_ms":98809,"temperature":0.7,"pith_summary":"NeuSEditor claims that text-guided editing of a neural implicit 3D scene can preserve the original scene's identity if the network separates the scene into background, source foreground, and target foreground renderers, with the target initialized from and conditioned on the source. This makes the edit a learned delta layered on top of the original geometry instead of a fresh optimization that can overwrite it. The paper adds a geometry-aware distillation loss that aligns diffusion latents of the source image with latents of a Phong-shaded target render, pushing gradient signal into surface geometry rather than appearance alone. The claimed consequence is a simpler workflow: no iterative dataset updates and no source prompt, with better identity preservation, fewer floaters, and superior CLIP/LPIPS and user-study results on DTU, Blender, and IN2N data.","feed_headline":"3D edits keep identity by splitting source, target, background","feed_subtitle":"A three-renderer net plus Phong-shaded distillation staves off catastrophic forgetting without dataset updates.","key_machinery":"The load-bearing mechanism is the three-renderer identity-preserving architecture together with the geometry-aware distillation loss. A background renderer uses an inverted-sphere density field; a source foreground renderer learns an SDF and the appearance of the original object; a target foreground renderer, initialized from and conditioned on the source, learns the edit. The Phong-aware loss term $\\mathcal{L}_{\\mathrm{PE}}$ computes stochastic latents (defined via the DDPM reverse-step mean, Eq. 9) for the source image and for a Phong-shaded version of the target surface, repeats the shading across three channels to fit the diffusion model input, and minimizes their squared difference; gradients are taken only through the renderer, omitting the diffusion U-Net Jacobian as in SDS. This is what carries the claimed geometric fidelity: the diffusion prior sees shading that varies with surface normal, so the gradient signal is tied to geometry rather than appearance alone.","core_discovery":"The central discovery is that identity preservation in text-guided implicit-surface editing is an architectural problem, not just a loss-tuning problem. The paper separates the scene into a background renderer (density field on an inverted sphere) and source and target foreground renderers (SDF-based, NeuS volume rendering), with the target conditioned on the source's geometry and feature outputs. On top of this, it defines a Phong-enhanced posterior distillation sampling objective, $\\mathcal{L}_{\\mathrm{PEPDS}} = \\lambda_{\\mathrm{PDS}}\\mathcal{L}_{\\mathrm{PDS}} + \\lambda_{\\mathrm{PE}}\\mathcal{L}_{\\mathrm{PE}}$, where $\\mathcal{L}_{\\mathrm{PDS}}$ aligns stochastic latents of source and target renderings and $\\mathcal{L}_{\\mathrm{PE}}$ aligns the source latent with the latent of a Phong-shaded target rendering computed from numerical SDF gradients. The paper argues that this combination keeps recognizable scene features, reduces floaters, and outperforms PDS and Instruct-NeRF2NeRF quantitatively and in a 41-participant user study across DTU, Blender, and IN2N datasets.","pith_inferences":["If the architecture is the active ingredient in identity preservation, this three-way split should transfer to explicit representations such as Gaussian splatting: a source-conditioned target set of splats plus a normal-based shading loss could reproduce the identity gains without SDF rendering or marching cubes.","The Phong-latent alignment rests on the diffusion encoder accepting repeated-channel shading as a real image; a direct test would swap the Phong input for a depth or normal map rendered in the same three-channel format and see whether the floater reduction persists, revealing whether the signal is geometric or merely a second, differently textured view.","Because the head-to-head comparison gives PDS two prompts and NeuSEditor one, the reported gains may mix architectural benefit with prompt-count benefit; running the proposed architecture with the original two-prompt PDS loss would decompose the two.","The additive parameter budget suggests a continual-editing regime in which edits compose in order; one could check whether 'make it a church' followed by 'make it snow' yields the same result as a single 'snowy church' prompt without retraining the identity."],"forward_implications":["Edits no longer require iterative dataset updates or a source text prompt; the identity is held in the source and background renderers while the target renderer learns only the change.","Foreground edits leave the background untouched because the background renderer is separate, and the paper also shows the separation works in reverse, letting background-only edits preserve the foreground.","The per-edit parameter cost is small and additive, about 14M parameters per edit on top of a 28M identity, so successive edits can be layered without retraining the identity.","Measured