{"id":"e628d25d-e186-46ea-9aec-0be258d4329b","arxiv_id":"2412.19720","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A learned prior maps low-frequency SDF observations to full-frequency coverage via disentangled latent codes and test-time optimization, sharpening 3D reconstructions.","lead":"This paper trains a neural network to turn blurry, low-detail 3D shapes into sharper versions by learning to add back high-frequency surface details. The method works by separating each shape into a low-frequency code and a full-detail code, then recovering the full-detail code for new shapes at test time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-time optimization in Eq. 3 is underdetermined: with no prior on eF/eC, the low-frequency loss can be satisfied by embeddings that decode to plausible but incorrect full-frequency shapes, so 'recovery' may be hallucination.","rationale":"Good faith reading: the paper proposes a novel frequency-domain prior with a clean disentanglement idea and releases code/data. The central claim, however, is not just that outputs look sharper but that full frequency coverage is recovered from low-frequency observations. The mechanism for this on unseen shapes is the test-time optimization of eF. I examined Eq. 3 and found it underdetermined: no latent prior, no regularization, and no identifiability argument. The empirical comparisons, while suggestive, use aggregate geometric metrics that reward plausible completion and cannot rule out hallucination. This is a more direct threat to the central claim than the train/test distribution shift noted by the reader, though the two are related: out-of-distribution inputs make hallucination more likely. The concrete test is designed to settle it: random restarts test identifiability, and frequency-band error tests recovery versus generic sharpening. If the test passes, the central claim is substantially supported; if it fails, the method would be better described as shape completion/consolidation rather than frequency recovery. Since this is unresolved, the conditional verdict stands.","tokens_in":12443,"tokens_out":6517,"duration_ms":70735,"concrete_test":"Take 50 held-out shapes from the ShapeNet test split; for each, create a low-frequency observation with the paper's protocol and also use one real NeuralTPS/OnSurf reconstruction. Run Eq. 3 ten times per input with different random initializations of eF and eC. For each run, decode fF(q,eF) and compute (i) CD/NC vs GT and (ii) an FFT-based high-frequency band error between the output occupancy and GT occupancy over frequencies above the cutoff. If the across-run CD variance is comparable to the mean CD, or if the high-frequency band error is not significantly below a category-mean 'sharp prior' baseline, then eF is not identifiable and the 'recover full frequency coverage' claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the test-time self-reconstruction in Eq. 3 recovering the true full-frequency embedding eF for an unseen shape. This is the linchpin: if the optimized eF is not the embedding of the observed shape, then fF(q,eF) produces a prior-plausible surface rather than the actual high-frequency structure. Eq. 3 minimizes only the low-frequency signed distance residual ||sL - s'_L||^2 over eF and eC with no regularization, no latent prior, and no uniqueness argument. Since eL = [eF,eC] and eC is a free per-observation code, a given blurred observation can often be fit by many (eF,eC) pairs; the optimizer may drift to a nearby training-set embedding. The reported CD/NC metrics reward completing missing structure and sharpening generic edges, so they cannot distinguish true frequency recovery from category-level hallucination. This concern applies even on the same synthetic degradation used in training, and real-input distribution shift compounds it. The paper's TSNE analysis covers training embeddings, not test-time optima, so it does not address identifiability.