{"id":"edffcbd0-6bf2-4afa-b4ed-39f594a64976","arxiv_id":"2608.10286","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Dense-view-conditioned teacher supervision applied along the student's own denoising trajectory improves sparse-view 3D Gaussian Splatting restoration across DL3DV, Mip-NeRF 360, and NeRFBusters.","lead":"TRACE-GS is a training method for restoring 3D scenes captured from only a few camera views. It trains a diffusion restorer with a teacher that sees extra views during training, and it supervises the student along the states it actually visits at inference time, without needing those extra views at deployment.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Dense-conditioned teacher's reliability at student-visited states is never measured; if its targets are inaccurate on those states, the on-policy gains in Table 3 could be an artifact of target noise rather than privileged geometry.","rationale":"The paper's central claim has two coupled requirements: the student must be supervised at its own rollout states, and the teacher targets at those states must be at least as reliable as the student's own predictions. The first requirement is plausibly supported by the E-vs-F ablation, which isolates the queried-state distribution. The second is never directly measured. The teacher is not a ground-truth oracle: it is a frozen LoRA adapter trained on forward-noised states of an offline pseudo-target. Querying it at student-visited states is therefore itself an off-policy use of the teacher, and the manuscript acknowledges the lack of uncertainty filtering in Sec. A.5 and lists uncertainty-aware supervision as future work. If the teacher targets are noisy or biased exactly where the student needs them most, the loss in Eq. (6) could have the same failure mode the paper attributes to off-policy supervision, just moved from the student to the teacher. The Table 3 margin between on-policy and off-policy variants is modest (0.17-0.31 dB), so the mechanism could in principle be supported by the numbers, but the absence of any teacher-error measurement or error bars means the conditional verdict is appropriate. The proposed one-step test would settle whether the dense-conditioned teacher actually provides better directions at visited states; until such a test is reported, the central mechanism is a plausible but unverified premise.","tokens_in":22395,"tokens_out":13326,"duration_ms":133209,"concrete_test":"On a held-out set of paired clips, roll out the trained student with the deployment 50-step schedule and record the N visited states. At each state, apply the scheduler once with the teacher's velocity and once with the student's velocity, and measure the distance (latent L2 or decoded PSNR/LPIPS) of each one-step successor to the offline pseudo-target bIT. If the teacher's successor is not closer to bIT than the student's in the late denoising steps where Fig. 6 shows the on-policy advantage, the targets in Eq. (6) are not reliably better than the student's own prediction, and the on-policy gain cannot be attributed to privileged teacher supervision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of TRACE-GS is that the dense-conditioned teacher supplies reliable velocity and retrieval targets at the student-visited states z^S_{t'_i}, which is assumed without measurement in Eqs. (5)-(8). The teacher is trained only on forward-noised states obtained from offline pseudo-targets bIT (Sec. A.4), so student-visited states generated by the sparse-conditioned student under a detached scheduler may be out-of-distribution for the teacher. At deployment, Sec. A.5 explicitly applies 'no uncertainty, perceptual-quality, or confidence-based filtering' to restored frames, and the Limitations section lists uncertainty-aware supervision only as future work. If the teacher's predictions are poor at these states, L_traj and L_ret in Eq. (9) anchor the student to those errors. The Table 3 ablation (F vs E) shows only net gains (+0.17 dB DL3DV, +0.31 dB Mip-NeRF), not teacher-error statistics; the small margin leaves room for the gains to come from extra state coverage or regularization rather than from reliable privileged targets. The paper provides no measurement of teacher error at visited states or of the correlation between teacher-target error and student improvement, so the load-bearing premise is unvalidated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TRACE-GS, a training-time distillation framework for sparse-view 3D Gaussian Splatting restoration. The method freezes a video-diffusion restoration backbone and trains LoRA adapters so that a dense-view-conditioned teacher and a sparse-view-conditioned student share the same architecture and capacity. After a