{"id":"449f09a3-847e-47ce-88b6-77dba5802c95","arxiv_id":"2412.08152","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"ProGDF trains a lightweight neural field to reproduce the optimization path of a 3D Gaussian edit, providing a real-time slider that interpolates between the original and edited scenes.","lead":"This paper introduces a 3D editing system for Gaussian splatted scenes that trains once and then lets users adjust the result in real time with a slider. It models the editing process as a neural field that outputs changes to scene attributes, enabling controllable and composable edits.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Controllable-slider claim rests on unvalidated interpolation: GDF is trained on sparse discrete trajectory frames but the user slider is continuous; no evidence that interpolated offsets are valid intermediate edits.","rationale":"The most load-bearing assumption is that the GDF generalizes continuously from the discrete trajectory samples. This is exactly what separates a user-controllable slider from a lookup table of precomputed frames. The reader identified this same concern; I concur and make it more concrete: the memory bank's sampling interval (100 steps) and the discretization into 10 bins mean the GDF sees only ~10-15 distinct control values, yet the user interface is continuous. The progressive constraint is the only stated reason the trajectory is smooth enough to interpolate, but it acts on PGS update magnitudes, not on the rendered-image manifold or the GDF's interpolation behavior. This is an empirical question that the paper does not address; there is no quantitative experiment measuring interpolation quality, monotonicity, or smoothness. I do not see a fatal internal inconsistency; the method could work given suitable smoothness, but the evidence is qualitative only. Other issues (best-of-three comparison, undefined H, missing code) are secondary and would be partially resolved by the proposed controllability experiments and code release. The correct verdict remains CONDITIONAL, matching the reader's assessment.","tokens_in":12319,"tokens_out":7365,"duration_ms":77527,"concrete_test":"Train ProGDF on a scene with T=1500. During GDF training, hold out every other sampled trajectory frame (with default 100-step interval, hold out t=200,400,...). After training, render GDF outputs at control signals corresponding to the held-out frames and compare against the actual PGS-rendered frames at those timesteps using PSNR/LPIPS. Also sweep the slider continuously from 0 to 1 in 50 increments, rendering each result, and check that CLIP directional similarity to the instruction and perceptual distance to the original are monotonically non-decreasing and that frame-to-frame differences show no discontinuities. If held-out predictions degrade significantly or the slider sweep is non-monotonic/jerky, the continuous controllability claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of real-time controllable editing (Abstract, Sec. 1) requires that the GDF, trained on a sparse set of discrete intermediate states, produces valid edited scenes for arbitrary continuous slider positions. Supervision comes from an online memory bank of trajectory frames sampled every 100 steps (Appendix), and control signals are further discretized into k=10 bins. At inference, a continuous slider is mapped through H(t/T) into Eq. (8), so the MLP must interpolate between these sparse supervision points. The paper's only stated support for smooth interpolation is the progressive constraint Eq. (3), which penalizes per-step parameter-update magnitudes during PGS training. That constraint does not guarantee that Euclidean interpolation between Gaussian offsets (which is what the MLP effectively produces) corresponds to semantically meaningful, artifact-free, or monotonic intermediate edits. No quantitative evaluation of controllability is provided: the user study scores only final results, not the coherence of the slider trajectory. Moreover, the 'learnable transformation H' is never specified or trained, so the mapping from slider position to control signal is undefined. If the GDF merely memorizes the sampled frames and interpolates in parameter space, the real-time controllability claim fails.