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes frequency consolidation priors (FCP) to sharpen low-frequency SDF observations by recovering their full-frequency counterparts. The method trains two SDF decoders, fL and fF, with per-shape embeddings eF shared between low- and full-frequency branches and an additional per-observation corruption code eC, so that eL = [eF, eC]. At test time, the learned fL is fixed and eF, eC are optimized by self-reconstruction from an unseen low-frequency observation (Eq. 3); the optimized eF is then decoded by fF to produce a sharper surface. Training pairs are generated by zeroing high-frequency FFT magnitudes of Poisson-reconstructed occupancy fields. The method is evaluated on ShapeNet, ABC, and ScanNet, reporting improved Chamfer distance and normal consistency over several recent baselines.","tokens_in":12742,"tokens_out":3790,"duration_ms":45366,"significance":"The conceptual contribution is interesting: disentangling a frequency-corruption code from a shape-identity code in order to enable test-time optimization is a plausible and generalizable mechanism for sharpening implicit functions, and it is supported by a useful release of code, data, and pretrained models. If the central claim is verified, the method would provide a practical post-processing prior for SDFs reconstructed from sparse point clouds or multi-view images. The paper also contains several positive elements: it tests across three datasets, includes visualizations of test-time optimization, and makes an effort to ablate the embedding design. However, the current evidence is not yet convincing enough to establish the central claim, because the test-time recovery step is insufficiently constrained and the evaluation protocol leaves important ambiguities.","major_comments":[{"comment":"The central claim that test-time self-reconstruction recovers the true full-frequency embedding eF is not supported by an identifiability argument. Eq. (3) minimizes only the low-frequency residual ||sL - sgt_L'||^2 over both eF' and eC', with no regularization, no prior on the latent codes, and no uniqueness guarantee. Since eC' is a free per-observation code, many (eF', eC') pairs can fit the same low-frequency observation, and the optimizer may settle on an eF' that decodes to a plausible but incorrect full-frequency shape. The TSNE analysis in Fig. 12 concerns training embeddings, not test-time optima, so it does not address this concern. I recommend adding quantitative tests of embedding recovery, such as comparing the optimized eF' with the nearest training-shape embedding, or evaluating whether the decoded shape is closer to the ground-truth shape than to a category-level template. Without such evidence, the reported metric improvements could reflect generic sharpening rather than true frequency recovery.","section":"Generalizing Frequency Consolidation Priors, Eq. (3)"},{"comment":"The evaluation protocol is under-specified. The paper states that for each test shape the authors 'generate low-frequency observations as described and use the worst observation to assess all methods,' but it does not define how 'worst' is determined, whether the same observation is used for all methods, or how the choice interacts with the six frequency subbands in Fig. 3. Table 1 reports a single mean value per metric per method, without error bars, per-band breakdowns, or significance tests. Given the large variation across frequency bands visible in Fig. 3, single means are not sufficient to support the claim of state-of-the-art performance. The authors should report per-band results, standard deviations, and the number of test shapes, and should specify the exact worst-observation protocol.","section":"Evaluation on ShapeNets, Table 1"},{"comment":"The paper's main ShapeNet evaluation is conducted on test observations produced by the same frequency-removal pipeline used for training, which limits the strength of the generalization claim. The transfer experiments to NeuralTPS, OnSurf, and ScanNet are more convincing in spirit, but the paper does not quantify the distribution shift between synthetic low-frequency observations and real sparse-point-cloud or scanned reconstructions. The statement in 'Supervisions for Learning Priors' that the over-smoothed surfaces are 'very similar' to spectral geometry results is only supported by a single visual example (Fig. 4). I recommend adding quantitative spectral comparisons between the training distribution and the real test inputs, and reporting results separately for each degradation type rather than only as averages over mixed cases.","section":"Supervisions for Learning Priors; Refining Reconstructions from Sparse Point clouds"},{"comment":"The paper cites BACON (Lindell et al. 2022) and SAP (Peng et al. 2021) as frequency-related methods but does not compare against them experimentally. Since the proposed method explicitly operates in the frequency domain and claims to recover high-frequency components, these are the most directly relevant baselines. Their absence from Tables 1–3 weakens the state-of-the-art claim. The authors should either include them in the comparisons or justify their exclusion with concrete technical reasons.","section":"Related Work, Learning with Frequency; Tables 