{"id":"1d950378-d9b4-4787-b551-22eb11619c6d","arxiv_id":"2501.12637","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"DWTNeRF improves few-shot INGP rendering by weighting low-frequency wavelet sub-bands higher in the loss and using multi-head attention across color and density, beating CombiNeRF by up to 0.26 dB PSNR on 3-shot LLFF.","lead":"DWTNeRF adds a wavelet-domain loss that prioritizes low image frequencies, plus attention-based color-density interactions, to a fast hash-encoding NeRF. It reports small but consistent gains over the INGP-based few-shot baseline CombiNeRF on 2-4 view benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The DW-loss ablation is confounded: it adds patch-based sampling and about 45% more training rays to the CombiNeRF baseline, so the reported 0.06 dB gain (Table 3) may reflect patch locality or ray budget rather than wavelet frequency weighting.","rationale":"After reading the full text, I view DWTNeRF as a plausible engineering contribution: CombiNeRF is a strong base, and the attention modules show consistent gains in Table 3. My concern is not that DWTNeRF fails, but that the evidence singled out as the method's novelty, the DWT loss, has not been isolated from two substantial confounds: patch-based sampling and a large increase in the number of training rays. The reader's weakest_assumption identified the patch-sampling confound; I agree and refine it by also making the extra-ray budget explicit. This matters because an orthogonal DWT preserves L2 distances, so the weighted subband loss differs from plain MSE only in the frequency weighting; a patch-MSE control is exactly the experiment needed. The attention module is not the focus of this concern: Table 3 suggests it is the larger contributor, and the mi-MLP comparisons support the claim that INGP is sensitive to architectural changes. Thus the verdict remains conditional: the method may work, but the specific wavelet mechanism is not yet established. Other issues, such as missing error bars and hyperparameter tuning on the same benchmark, are real but less directly testable and less damaging to the stated contribution.","tokens_in":14166,"tokens_out":6887,"duration_ms":74811,"concrete_test":"Re-run the +DW ablation with the wavelet term replaced by plain L2 on the same 192x192 ground-truth patches, keeping patch locations, the 36,864-ray budget, the every-10-iteration schedule, the 5K cutoff, and all CombiNeRF settings unchanged. Evaluate on the same 3-shot LLFF scenes with per-scene PSNR/SSIM/LPIPS. If patch-MSE reaches or exceeds 20.25 PSNR, the wavelet decomposition contributes nothing beyond patch-based MSE; if patch-MSE stays more than 0.05 dB below 20.25, the frequency weighting is empirically supported. A secondary control with the same number of extra uniformly random rays (no patch) would further separate patch locality from the extra-ray effect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 3 isolates the DW loss by comparing CombiNeRF (-freq mask.) at 20.19 PSNR with DWTNeRF (+DW.) at 20.25, but the +DW configuration changes three things at once: (1) it renders an extra 192x192 patch (36,864 rays) every 10 iterations until iteration 5K; (2) that patch consists of adjacent rays rather than random rays; and (3) it applies the weighted wavelet subband MSE to those patches. No control uses plain MSE on the same patches with the same extra-ray budget, and no control adds the same number of extra random rays. Because the Haar/Daubechies DWT is orthogonal, the DW loss is mathematically equivalent to ordinary MSE with frequency-dependent subband weights; the only thing separating it from a patch-MSE control is the weighting. Without that control, the 0.06 dB increment in Table 3, and the larger headline gaps over CombiNeRF in Table 1, can be explained by patch-based sampling (a known variance-reduction technique) or simply by the optimizer seeing about 45% more gradients. The paper's own limitation statement concedes that the standard photometric MSE remains in the objective, so the proposed low-frequency emphasis is a mild reweighting rather than a mechanism that has been cleanly validated. This is load-bearing because the paper's title and central claim attribute the few-shot improvement to the wavelet decomposition.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DWTNeRF, a few-shot NeRF method built on Instant-NGP and CombiNeRF. Its two contributions are (1) a Discrete Wavelet (DW) loss that decomposes rendered and ground-truth patches into sub-bands and weights low frequencies more heavily, applied with patch-based ray sampling, and (2) multi-head attention modules at the input and output levels of the color/density branches to model cross-branch interactions without modifying the MLP architecture. The method is evaluated on LLFF (2/3/6/9 views) and NeRF-Synthetic (4 views), reporting consistent but small improvements over CombiNeRF, and ablations are provided for the DW loss, attention modules, and wavelet choice.","tokens_in":14442,"tokens_out":2596,"duration_ms":28159,"significance":"If the improvements are attributable to the proposed components, the paper would make a useful contribution to few-shot INGP by showing that frequency-domain supervision can be adapted to hash-based encodings and that attention-based cross-branch interactions avoid the fragility of architectural changes. The paper also provides a solid negative result that FreeNeRF-style frequency masking and mi-MLP-style modifications degrade or fail to improve INGP. However, the central claim that the DW loss drives the improvement is not cleanly established because the +DW ablation changes patch sampling and ray budget simultaneously, and the reported gains are small with no repeated-run statistics. The significance is therefore conditional on additional controlled experiments.","major_comments":[{"comment":"The ablation