{"id":"e17c24f7-a6a5-41be-a186-34eb112a09dd","arxiv_id":"2412.05279","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Perturb-and-Revise edits 3D scenes by mixing a NeRF's trained parameters with random ones, running multi-view score distillation toward the edit prompt, and refining with identity-preserving gradients.","lead":"This paper introduces a method for editing 3D scenes with text prompts, including big changes like new poses, added objects, or clothes. It works by partially randomizing the scene's internal parameters, re-optimizing with a text-guided image model, then pulling the result back toward the original so the subject stays recognizable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central mechanism, Eq. (3), is only justified distributionally in Supp.","rationale":"The reader's weakest-assumption analysis correctly identifies the load-bearing point: the paper's theoretical support for parameter perturbation (Supp. B) only proves a distributional interpolation result and does not establish that linear interpolation in raw InstantNGP parameter space produces a controlled semantic reversion of the scene. The central claim that perturbation 'enables the particles to escape local minima and follow a natural generative path' requires exactly this semantic-reversion property. Without it, the method's advantage over the η=0 baseline and over full regeneration is not explained by the proposed mechanism, and the novelty of the contribution is weakened. I agree with the conditional verdict rather than moving to reject, because the qualitative demonstrations and ablations provide empirical support for the practical claim that perturbation helps; the concern is about the mechanism and generalizability, not about whether the reported results were obtained. The concrete test I propose would directly probe whether perturbed renders are coherent partial reverses of the source, and whether the editing outcome is stable across random draws. If the test shows coherent degradation and stable success, the concern is resolved. If it shows incoherence or high draw-to-draw variance, the paper's central mechanism would need to be reframed as 'regeneration with source anchoring,' which is a materially weaker claim. The reader's original verdict of CONDITIONAL is appropriate, and this stress-test does not change it.","tokens_in":17415,"tokens_out":5430,"duration_ms":61374,"concrete_test":"Take a set of source NeRFs from the paper's fashion and Objaverse evaluations. For η ∈ {0, 0.2, 0.4, 0.6, 0.8}, render θ_perturbed before any editing and measure: (i) presence/pose of the expected object via a keypoint or detector model, (ii) occupancy and density statistics, and (iii) CLIP similarity of the render to the source caption and to a random-init render. If η=0.6 renders are no more recognizable than random-init renders, or if semantic attributes vanish abruptly rather than degrading smoothly, the 'partially undoing optimization' premise fails. Then run the full edit pipeline for a fixed η=0.6 with 10 independent θ0 draws per prompt and record success rates; high variance across draws would show that editing success depends on accidental properties of the random interpolation, not on a reliable generative trajectory.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that parameter perturbation θ_perturbed = (1−η)·θ_src + η·θ_0 (Eq. 3) lets the multi-view score-distillation ODE (Eq. 4) follow a natural generative trajectory from a partially-undone optimization state, enabling pose changes and object addition that no-perturbation editing cannot achieve. The only formal support, Supp. B, proves that the distribution of θ_perturbed interpolates between a point mass at θ_src and the initialization distribution P(Θ0). That is a statement about ensembles, not about individual parameter vectors: it does not show that a specific interpolated InstantNGP hash grid or MLP corresponds to a semantically coherent, mildly-degraded version of the source scene. Raw parameter space is not organized by spatial frequency or semantic scale, so linear interpolation may instead produce foggy or mixed geometry that the diffusion model must completely regenerate. If so, the method's observed success is not evidence for the proposed 'generative trajectory' mechanism; it would be evidence for regeneration from a nearly-random start followed by identity-preserving refinement. This distinction is load-bearing because the paper explicitly contrasts perturbation with both η=0 and full regeneration, and the adaptive η selection (Sec. 4.2) is motivated by the