1–3"}],"minor_comments":[{"comment":"There is a typo in the first sentence: 'Singed Distance Functions' should be 'Signed Distance Functions.'","section":"Introduction"},{"comment":"Several citations are incomplete or malformed, including 'Takikawa et al. 2021,?' and 'IDR (?)'; these should be resolved before publication.","section":"Related Work"},{"comment":"The text says 'we use the first 5 low frequency observations' for each training shape, but Fig. 3 shows six subbands and the sampling procedure in 'Supervisions for Learning Priors' is described as random. The relationship between the six illustrated bands and the 'first 5' observations is unclear and should be clarified.","section":"Learning Frequency Consolidation Priors"},{"comment":"In the NeuralTPS row, the normal consistency value appears as '50.899', which is likely a typo for '0.899'. Please check the table formatting.","section":"Table 2"},{"comment":"The paper does not specify the FFT resolution used for the Poisson solver, the number of training shapes per dataset, or the initialization of the embeddings. These details are needed for reproducibility.","section":"Implementation Details"},{"comment":"The t-SNE visualization in Fig. 12 is described as showing optimization paths, but the figure as printed is difficult to read. A larger figure with labeled axes and a legend would help the reader verify the claimed semantic structure.","section":"Ablation Studies and Analysis"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is promising and the code release is a positive factor. The main risk is that the test-time optimization in Eq. (3) may be recovering plausible full-frequency shapes rather than the true high-frequency structure of the input; the authors should be asked to provide direct evidence on embedding recovery and to tighten the evaluation protocol. The missing comparison to BACON and SAP is also a scope and completeness issue that should be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on the FCP paper. The central novelty is the construction of paired low/full frequency SDFs via FFT-domain filtering, the disentanglement of shape identity eF from corruption eC, and the test-time optimization to recover eF. That combination is genuinely new as far as I know, and the authors back it with a clean story and released code and data. The empirical results across ShapeNet, ABC, and ScanNet are suggestive; the improvements on NeuralTPS/OnSurf are especially interesting because they go beyond the synthetic degradation used in training.\n\nBut the load-bearing eval is thin in a few places. Tables 1 and 2 report single means with no error bars. The 'worst observation' protocol isn't defined. There are no frequency-domain baselines (BACON, SAP) despite them being cited as relevant. The abstract promises multi-view results but the experiments are point clouds and ScanNet only. None of these are fatal on their own, but they weaken the state-of-the-art claim.\n\nThe more serious issue is the test-time identifiability. Eq. 3 optimizes eF and eC against only the low-frequency residual, with no prior or regularization. The paper does not show that the optimized eF is unique or that it corresponds to the observed shape's identity. eC is free to absorb the low-frequency details, so a range of (eF,eC) can fit the same observation; the decoder may then produce a shape typical of the class rather than the actual high-frequency structure. The TSNE plot is on training embeddings, not test-time optima, so it doesn't address this. I'm not asserting the method is broken--the generalization on NeuralTPS/OnSurf suggests real transfer--but the paper has not ruled out hallucination as a partial explanation for the gains.\n\nOverall: this deserves a serious referee, but the referee should push hard on identifiability--adding a latent prior, analyzing uniqueness, or testing with a dataset where the true high-frequency is known. I'd bring it to a reading group; I likely won't cite it until these points are settled.","headline":"The core idea is novel and the gains are plausible, but the test-time recovery step is underdetermined and the evaluation doesn't pin down whether the prior is truly recovering instance-specific frequencies or just sharpening generically.","tokens_in":13222,"tokens_out":3235,"would_cite":false,"duration_ms":33549,"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":"This paper claims that a frequency consolidation prior, learned from FFT-generated low-frequency and full-frequency SDF pairs, can recover high-frequency surface detail from low-frequency observations via a disentangled embedding