on 34 tasks across DTU, Blender, and IN2N data, identity preservation (LPIPS against a separately trained NeRF) and user preference improve over PDS and Instruct-NeRF2NeRF, with the largest gains on DTU and Blender.","Numerical SDF gradients used in the Phong shading produce cleaner meshes and fewer floaters than analytical gradients at similar CLIP alignment with the text prompt."],"supporting_citations":[{"why":"It supplies the main baseline (Instruct-NeRF2NeRF) and the IN2N dataset, and its iterative dataset-update workflow is what NeuSEditor claims to eliminate.","marker":"[15]"},{"why":"It defines posterior distillation sampling, the baseline loss and NeRF comparison that the Phong-aware term extends.","marker":"[25]"},{"why":"It supplies progressive hash encoding for SDF geometry, which stage one uses to learn scene identity.","marker":"[28]"},{"why":"It provides the NeuS volume-rendering equations used by both the source and target foreground renderers.","marker":"[58]"},{"why":"It introduces SDS and the gradient rule that omits the U-Net Jacobian, which both distillation terms adopt.","marker":"[40]"},{"why":"It grounds the additive-learning strategy used to initialize the target renderer from the source.","marker":"[19]"},{"why":"It motivates treating stochastic posterior latents as structural signals worth aligning between source and target.","marker":"[21]"},{"why":"It supplies the inverted-sphere parameterization for the background density-field renderer.","marker":"[68]"}],"fun_headline_variants":["NeuSEditor: split-scene neural edits preserve identity","Three renderers keep 3D identity in text-guided editing","Background-foreground split stops identity drift in 3D edits","Phong-shaded distillation keeps surface edits on point","Geometry-aware loss beats PDS and IN2N in 3D editing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the diffusion model treats a Phong-shaded, three-channel copy of the target surface as a legitimate image, so the noise-recovery latent computed from it carries real geometric information that can steer the edit; if that input is out-of-distribution, the geometry-aware loss contributes no useful gradient and the architecture alone must explain the results.","fun_headline_variants_meta":{"raw":{"variants":["NeuSEditor: split-scene neural edits preserve identity","Three renderers keep 3D identity in text-guided editing","Background-foreground split stops identity drift in 3D edits","Phong-shaded distillation keeps surface edits on point","Geometry-aware loss beats PDS and IN2N in 3D editing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1400,"prompt_tokens":918,"completion_tokens":482,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":394}},"tokens_in":534,"tokens_out":482,"duration_ms":4847,"temperature":1.0,"reasoning_tokens":394,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:02:36.145711+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the Phong-shaded target render in $\\mathcal{L}_{\\mathrm{PE}}$ with a constant gray image of the same shape and rerun the ablation; if the multi-view LPIPS and floater counts stay the same, the geometric signal in the loss is not doing the claimed work. A complementary check is to render a set of shapes with Phong shading, push them through the diffusion encoder, and see whether latent-space distances between renderings correlate with surface-normal difference rather than with albedo difference.","supporting_citations":[{"cited_title":"Instruct-NeRF2NeRF: Editing 3D scenes with instructions","cited_arxiv_id":null,"evidence_quote":"It supplies the main baseline (Instruct-NeRF2NeRF) and the IN2N dataset, and its iterative dataset-update workflow is what NeuSEditor claims to eliminate."},{"cited_title":"Posterior distil- lation sampling","cited_arxiv_id":null,"evidence_quote":"It defines posterior distillation sampling, the baseline loss and NeRF comparison that the Phong-aware term extends."},{"cited_title":"Neuralangelo: High-fidelity neural surface reconstruction","cited_arxiv_id":null,"evidence_quote":"It supplies progressive hash encoding for SDF geometry, which stage one uses to learn scene identity."},{"cited_title":"Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction","cited_arxiv_id":null,"evidence_quote":"It provides the NeuS volume-rendering equations used by both the source and target foreground renderers."},{"cited_title":"Barron, and Ben Milden- hall","cited_arxiv_id":null,"evidence_quote":"It introduces SDS and the gradient rule that omits the U-Net Jacobian, which both distillation terms adopt."},{"cited_title":"LoRA: Low-rank adaptation of large language models","cited_arxiv_id":null,"evidence_quote":"It grounds the additive-learning strategy used to initialize the target renderer from the source."},{"cited_title":"An edit friendly DDPM noise space: Inversion and manipulations","cited_arxiv_id":null,"evidence_quote":"It motivates treating stochastic posterior latents as structural signals worth aligning between source and target."}],"review_version":1}