basin-escaping trajectory story. Without evidence that interpolated parameters degrade smoothly and semantically, the mechanism remains an unsupported modeling assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Perturb-and-Revise (PnR), a training-free method for text-guided editing of NeRF scenes. The method first perturbs the optimized NeRF parameters by linearly interpolating them with a random initialization (Eq. 3), with the perturbation amount eta chosen adaptively via a loss-landscape heuristic (Sec. 4.2). It then runs a multi-view score-distillation ODE (Eq. 4) with timestep annealing, and finally applies identity-preserving gradients (IPG) to refine the result toward the source. The authors evaluate on synthetic fashion objects and Objaverse objects, comparing against SDS, PDS, and Instruct-NeRF2NeRF, with additional qualitative comparisons to Instruct-GS2GS and MVDream regeneration. The central claims are that parameter perturbation enables geometric edits (pose changes, object addition) that prior methods cannot handle, and that the adaptive eta selection and IPG refinement improve the quality and fidelity of the edits.","tokens_in":17769,"tokens_out":5663,"duration_ms":59144,"significance":"If the claimed results hold, the paper proposes a simple and effective mechanism for 3D editing that goes beyond texture/color changes and handles geometric edits such as pose changes and object addition. The idea of hijacking the generative ODE trajectory by parameter-space perturbation is original and conceptually appealing. The paper provides extensive qualitative results, an ablation on eta selection, and ablations for the IPG component. The main strengths are the breadth of edit types demonstrated and the efficiency of the approach (1.5k editing steps vs. 10k for regeneration). However, the quantitative evaluation is not yet fully convincing: only three quantitative baselines are compared, point estimates are reported without error bars, and the adaptive eta selection is tuned on the same evaluation benchmark, which weakens the claims of optimality and state-of-the-art performance.","major_comments":[{"comment":"The adaptive eta selection relies on two constants, Delta_min and eta_max, which are set based on observations on the same evaluation benchmark that is later used to claim that the adaptive scheme is near-optimal (Fig. 8). Specifically, Supp. C states that Delta_min is chosen from Table 3 (CLIP directional similarity on the same dataset) and eta_max from the success-rate plot in the main paper. This makes the comparison against fixed eta values circular. To support the claim, the authors should either tune on a disjoint development set and evaluate on a separate test set, or provide a sensitivity analysis showing that the final editing metrics are robust to the choice of these constants across a plausible range.","section":"Sec. 4.2 / Supp. C / Table 3"},{"comment":"The theoretical justification for parameter perturbation only establishes that the distribution of theta_perturbed interpolates between a point mass and the initialization distribution. It does not show that a particular interpolated parameter vector corresponds to a semantically coherent, partially 'undone' version of the source scene, which is the premise of the 'generative trajectory' story. The authors should provide direct evidence, for example, renderings of the perturbed NeRF at eta values 0, 0.2, 0.4, 0.6, 0.8 before any editing steps, or a plot of the source-prompt loss along the interpolation path. Such evidence would clarify whether the perturbation produces gradual semantic degradation or an incoherent state that is then regenerated from scratch.","section":"Sec. 4.1 / Supp. B"},{"comment":"The quantitative comparison lacks error bars, significance tests, and multiple seeds. Many of the reported differences are tiny: the average CLIP-Dir-Sim is 0.0565 for PnR versus 0.0557 for Instruct-NeRF2NeRF, and for ViT-B/16 Instruct-NeRF2NeRF is actually higher (0.0549 vs. 0.0534). Without variance estimates, these numbers do not support the claim that PnR 'achieves state-of-the-art results'. In addition, Instruct-GS2GS is included in the qualitative comparison (Fig. 2) but omitted from Table 1. The authors should report standard deviations over multiple runs/seeds and include Instruct-GS2GS in the quantitative evaluation, or moderate the strength of the claim.","section":"Sec. 5.2 / Table 1"},{"comment":"The y-axis of Fig. 8(a) is labeled 'percentage of successful experiments (without errors)', but the text never defines what constitutes a success or an