and…","keywords":["frequency consolidation prior","neural implicit functions","signed distance functions","frequency-domain reconstruction","test-time optimization","point cloud reconstruction","sharp surface reconstruction","FFT Poisson solver"],"falsifier":"Evaluate the trained prior on low-frequency SDFs created by a different degradation mechanism than FFT truncation, for example SDFs learned by an MLP with spectral bias from sparse point clouds, and compare reconstruction error before and after sharpening; if the gain over the unsharpened input is much smaller than on FFT-truncated test pairs, the prior has learned to invert its own training degradation rather than to consolidate frequency coverage in general.","tokens_in":12263,"feed_emoji":"🧊","tokens_out":6403,"duration_ms":60119,"temperature":0.7,"pith_summary":"Neural signed distance functions tend to capture low-frequency geometry, so surfaces reconstructed from sparse point clouds or multi-view images come out smooth and incomplete. This paper argues that a prior learned in the frequency domain can restore the missing high-frequency content: given a low-frequency SDF observation, it recovers a full-frequency SDF with sharper edges and more complete structures. The paper's central claim is that this frequency consolidation prior generalizes to unseen shapes when frequency content is stored as disentangled embeddings, because test-time optimization can recover the full-frequency embedding from the observed low-frequency SDF alone. If correct, the method offers a post-hoc sharpening step for existing neural implicit reconstruction pipelines.","feed_headline":"A learned prior restores sharp edges to blurry 3D shapes","feed_subtitle":"Neural SDFs recover missing high-frequency detail from sparse point clouds and scans via test-time embedding recovery.","key_machinery":"The central object is the frequency consolidation prior: a data-driven mapping from a low-frequency SDF observation to its full-frequency coverage, represented by two neural SDF decoders $f_L(q, e_L)$ and $f_F(q, e_F)$ that share query conditioning and are trained jointly on pairs of low- and full-frequency occupancy fields. Low- and full-frequency fields are generated efficiently by an FFT-based Poisson surface reconstruction, avoiding the cost of eigen-decomposition of Laplace-Beltrami operators. The load-bearing identity is the embedding disentanglement $e_L = [e_F, e_C]$, which separates shape identity from frequency-corruption information; this is what allows a frozen $f_L$ to recover $e'_F$ at test time through self-reconstruction, after which $f_F(q, e'_F)$ yields the sharpened surface.","core_discovery":"The paper claims that full frequency coverage can be recovered from a low-frequency observation of a signed distance function, and that this recovery is the mechanism behind sharper and more complete surfaces. It builds a training set by solving Poisson surface reconstruction with an FFT-based solver, zeroing out high-frequency magnitudes to create low-frequency observations, and pairing each with its full-frequency reconstruction. It then learns two SDF decoders, one for low frequencies and one for full frequencies, connected by embeddings where the low-frequency embedding is the concatenation of a shape-identity/full-frequency embedding and a corruption embedding, $e_L = [e_F, e_C]$. At test time, the parameters of both decoders are frozen and only the embeddings are optimized to reproduce the observed low-frequency SDF; the optimized full-frequency embedding is decoded to produce the sharpened surface. The paper reports that this pipeline lowers Chamfer distances and raises normal consistency relative to the compared methods on ShapeNet, ABC, and ScanNet reconstructions.","pith_inferences":["A natural extension is to apply the same frequency-consolidation idea to other implicit fields, such as occupancy or unsigned distance functions, where the same spectral-bias problem appears; nothing in the method is strictly tied to signed distance.","Because the synthetic training degradation is produced by FFT magnitude truncation, the prior's success on real reconstructions likely depends on how well that degradation mimics the spectral bias of neural networks; a reader should treat cross-domain results as the stronger evidence.","The method could be combined with retrieval-based shape completion: the optimized full-frequency embedding sits in a semantic latent space, so it could serve as a shape descriptor for matching or interpolation rather than only for sharpening.","Test-time self-reconstruction requires hundreds of optimization