direction-alignment warm-up, the student performs its own reverse-diffusion rollout, and the frozen teacher is queried at each student-visited state, providing trajectory-distillation and retrieval-alignment targets. At deployment only the sparse-conditioned student adapter is retained and its restored renderings are used as pseudo-observations in an iterative 3DGS refinement loop. Experiments on DL3DV-Benchmark, Mip-NeRF 360, and NeRFBusters report consistent PSNR/SSIM/LPIPS gains over diffusion-based baselines under 3, 6, and 9 input views, and ablations isolate the effect of privileged geometry and of on-policy versus off-policy state selection.","tokens_in":22680,"tokens_out":4594,"duration_ms":47209,"significance":"If the central claim is validated, the paper makes a useful conceptual contribution: it identifies the training-inference state mismatch in diffusion-based sparse-view 3DGS restoration and addresses it with a LUPI-style dense-conditioned teacher, without adding deployment-time cost. The strengths of the paper are the clear problem formulation, the capacity-matched teacher-student design, and the well-constructed ablation (E) versus (F) that isolates on-policy state selection from training budget. The deployment procedure is carefully matched to GSFixer, and the authors state explicit limitations. The main weakness is empirical verification: all quantitative claims rest on single-run means, several baseline numbers are copied from a previous evaluation, and the reliability of teacher targets at student-visited states is never measured. These issues are fixable and do not invalidate the idea, but they currently prevent the central claim from being considered established.","major_comments":[{"comment":"The load-bearing premise that the dense-conditioned teacher supplies reliable targets at student-visited states is asserted but never measured. The teacher is trained on forward-noised states derived from offline pseudo-targets (Sec. A.4), whereas at Stage 2 it is queried at states produced by the sparse-conditioned student under a detached scheduler; these states can lie outside the teacher's training distribution. Sec. A.5 explicitly applies no uncertainty or confidence filtering, and the Limitations section relegates uncertainty-aware supervision to future work. Because the Table 3 (E)-versus-(F) comparison shows only net gains (+0.17 dB on DL3DV, +0.31 dB on Mip-NeRF 360), the observed improvement could in principle come from additional state coverage or from implicit regularization rather than from the reliability of privileged targets. Please provide a diagnostic of teacher error at visited states, for example the teacher velocity error against the warm-up reference at matched timesteps or against ground-truth clean latents where available, and report how this error evolves along the rollout. This is needed to substantiate the mechanism claimed in Eqs. (5)-(8).","section":"Sec. 3.4 / Eqs. (5)-(8)"},{"comment":"All reported numbers are single-run means without error bars, standard deviations, or significance tests. Several decisive margins are small: in Table 3, variant (E) versus (F) differs by 0.17 dB PSNR on DL3DV and 0.31 dB on Mip-NeRF 360; in Table 2, the 9-view PSNR gain over GSFixer is at most 0.57 dB. Without a measure of run-to-run variance, the ordering of variants and the claim of consistent gains are not statistically established. Please report results over at least three seeds for the main comparisons and the core ablations, or provide confidence intervals, and state whether the same pseudo-targets and evaluation scenes are reused across seeds.","section":"Sec. 4.2 / Tables 1-3"},{"comment":"A substantial fraction of the baseline numbers in Tables 1 and 2 is taken from the GSFixer common evaluation rather than re-run under the exact protocol used for TRACE-GS. The supplement states that baseline values from GSFixer are reported unchanged. Because the central empirical claim rests on comparison with these baselines, and because the restore-refine pipeline (trajectory generation, repair schedule, pseudo-observation weighting) can change results by amounts comparable to the reported gaps, the paper should clearly mark which cells are copied, verify protocol equivalence, and ideally re-run at least the closest baseline (GSFixer) in the same code environment. Without this, the headline improvements over prior methods are not fully verifiable.","section":"Sec. B / Tables 1-2"}],"minor_comments":[{"comment":"Equation (2) contains underbrace annotations that are difficult to read in the submitted PDF; please typeset the teacher-role and student-role annotations more cleanly.","section":"Eq. (2)"},{"comment":"The relationship between training-time