error. This metric is used to justify eta_max=0.6 and to compare adaptive eta with fixed eta, so a precise operational definition (e.g., based on a CLIP threshold, manual inspection, or a convergence criterion) is necessary. Without it, the robustness claim is not fully verifiable.","section":"Sec. 5.4 / Fig. 8"}],"minor_comments":[{"comment":"The figure layout is confusing: the horizontal axis appears to represent optimization steps with irregular spacing (100 to 900) and then 10000 for the regeneration row. Please label the axes clearly and specify whether the perturbation rows share the same step sequence.","section":"Fig. 4"},{"comment":"The statement 'adding more noise to the parameters is equivalent to undoing more of the optimization process' is a key intuition but is presented without support. Either provide a reference or an experiment (e.g., showing that perturbed renderings resemble earlier optimizer states) or soften the wording.","section":"Sec. 4.1"},{"comment":"The asterisk in the figure captions says 'we use an identical update rule and schedule' for PDS and PnR, but the methods differ in parameter perturbation and IPG. Please clarify which components are identical (e.g., the score-distillation backbone and timestep schedule) to avoid confusion.","section":"Sec. 5.1 / Figs. 2, 9"},{"comment":"The values of lambda_L1 and lambda_p are only given in the supplementary material. Consider reporting them in the main text, since they are part of the method definition.","section":"Sec. 4.3 / Supp. E"},{"comment":"The abbreviation CLIP-Dir-Con is not defined in the main text; define it when first used (it is defined in the supplement as CLIP directional consistency).","section":"Table 2"},{"comment":"The phrase 'makes possible a variety of NeRF editing' is awkward; consider 'enables a variety of NeRF edit types' for clarity.","section":"Abstract / conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and relevant problem, and the qualitative results are often impressive. The main concerns are the circularity in the adaptive eta evaluation and the absence of error bars/standard deviations in the quantitative comparison, which weakens the state-of-the-art claim. The mechanism story (generative trajectory via parameter perturbation) is plausible but not yet directly evidenced; the requested renderings of the perturbed scene would substantially increase confidence. The paper's scope and experimental breadth are appropriate for the journal, but the evaluation should be tightened before acceptance. I see no grounds for rejection, as the core idea is novel and the method appears to work well in many cases."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"My take: this paper deserves a serious referee. It introduces a simple and effective mechanism—interpolate a trained NeRF's parameters toward a random initialization, then run multi-view score distillation with an identity-preserving refinement. The qualitative results for pose changes and object addition are visibly beyond what Instruct-NeRF2NeRF and PDS can do, and that is the main claim. Credit where due: the idea is clearly stated, the ablation in Fig. 4 shows the perturbation has a real effect, and the extension to real scenes with PDS is a nice bonus. The supplementary is honest about limitations and about the analogy to SDEdit.\n\nThe theory is thin. Supp. B proves only that the distribution of perturbed parameters interpolates between a point mass at the source and the prior; it says nothing about whether a specific interpolated hash grid corresponds to a semantically meaningful partial undo. That is a fair criticism, but it targets the explanation, not the method. The method works in practice regardless, and the paper would be stronger if it simply framed perturbation as a heuristic initialization rather than as hijacking a 'generative trajectory.'\n\nThe bigger issue is evaluation hygiene. The adaptive eta rule uses Δ_min=1000 and η_max=0.6, and the supplementary says these were set based on observations on the same benchmark. That is fitting the selection rule to the eval set. There are no error bars or significance tests, and the quantitative comparison omits Instruct-GS2GS, which appears in the qualitative table. The pseudocode for parameter perturbation doesn't match Eq. (3): Alg. 1 interpolates between θ_init and θ_new, not between θ_src and θ0. These are fixable in revision, but they make the current numbers less convincing than the qualitative demo.