iterations per shape; an amortized predictor of the full-frequency embedding could remove this cost while retaining the prior."],"forward_implications":["The paper claims that an existing low-frequency SDF can be sharpened as a post-processing step by optimizing an embedding against it, without retraining the decoders.","The paper claims the prior transfers to unobserved frequency bands, since it sharpens reconstructions produced by sparse-point-cloud methods such as NeuralTPS and OnSurf.","The paper claims the recovered surfaces are not only sharper but also more complete, because the full-frequency supervision comes from watertight Poisson reconstructions.","The paper claims the approach works on CAD models with sharp edges, where it reports lower mean and variance of Chamfer distance than the compared methods.","The paper claims scene-level objects segmented from ScanNet scans are sharpened using priors learned from ShapeNet classes."],"supporting_citations":[{"why":"Supplies the auto-decoding formulation used for test-time embedding optimization and the decoder architecture.","marker":"(Park et al. 2019)"},{"why":"Provides the FFT-based Poisson surface reconstruction used to decompose shapes into frequency components.","marker":"(Peng et al. 2021)"},{"why":"NeuralPull converts point clouds and meshes into SDFs so the prior can accept different observation types.","marker":"(Ma et al. 2021)"},{"why":"Generates the sparse-point-cloud reconstructions used as unseen low-frequency test inputs and provides the ShapeNet test split.","marker":"(Chen, Han, and Liu 2023)"},{"why":"Provides the OnSurf sparse reconstruction baseline and its outputs as low-frequency test inputs.","marker":"(Ma et al. 2022)"},{"why":"Provides the ABC dataset train/test split and the Points2Surf baseline for CAD evaluation.","marker":"(Erler et al. 2020)"},{"why":"Supplies Scan2CAD alignment and ground-truth shapes for ScanNet scene evaluation.","marker":"(Avetisyan et al. 2019)"},{"why":"Supplies the ShapeNet dataset used for training and evaluation.","marker":"(Chang et al. 2015)"}],"fun_headline_variants":["Frequency prior sharpens neural 3D shapes","Sharpening 3D shapes via frequency consolidation","Recovering high-freq details in neural SDFs","Test-time embedding recovery sharpens 3D shapes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that low-frequency observations made by zeroing high-frequency FFT magnitudes of a Poisson-reconstructed occupancy field are representative of the low-frequency artifacts that real neural SDFs acquire from sparse point clouds or multi-view images; if that match fails, the learned prior would fix synthetic truncation rather than real spectral bias.","fun_headline_variants_meta":{"raw":{"variants":["Frequency prior sharpens neural 3D shapes","Sharpening 3D shapes via frequency consolidation","Recovering high-freq details in neural SDFs","Test-time embedding recovery sharpens 3D shapes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000237,"raw_usage":{"total_tokens":1547,"prompt_tokens":1024,"completion_tokens":523,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":460}},"tokens_in":640,"tokens_out":523,"duration_ms":5539,"temperature":1.0,"reasoning_tokens":460,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:55:17.870091+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the trained prior on low-frequency SDFs created by a different degradation mechanism than FFT truncation, for example SDFs learned by an MLP with spectral bias from sparse point clouds, and compare reconstruction error before and after sharpening; if the gain over the unsharpened input is much smaller than on FFT-truncated test pairs, the prior has learned to invert its own training degradation rather than to consolidate frequency coverage in general.","supporting_citations":[{"cited_title":"J.; Florence, P.; Straub, J.; Newcombe, R.; and Lovegrove, S","cited_arxiv_id":null,"evidence_quote":"Supplies the auto-decoding formulation used for test-time embedding optimization and the decoder architecture."},{"cited_title":"M.; Liao, Y.; Niemeyer, M.; Pollefeys, M.; and Geiger, A","cited_arxiv_id":null,"evidence_quote":"Provides the FFT-based Poisson surface reconstruction used to decompose shapes into frequency components."},{"cited_title":"J.; and Wimmer, M","cited_arxiv_id":null,"evidence_quote":"Provides the ABC dataset train/test split and the Points2Surf baseline for CAD evaluation."},{"cited_title":"X.; and Niessner, M","cited_arxiv_id":null,"evidence_quote":"Supplies Scan2CAD alignment and ground-truth shapes for ScanNet scene evaluation."}],"review_version":1}