sparse renderings taken at 7K and 17K iterations and the deployment repair schedule at 7K, 17K, and 27K iterations is not explained; clarify whether the two sets of iteration numbers are intended to correspond.","section":"Sec. 4.1"},{"comment":"The caption does not state how many clips are used for the per-step evaluation or whether intermediate PSNR/LPIPS are computed on decoded latents or on features; please provide these details in the caption or in the supplement.","section":"Figure 6"},{"comment":"The definition of H_i^b is given after the equation that uses it; reorder the presentation so that H_i^b is defined before the attention expression.","section":"Eq. (7)"},{"comment":"The phrase 'colors from yellow to red' is not accessible in grayscale printing; consider adding numeric emphasis or a separate table note.","section":"Tables 1-2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for the journal and the core idea is timely. The main risk is empirical verification rather than conceptual soundness; I encourage the authors to add repeated-run statistics, a teacher-error diagnostic at on-policy states, and a clear disclosure of which baseline numbers were reused. The reuse of the GSFixer common evaluation should be checked for consistency and, if possible, supplemented with at least one independently re-run baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is a solid, well-scoped contribution to diffusion-based sparse-view 3DGS restoration. The central move is to supervise the student at the states it actually visits during rollout, using a teacher conditioned on dense-view renderings available only at training time. The components are known, but the combination is new, and the experiments support its value. I don't think the stress-test worry about teacher reliability breaks the paper, but it is the right question to press.\n\nWhat's genuinely good: the ablation (E) vs (F) shares the same teacher and training budget, varying only whether the queried states are independently sampled or student-visited. That isolates the claimed effect. The results are consistent across three datasets and three sparsity levels, and the per-step quality curves (Fig. 6) show the predicted late-emerging gap. The paper is also honest about where baseline numbers come from and about matching the GSFixer protocol.\n\nSoft spots, in proportion: the biggest is that the paper never measures how good the teacher's targets actually are at student-visited states. The teacher was trained on forward-noised states from dense-conditioned pseudo-targets; the student's rollout states may be outside that distribution. If the teacher is wrong there, the trajectory loss anchors the student to those errors. The stress-test flags this as load-bearing, and I largely agree, but I don't think it is fatal. The teacher is the same architecture and the states are just latents; the real diagnostic is cheap: report teacher error at visited states vs. off-policy states, and show the correlation with student improvement. Without that, the small margins in Table 3 could come partly from extra state coverage or regularization. That is a moderate concern, not a rejection.\n\nOther, smaller issues: all tables report single-run means with no error bars or significance tests; baseline numbers are mostly taken from GSFixer's public eval rather than re-run; and no code or checkpoints are released. Standard for the field, but they limit how strongly we can read the numbers.\n\nThe citation pattern looks fine: no circularity, and the related work on on-policy distillation and LUPI is relevant.\n\nBottom line: this deserves a serious referee. The core idea is plausible, the ablations are the right ones, and the writing is clear. Ask the authors to measure teacher error at visited states and to report variance; then this becomes a convincing paper.\n\nRecommendation: send to peer review.","headline":"A solid, well-scoped training-time recipe for sparse-view 3DGS restoration; the on-policy/privileged-geometry combination is new and the ablations are well-designed, but the teacher-reliability question needs direct measurement before the mechanism is fully established.","tokens_in":23169,"tokens_out":2401,"would_cite":true,"duration_ms":24390,"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":"Supervising sparse-view 3DGS diffusion at the states it actually visits, with targets from a dense-view teacher, reduces error accumulation and lifts restoration quality.","keywords":["sparse-view 3D Gaussian splatting","on-policy trajectory distillation","privileged information","diffusion-based restoration","off-policy mismatch","LoRA adaptation","novel view synthesis","3DGS refinement"],"falsifier":"A direct check is to record