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"ProGDF proposes to replace iterative in-the-loop training for 3D Gaussian Splatting (3DGS) editing with a single out-of-loop training session. The method first runs a Progressive Gaussian Splatting (PGS) stage that optimizes 3D Gaussians from the original scene toward a target scene under a progressive constraint and a Laplacian-based rendering-quality regularizer, while recording intermediate states along the optimization trajectory. A lightweight Gaussian Differential Field (GDF), an MLP with about 6M parameters, is then trained in parallel to map a spatial position and a discretized control signal to attribute offsets, with supervision from resampled trajectory frames. At inference, a user moves a slider to select a control signal, and the edited Gaussians are produced in 0.02 seconds. The paper also demonstrates fine-grained 3D manipulation by composing multiple region-specific GDFs. Experiments compare ProGDF with Instruct-N2N and GaussianEditor on user study scores and CLIP directional similarity, and ablations study the progressive constraint and rendering-quality regularization.","tokens_in":12589,"tokens_out":4923,"duration_ms":52272,"significance":"The central idea of process-oriented modeling for 3D editing is timely and potentially valuable. If validated, the out-of-loop training paradigm would be a practical step beyond per-edit optimization, and the claimed 0.02-second slider-based control would substantially improve the user experience of 3DGS editing. The paper has clear strengths: the GDF is lightweight, training is parallel with the PGS stage, inference is fast, and the two applications (controllable editing and fine-grained manipulation) are novel and well motivated. However, the current evidence is weakened by an evaluation protocol that gives ProGDF an advantage over baselines, by the absence of quantitative validation of the central slider-interpolation claim, and by an underspecified control-signal transformation. These issues are localized and fixable, so the contribution is worth pursuing after substantial revision.","major_comments":[{"comment":"The quantitative comparison is not apples-to-apples. The text states that for ProGDF the authors 'randomly adjust its slider for three inference runs and take the best results', while Instruct-N2N and GaussianEditor are evaluated with a single deterministic run. Taking the best of three runs introduces a selection bias that can inflate both the user study score and the CLIP Sdir value. To support the claimed superiority, report results for a fixed protocol, such as one predefined slider position (e.g., mid-range), or report the mean and standard deviation over multiple slider positions for all methods, with appropriate significance tests.","section":"Quantitative Results, Table 1"},{"comment":"The central controllability claim requires that the GDF generalizes between the discrete control signals used in training. The control signal is discretized into k=10 bins and supervised by trajectory frames sampled every 100 steps, yet the user slider is continuous. The paper provides no direct evidence that slider positions between bins produce valid, artifact-free, semantically meaningful, or monotonic intermediate edits; the user study and CLIP scores evaluate only final results, not the coherence of the slider trajectory. I recommend adding a quantitative interpolation experiment, for example rendering a fixed view across many slider values and reporting per-slider CLIP directional similarity, rendering-quality metrics, or a dedicated user study on trajectory coherence.","section":"Gaussian Differential Field, Eq. (8)"},{"comment":"The 'learnable transformation H' in Eq. (8) is never defined. The paper says that the goal is to construct a learnable transformation H to map the optimization trajectory to a slider, but it does not specify the parameterization of H, its input and output dimensions, how it is trained, or whether it is the identity map on t/T. Without this information, the mapping from slider position to control signal is underspecified and the experiments cannot be reproduced. Please define H explicitly and include its implementation details, or state that H is fixed and describe its form.","section":"Gaussian Differential Field, Controllable 3D Editing"},{"comment":"The GDF is trained on, and evaluated against, the same PGS trajectory frames that it is designed to reproduce. This does not by itself invalidate the method, but it means the reported variable editing results are essentially reconstructions of training samples. There is no evidence that the GDF generalizes to control values not seen during training or to trajectories generated under different conditions. I recommend a held-out evaluation, such as training the GDF on a subset of trajectory frames and testing on the remaining frames, or comparing a slider path used during training with an unseen slider path, to demonstrate that the GDF learns a continuous process rather than memorizing a discrete set.","section":"Gaussian Differential Field, Trajectory Resampling and Eq. (10)"},{"comment":"Equation (3) is ambiguous. The term ΔΘ_t_i is described as the 'variation' of the i-th Gaussian at time step t, but no norm is specified; as written, the sum over signed updates could cancel and would not penalize the magnitude