\n\nThis paper is for people working on 3D editing and distillation. It deserves peer review, not desk rejection. If I were the editor, I would send it out, asking reviewers to focus on the eval protocol and to request code and error bars. I would cite it for the perturbation trick, but I would be careful about the theory.","headline":"Parameter-space perturbation is a genuinely useful trick for NeRF editing; the weak theory and eval hygiene don't sink it, but the adaptive-eta rule is fitted to the same benchmark it is evaluated on.","tokens_in":18228,"tokens_out":2300,"would_cite":true,"duration_ms":22860,"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":"Perturb-and-Revise claims that pushing an optimized NeRF partway back to random initialization lets text-driven editing change pose, appearance, and geometry in ways previous methods cannot.","keywords":["neural radiance fields","3D editing","score distillation","parameter perturbation","diffusion models","particle-based variational inference","multi-view consistency","text-guided editing"],"falsifier":"Render the perturbed NeRF at a moderate perturbation level (around $\\eta = 0.3$) before any distillation runs; if the geometry is already unrecognizable or incoherent, the claimed re-entry into a generative trajectory fails. A second check is to compare edits obtained by interpolating toward two different random initializations with the same $\\eta$: if the outputs diverge sharply, the method is sensitive to which random parameters are drawn.","tokens_in":17251,"feed_emoji":"🎨","tokens_out":7823,"duration_ms":71905,"temperature":0.7,"pith_summary":"This paper claims that a text-guided 3D editing system can handle large geometric changes—pose changes, added objects, and shape edits—if it first pushes an optimized neural radiance field (NeRF) partway back toward random initialization. The method, Perturb-and-Revise, interpolates the source NeRF's parameters with a random NeRF's parameters, chooses the interpolation amount from a short analysis of the local loss landscape, and then runs multi-view score distillation with an identity-preserving refinement term. The paper argues that the perturbed parameters leave the source scene's local minimum and follow a natural coarse-to-fine generative trajectory toward the edit prompt, so the final object stays recognizable while undergoing changes that earlier text-based editing methods could not produce.","feed_headline":"Mixing a NeRF with random weights unlocks hard 3D edits","feed_subtitle":"Perturb-and-Revise reuses the generative path of score distillation to change pose and add objects, not just colors.","key_machinery":"The load-bearing object is the edited NeRF treated as a particle in parameter space, moved by the Wasserstein gradient-flow ODE that underlies score distillation. The central intervention is parameter perturbation, a linear interpolation between the optimized parameters and a random initialization, with the interpolation weight $\\eta$ selected adaptively from the measured loss decrease over a few simulated distillation steps. This is what lets the particle escape the source local minimum and re-enter the coarse-to-fine generative path. The second mechanism is the identity-preserving gradient, a weighted combination of L1 and perceptual losses between renders of the current and source NeRF, added in later refinement steps as a tug-of-war term that balances the edit direction against fidelity to the source object.","core_discovery":"On the paper's own terms, the discovery is that a fully optimized NeRF's parameters are not a fixed point for editing: linearly interpolating them toward a random initialization, $$\\theta_{\\mathrm{perturbed}} = (1-\\eta)\\,\\theta_{\\mathrm{src}} + \\eta\\,\\theta_0, \\quad \\theta_0 \\sim P(\\Theta_0),$$ puts the particle in a less-optimized basin so that the score-distillation ODE can follow a generative path toward the target distribution of the edit prompt. The paper then shows that the needed perturbation level can be inferred from the local loss landscape rather than searched for, and that identity-preserving gradients added in later refinement steps pull the result back toward the source object. This combination is what the paper claims makes color, appearance, and geometry edits, including pose changes and object additions, possible from a single text prompt.","pith_inferences":["If interpolation is truly partial de-optimization, the same principle should transfer to other representations trained from random initialization, such as voxel grids or meshes, with the loss-decrease rule setting $\\eta$.","The effect may depend on the parameter geometry of the chosen NeRF implementation; a