the teacher's velocity error against ground-truth targets at the student-visited states used by TRACE-GS and compare it with the teacher's error at independently forward-noised states; if the on-policy states are not supervised more accurately than the off-policy states, the claimed benefit of trajectory distillation would be expected to disappear or reverse.","tokens_in":22221,"feed_emoji":"🎯","tokens_out":9707,"duration_ms":77333,"temperature":0.7,"pith_summary":"TRACE-GS argues that diffusion-based sparse-view 3DGS restoration fails less from architecture than from where supervision is applied: training at independently noised states leaves the student's own denoising states unsupervised, and under-constrained sparse geometry makes errors compound along the rollout. To fix this, it trains an identical-capacity teacher conditioned on dense-view renderings (geometry available only during training) and queries that frozen teacher at each state the sparse-view student visits, aligning denoising velocities and cross-view retrieval responses there. Only the sparse-view student is deployed, so the richer geometry costs nothing at inference. On DL3DV-Benchmark, Mip-NeRF 360, and NeRFBusters, the paper reports consistent PSNR, SSIM, and LPIPS gains under 3, 6, and 9 input views, with the largest gains at 3 views. The paper's central point is that coupling reliable targets with student-visited states is the transferable ingredient, not a more sophisticated restoration network.","feed_headline":"On-policy denoising sharpens sparse-view 3D Gaussian reconstructions","feed_subtitle":"A dense-view teacher supervises the student only where it actually visits, cutting geometry drift at 3, 6, and 9 views.","key_machinery":"The load-bearing object is a geometry-asymmetric teacher–student pair built on a frozen video-diffusion restoration backbone with reference cross-attention. The two roles share the frozen weights and rank-32 LoRA adapters cloned from a shared Stage-1 alignment; the teacher is conditioned on dense-view renderings $R^d$ and the student on sparse-view renderings $R^s$, so the asymmetry lies entirely in the conditioning geometry. During on-policy trajectory distillation, the student rolls out a resampled 10-step reverse schedule from noise, the scheduler transitions are detached, and the frozen teacher is queried at each visited state $z^S_{t'_i}$ to align velocity predictions (loss $\\mathcal{L}_{\\mathrm{traj}}$) and retrieval responses from the final-block reference cross-attention (loss $\\mathcal{L}_{\\mathrm{ret}}$). Detaching the transitions lets gradients reach the student adapter at every visited state without backpropagating through the rollout, and the teacher's targets are applied exactly where the student needs them. Deployment keeps only the student adapter and feeds its restored, posed pseudo-observations back into iterative 3DGS optimization.","core_discovery":"On the paper's own terms, the central discovery is that supervision at independently forward-noised states does not cover the states reached during inference, and in sparse-view 3DGS this off-policy mismatch interacts with geometry-induced bias to accumulate errors along the denoising rollout. TRACE-GS therefore performs on-policy trajectory distillation: a dense-view-conditioned teacher, with the same frozen diffusion backbone and the same capacity as the student and differing only in conditioning, is queried at the sparse-view student's own rollout states to supply velocity targets and normalized reference cross-attention responses, through losses $\\mathcal{L}_{\\mathrm{traj}}$ and $\\mathcal{L}_{\\mathrm{ret}}$. The teacher's advantage comes only from renderings of a 3DGS fitted with more training views, putting the method in the learning-using-privileged-information setting; at deployment the student alone restores renderings that serve as pseudo-observations for 3DGS refinement. The paper reports that this consistently improves PSNR, SSIM, and LPIPS on DL3DV-Benchmark, Mip-NeRF 360, and NeRFBusters, with the largest gains under 3-view input and a late-rollout gap between the on-policy and off-policy variants that the authors interpret as error accumulating at uncovered states.","pith_inferences":["Editorial inference: the same on-policy-distillation mechanism should transfer to other conditioning asymmetries, such as depth or multi-modal input, whenever a privileged signal exists only during training; the essential step is querying the privileged model at the deployable model's visited states.","Editorial inference: since the teacher is queried without any confidence or uncertainty filtering, a natural strengthening is to weight on-policy targets by estimated teacher error, and a failure analysis