of parameter changes. Please define ΔΘ_t_i explicitly, e.g., as the L2 norm of the per-step parameter update, and clarify how the constants α, β, and s are chosen so that the progressive constraint has the intended dampening effect on the trajectory.","section":"Progressive Gaussian Splatting, Eq. (3)"}],"minor_comments":[{"comment":"The notation L_render = -∇²_Laplacian R(Θ_t, v) is confusing because the Laplacian is already a second-derivative operator. Please state precisely what ∇²_Laplacian means, for example that L_render is the negative mean Laplacian of the rendered image, and define the averaging over pixels.","section":"Eq. (4) and Eq. (11)"},{"comment":"The user study reports a 95% confidence interval but does not describe the participant pool, the number of scenes and instructions per participant, the order of presentation, or whether any statistical test was performed between methods. Including these details would strengthen the comparison, especially given the best-of-three protocol.","section":"Appendix, User Study"},{"comment":"With T between 1500 and 2000 steps and a trajectory resampling interval of 100, only about 15 to 20 discrete time stamps are actually used for supervision, and these are further collapsed into k=10 bins. The paper should discuss how this limited number of distinct control values is reconciled with the claim of continuous slider control.","section":"Appendix, Implementation Details"},{"comment":"For the flexible fine-grained manipulation results, the caption notes that some instructions are 'not real inputs' but are used only to indicate which editing results are combined. It would be clearer to describe the actual user interaction for composition, including how multiple GDFs are blended when their masks overlap.","section":"Figure 2 and Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The paper introduces an interesting and potentially impactful paradigm, but the evaluation protocol and the underspecification of the control-signal mapping are substantial. The best-of-three comparison is the most serious issue because it directly affects the headline quantitative claims; it should be corrected in any revision. The interpolation concern is central to the method's core promise, and the proposed extra experiments should be made mandatory by the editor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look, but read Table 1 with suspicion. The core idea—train once, then let the user slide through edit strength—is genuinely new among the cited 3DGS editing works, and the PGS/GDF split is a clean way to get there. The problem is real; in-the-loop retraining for each edit is indeed a pain, and 0.02s slider response after a single session would be practically useful. The method is described clearly enough to reproduce in principle: progressive constraint in Eq. (3), rendering quality regularization in Eq. (4), memory-bank supervision, and the GDF MLP. The qualitative results in Figures 4–6 and the ablations in Figures 7–8 show each component does something. Credit where due: the authors identify an actual usability bottleneck and propose a plausible fix.\n\nNow the soft spots, in order of size. First, the quantitative comparison is unfair in a way that matters: ProGDF gets best-of-three slider runs, baselines get one deterministic run. That selection bias alone could explain the CLIP gap, and the user study inherits the same inflation. Second, the continuous-slider claim is under-supported. Eq. (8) feeds a discretized control signal H(t/T) into the MLP; supervision comes from sparse frames sampled every 100 steps and then binned into k=10. The smoothness of the slider trajectory is never evaluated. The progressive constraint is a reasonable prior, but it does not guarantee that Euclidean interpolation between Gaussian offsets gives meaningful intermediate edits. The stress-test note is right: this is the load-bearing assumption, and the paper leaves it qualitative. Third, 'H' is called a learnable transformation but no parameterization, training, or ablation is given anywhere—it reads as an unspecified component. Fourth, no code or data is released, which hurts because the interpolation issue is empirically checkable. The paper's own limitations section mentions drag/click/scribble and decoupled texture control, but not the unvalidated interpolation; the limitations are honest but incomplete.