controlled comparison across architectures would separate a general principle from an architecture-specific accident.","Sequential edits could reuse the perturb-and-revise loop to preserve identity across multiple prompts, treating each edit as a fresh generative trajectory anchored to the original object."],"forward_implications":["A single text prompt can drive pose changes and object additions in NeRFs, not just recoloring or restyling.","Users do not need to describe the source object, and the perturbation level is picked automatically rather than by grid search.","Edits complete in roughly 1.5k distillation steps plus 1k refinement steps, about 7 minutes in the paper's setup, compared with about 26 minutes for full regeneration.","The same perturbation idea extends to real-scene editing and, in the paper's supplementary experiments, shows preliminary gains when applied to 3D Gaussian splatting."],"supporting_citations":[{"why":"Supplies the particle-based variational inference view of score distillation that the paper's parameter ODE builds on.","marker":"[62]"},{"why":"Gives the Score Distillation Sampling update rule that serves as the base editing engine and as a comparison baseline.","marker":"[41]"},{"why":"Supplies the multi-view consistent diffusion model used in the editing ODE and in regeneration comparisons.","marker":"[48]"},{"why":"Defines the instruction-based 3D editing baseline and the CLIP directional metrics used in evaluation.","marker":"[8]"},{"why":"Provides the posterior distillation update rule that the paper perturbs and compares against on real scenes.","marker":"[25]"},{"why":"Is the NeRF implementation whose optimized hash-grid parameters are the space in which interpolation happens.","marker":"[37]"},{"why":"Introduces the neural radiance field representation being edited and the rendering-based losses used throughout.","marker":"[35]"}],"fun_headline_variants":["Randomize NeRF weights to unlock tough 3D edits","Auto-tune perturbation to revise NeRF for any edit","Perturb the loss landscape, revise the NeRF","Random starts make NeRF editing flexible","Hard 3D edits from random NeRF perturbations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method depends on the assumption that mixing the optimized network weights with random weights acts like partially undoing the optimization, so the diffusion model sees a scene that is less finished but still structurally repairable rather than a broken jumble of geometry.","fun_headline_variants_meta":{"raw":{"variants":["Randomize NeRF weights to unlock tough 3D edits","Auto-tune perturbation to revise NeRF for any edit","Perturb the loss landscape, revise the NeRF","Random starts make NeRF editing flexible","Hard 3D edits from random NeRF perturbations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000366,"raw_usage":{"total_tokens":1947,"prompt_tokens":905,"completion_tokens":1042,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":965}},"tokens_in":521,"tokens_out":1042,"duration_ms":8688,"temperature":1.0,"reasoning_tokens":965,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:50:21.526917+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render the perturbed NeRF at a moderate perturbation level (around $\\eta = 0.3$) before any distillation runs; if the geometry is already unrecognizable or incoherent, the claimed re-entry into a generative trajectory fails. A second check is to compare edits obtained by interpolating toward two different random initializations with the same $\\eta$: if the outputs diverge sharply, the method is sensitive to which random parameters are drawn.","supporting_citations":[{"cited_title":"Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distilla- tion","cited_arxiv_id":null,"evidence_quote":"Supplies the particle-based variational inference view of score distillation that the paper's parameter ODE builds on."},{"cited_title":"Posterior dis- tillation sampling","cited_arxiv_id":null,"evidence_quote":"Provides the posterior distillation update rule that the paper perturbs and compares against on real scenes."},{"cited_title":"Instant neural graphics primitives with a mul- tiresolution hash encoding","cited_arxiv_id":null,"evidence_quote":"Is the NeRF implementation whose optimized hash-grid parameters are the space in which interpolation happens."},{"cited_title":"Nerf: Representing scenes as neural radiance fields for view syn- thesis","cited_arxiv_id":null,"evidence_quote":"Introduces the neural radiance field representation being edited and the rendering-based losses used throughout."}],"review_version":1}