should track whether teacher error at visited states is the limiting factor.","Editorial inference: the per-step results imply that later rollout states carry most of the signal, so annealing the Stage-2 rollout length (short early, long late) is a testable way to cut memory cost without losing the late-state benefit.","Editorial inference: the off-policy variant's per-step plateau suggests a general lesson for diffusion restoration beyond 3DGS: replaying real student states during training is valuable whenever the restoration input itself changes across refinement rounds."],"forward_implications":["If the central claim holds, training-time dense views can transfer geometric reliability to a deployment-time sparse-view model at zero extra inference cost: teacher, dense reconstruction, and additional views are all discarded at test time.","The largest gains under 3 input views follow directly: the sparser the conditioning geometry, the wider the teacher–student gap and the more value on-policy supervision has.","Because the on-policy and off-policy variants separate only in the later denoising steps, the method implies that error accumulation, not capacity, is the bottleneck in sparse-view restoration, so later denoising states deserve the most supervision.","Retrieval alignment contributes beyond velocity matching, suggesting that constraining how the student attends to reference views along its own trajectory is a usable lever for cross-view consistency.","The framework is a drop-in refinement stage for any restore–refine pipeline where extra training views exist, not just the GSFixer-style backbone used here."],"supporting_citations":[{"why":"Supplies the frozen video-diffusion restoration backbone, the reference-token conditioning interface, and the restore-refine protocol that TRACE-GS inherits and compares against.","marker":"Yin et al. 2025"},{"why":"Provides the pretrained CogVideoX-style video DiT and VAE that both teacher and student share.","marker":"Yang et al. 2024"},{"why":"Documents the off-policy exposure-bias problem that motivates supervising student-visited states.","marker":"Li et al. 2024b"},{"why":"Shows how deviations compound along diffusion rollouts, the phenomenon the paper says sparse-view geometry amplifies.","marker":"Ning et al. 2024"},{"why":"Gives the imitation-learning reduction that justifies on-policy supervision at states the student actually visits.","marker":"Ross, Gordon, and Bagnell 2011"},{"why":"Defines learning using privileged information, the training-only/deployment-free setting TRACE-GS adopts.","marker":"Vapnik and Vashist 2009"},{"why":"Supplies the generalized-distillation formulation for transferring knowledge from a privileged teacher to a deployable student.","marker":"Lopez-Paz et al. 2015"},{"why":"Provides the DL3DV scenes used to build paired sparse- and dense-view renderings and the in-domain benchmark.","marker":"Ling et al. 2024"}],"fun_headline_variants":["On-policy denoising fixes sparse-view 3DGS geometry drift","Teacher guides student only on its own path in 3DGS","Privileged geometry distills better sparse-view 3DGS","Sparse-view 3DGS sharpened by on-policy trajectory distillation","Train with rich views, deploy sparse: TRACE-GS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the dense-view teacher supplies reliable targets at the states the sparse-view student actually visits, even though those states come from a weaker model's rollout and the paper applies no confidence or uncertainty filtering to the teacher's pseudo-observations.","fun_headline_variants_meta":{"raw":{"variants":["On-policy denoising fixes sparse-view 3DGS geometry drift","Teacher guides student only on its own path in 3DGS","Privileged geometry distills better sparse-view 3DGS","Sparse-view 3DGS sharpened by on-policy trajectory distillation","Train with rich views, deploy sparse: TRACE-GS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1391,"prompt_tokens":1064,"completion_tokens":327,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":680,"completion_tokens_details":{"reasoning_tokens":232}},"tokens_in":680,"tokens_out":327,"duration_ms":3502,"temperature":1.0,"reasoning_tokens":232,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:10:50.182531+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct check is to record the teacher's velocity error against ground-truth targets at the student-visited states used by TRACE-GS and compare it with the teacher's error at independently forward-noised states; if the on-policy states are not supervised more accurately than the off-policy states, the claimed benefit of trajectory distillation would be expected to disappear or reverse.","supporting_citations":[],"review_version":1}