\n\nBottom line: the paper deserves a serious referee. The out-of-loop training paradigm is a real step for the 3D editing subfield, and the weaknesses are addressable rather than fatal. I'd send it out asking specifically for an honest comparison protocol (matched inference runs) and quantitative evidence of slider continuity—e.g., interpolation quality at intermediate slider values, or agreement between predicted and actual trajectory frames. Who is this for: anyone working on interactive 3D editing with Gaussians. It won't reshape the field, but it is a legitimate contribution.","headline":"Real idea (out-of-loop training) with an unvalidated continuous-slider claim and an unfair comparison protocol; send to review but ask for the missing evidence.","tokens_in":13102,"tokens_out":2609,"would_cite":false,"duration_ms":26044,"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":"ProGDF turns 3D editing into a learned offset field, so one offline training pass yields real-time control.","keywords":["3D editing","Gaussian splatting","progressive modeling","differential field","instruction-guided editing","real-time interaction","controllable editing","fine-grained 3D manipulation"],"falsifier":"Render GDF outputs at several control values that fall strictly between the ten discretized bins, such as 0.23, 0.47, and 0.81, in a scene used for training, and independently run a fresh PGS optimization that records the true intermediate Gaussians at those same progress fractions; if the rendered interpolation images differ materially in semantic content or texture quality from the true intermediates, the central generalization assumption is refuted.","tokens_in":12137,"feed_emoji":"🎛️","tokens_out":7561,"duration_ms":67501,"temperature":0.7,"pith_summary":"ProGDF claims that a 3D editing session can be split into an offline training stage and an online interaction stage, so that after one training session users can sweep an editing-strength slider and see modified 3D scenes in 0.02 seconds. The paper frames this as process-oriented modeling: instead of producing only a final edited scene, it treats the optimization path from original to edited Gaussian splatting as the object of learning. A Progressive Gaussian Splatting (PGS) stage gathers diverse intermediate frames along that path, and a lightweight Gaussian Differential Field (GDF) maps a spatial position plus a slider value to per-Gaussian attribute offsets. If the claim holds, interactive trial-and-error editing, which currently costs tens of minutes per attempt, becomes a real-time user interface.","feed_headline":"One training session makes 3D editing a 0.02-second slider","feed_subtitle":"Instead of re-optimizing for tens of minutes per try, users scrub between original and edited scenes instantly.","key_machinery":"The load-bearing object is the Progressive Gaussian Differential Field, consisting of Progressive Gaussian Splatting (PGS) plus a lightweight neural network called GDF. PGS makes the editing trajectory visible and well-behaved: a progressive constraint $\\alpha \\beta^{t/s} \\sum_i \\Delta \\Theta_i^t$ keeps early Gaussian updates from rushing, and the Laplacian rendering-quality term $-\\nabla^2_{\\text{Laplacian}} R(\\Theta_t, v)$ preserves sharp texture in intermediate frames. GDF is the interactive machinery: it receives a 3D location and the scalar control signal $H(t/T)$, discretized into $k=10$ bins, and emits the per-Gaussian deltas $(\\Delta \\mu_i, \\Delta \\Sigma_i, \\Delta \\sigma_i, \\Delta c_i)$, which are added to the original scene. Trajectory resampling from a memory bank prevents the neural network from forgetting earlier editing states while it is trained in parallel with the Gaussians.","core_discovery":"The central discovery is that controllability in 3D editing can be modeled out-of-loop as a continuous offset field rather than by re-optimizing 3D Gaussians for each user request. Given an instruction, PGS optimizes 3D Gaussians from the original scene toward the edited scene while a progressive constraint dams early large updates and a Laplacian-based rendering-quality regularizer keeps intermediate frames sharp; those frames are stored in a memory bank. GDF, a roughly six-million-parameter neural network, takes any spatial coordinate and a normalized control signal H(t/T), discretized into ten bins, and predicts the offsets in position, scale, rotation, opacity, and color for the Gaussian at that point. The predicted offsets are added to the original Gaussians, rendered, and supervised by the same editing loss used for the target. After training, a user moves a slider and the GDF evaluates all affected Gaussians in 0.02 seconds, producing continuously variable edits, and separate GDFs trained on different masked regions can be composed for fine-grained manipulations.","pith_inferences":["The same offset-field idea could be carried beyond a scalar slider: replacing H(t/T) with a low-dimensional control vector, for example separate sliders for texture and color, is a natural extension the paper does not demonstrate.","If the GDF truly interpolates between stored frames, the representation implies edit reversibility: subtracting the predicted offset should return the scene toward the original, which the paper does not test.","A direct stress test of the generalization claim would compare GDF predictions at slider values that fall between the ten training bins against independently optimized intermediate states, separating true interpolation from memorization of sampled frames.","The composition of multiple GDFs suggests a library of reusable partial edits: training an edit once and combining regions later could make asset editing more modular than current end-to-end pipelines."],"forward_implications":["After one training session for an instruction, a user can generate a range of edit strengths from the same model, with each slider adjustment rendering in 0.02 seconds.","The same trained GDF can be reused for controllable editing without retraining, including in a session where the user decides a different amount of the edit is better.","By assigning different 3D masks to different GDFs, edits such as Batman shirt with yellow pants can be produced by composing region-specific trained fields.","If the progressive constraint is removed, the intermediate frames collapse toward the final result, so slider positions produce nearly identical outputs; if the Laplacian regularizer is removed, intermediate edits blur.","The out-of-loop design means training GDF in parallel with Gaussian optimization adds only modest compute, making the one-time training cost acceptable."],"supporting_citations":[{"why":"Supplies the 3D Gaussian Splatting representation and differentiable renderer used to render original, edited, and predicted scenes.","marker":"[Kerbl et al. 2023]"},{"why":"Provides the instruction-following image editing model E in Eq. (1), used to produce multi-view edited images from rendered originals.","marker":"[Brooks, Holynski, and Efros 2023]"},{"why":"GaussianEditor supplies the 3D mask unprojection procedure and editing-loss setup that ProGDF adapts, and serves as the 3DGS baseline for comparison.","marker":"[Chen et al. 2024]"},{"why":"Instruct-N2N is the NeRF-based instruction editing baseline against which ProGDF is compared.","marker":"[Haque et al. 2023]"},{"why":"Segment Anything is used to segment multi-view images and extract the 3D manipulation region.","marker":"[Kirillov et al. 2023]"},{"why":"Defines the Laplacian edge detector used in the rendering-quality regularization of Eq. (4) and Eq. (11).","marker":"[Wang 2007]"},{"why":"Is the source of the k-bin discretization strategy applied to the GDF control signal.","marker":"[Zhao et al. 2024]"}],"fun_headline_variants":["Pre-train once, then scrub 3D edits in 0.02s","ProGDF: one training run, slider-controlled 3D edits","Out-of-loop training: one session, real-time 3D control","One training pass, then a 0.02s slider for 3D edits","ProGDF: train once, then slide between 3D scene edits in real time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the small neural network inside GDF generalizes across the slider, rather than merely recalling the discrete intermediate frames it was trained on; if slider values between those sampled frames produce broken or non-semantic edits, the real-time controllability claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Pre-train once, then scrub 3D edits in 0.02s","ProGDF: one training run, slider-controlled 3D edits","Out-of-loop training: one session, real-time 3D control","One training pass, then a 0.02s slider for 3D edits","ProGDF: train once, then slide between 3D scene edits in real time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001753,"raw_usage":{"total_tokens":6950,"prompt_tokens":1005,"completion_tokens":5945,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":5839}},"tokens_in":621,"tokens_out":5945,"duration_ms":37975,"temperature":1.0,"reasoning_tokens":5839,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:08:15.203192+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render GDF outputs at several control values that fall strictly between the ten discretized bins, such as 0.23, 0.47, and 0.81, in a scene used for training, and independently run a fresh PGS optimization that records the true intermediate Gaussians at those same progress fractions; if the rendered interpolation images differ materially in semantic content or texture quality from the true intermediates, the central generalization assumption is refuted.","supporting_citations":[{"cited_title":"A.; Holynski, A.; and Kanazawa, A","cited_arxiv_id":null,"evidence_quote":"Instruct-N2N is the NeRF-based instruction editing baseline against which ProGDF is compared."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Laplacian edge detector used in the rendering-quality regularization of Eq. (4) and Eq. (11)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the source of the k-bin discretization strategy applied to the GDF control signal."}],"review_version":1}