isolating the DW loss is confounded. Comparing 'CombiNeRF (-freq mask.)' at 20.19 PSNR with 'DWTNeRF (+DW.)' at 20.25 changes three things at once: the +DW configuration renders an extra 192x192 patch (36,864 rays) every 10 iterations until iteration 5K, those rays are adjacent rather than random, and the loss applied to the patch is a weighted wavelet MSE. Because the Haar/Daubechies DWT is orthogonal, the DW loss is mathematically equivalent to ordinary MSE with frequency-dependent subband weights; hence the only component unique to the DW loss is the subband weighting. Without a control that applies plain MSE to the same patches with the same extra-ray schedule, the 0.06 dB gain in Table 3, and the larger gaps over CombiNeRF in Table 1, can be explained by patch-based sampling or simply by the optimizer receiving roughly 45% more gradients. This control must be added before the paper's title-level claim about the wavelet loss can be accepted.","section":"Table 3 and Section 4.2 (DW loss, patch-based rendering)"},{"comment":"No repeated runs, seeds, or error bars are reported, yet the headline improvements are very small: e.g., 3-shot LLFF PSNR improves from 20.12 to 20.38, and 4-shot NeRF-Synthetic PSNR from 19.15 to 19.25. Given that the ablation rows in Table 3 differ by 0.01-0.06 dB PSNR and 0.001-0.007 SSIM, single-run results cannot establish that these differences are not optimization noise. The authors should report mean and standard deviation over at least three seeds for the main tables and the key ablations, or otherwise justify why single runs are sufficient.","section":"Tables 1-5 and Section 5.3-5.4"},{"comment":"The DW loss depends on several per-dataset hyperparameters (lambda_LL=0.4 vs 0.04, other subband weights 0.2 vs 0.02, K_DW=10 vs 150, T_DW=5K, patch size 192, and 2 vs 1 attention heads), and no sensitivity analysis is provided. Since the final result is only 0.1-0.3 dB above the baseline, it is unclear whether the reported configuration is a robust operating point or a carefully selected one. A sensitivity study over the subband weights and the loss schedule would substantially strengthen the empirical claim.","section":"Tables 1, 3, 5 and Section 5.1 (hyperparameters)"}],"minor_comments":[{"comment":"The text contains a typo: 'cross-banch interactions' should be 'cross-branch interactions'.","section":"Section 5.4"},{"comment":"The caption says 'much noiser' and 'noiser' should be 'noisier'.","section":"Figure 7 caption"},{"comment":"The venue column formats some entries inconsistently (e.g., 'ARXIV‘22', 'ICML ‘23', 'CVPR ‘23'); these should be normalized.","section":"Table 1 venue column"},{"comment":"The sentence 'This is a concern in applying the DW loss, which we will explain later' is vague; the concern about sub-band size could be stated explicitly at that point.","section":"Section 4.2"},{"comment":"The wavelet comparison reports differences of 0.01-0.02 dB PSNR without any variance information; the conclusion that the Haar wavelet is sufficient would be more convincing with repeated runs.","section":"Section 5.4 and Table 5"}],"recommendation":"major_revision","confidential_remarks":"The key issue is the confound in the DW-loss ablation. If the authors can show that patch-based MSE with the same ray budget does not reproduce the +DW gain, the paper's central claim is credible. If not, the contribution reduces to a small empirical gain from patch sampling, which is not sufficient for the claimed novelty. I would ask for the patch-MSE control and repeated-seed statistics as mandatory revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. First, this is a genuinely incremental but competent methods paper: it builds on CombiNeRF, adds a wavelet-domain loss that reweights low frequencies, and adds multi-head attention at the inputs/outputs of INGP's two MLP branches. The results are modest—0.1–0.26 dB PSNR over CombiNeRF on LLFF 2/3-shot, and 0.1 dB on NeRF-Synthetic 4-view—but the paper is clear about what it does and does not claim. Second, the ablation that is supposed to validate the DWT loss is confounded, and the confound is load-bearing. The +DW configuration adds patch-based ray sampling and about 37k extra rays every 10 iterations up to iteration 5k, and there is no control that uses plain MSE on the same patches or adds the same extra random rays. Since Haar/Daubechies DWT is orthogonal, the DW loss is just MSE with per-subband weights, so the only thing separating it from a patch-MSE baseline is the weighting. Without that control, the 0.06 dB gain in Table 3 could easily come from patch locality (a known variance-reduction trick) or from the optimizer simply seeing more gradients. The paper's own limitation statement concedes that MSE stays in the objective, which makes the low-frequency emphasis a mild reweighting rather than a validated mechanism.\n\nWhat I credit: the negative results are genuinely useful. The paper shows that removing FreeNeRF-style frequency masking from CombiNeRF helps (20.12→20.19), and that mi-MLP's residual and element-wise cross-branch additions degrade INGP—that is a real, reproducible observation about transferability. The attention-based interactions, by contrast, are evaluated more cleanly: they change the model without touching the MLP layers, and the ablation shows a consistent improvement (20.19→20.36 for full cross-branch). I'd have liked error bars or multiple seeds, because 0.1–0.3 dB PSNR differences can easily flip with one seed, but the attention story is plausible.\n\nThe soft spots, in order of seriousness: the missing patch-MSE control; no variance estimates; hyperparameters (sub-band weights, attention heads, patch size, loss interval) tuned on the same benchmark scenes with no sensitivity analysis. The wavelet-family ablation (Haar vs db2 vs db3) shows no meaningful differences, which is consistent with their low-frequency-prioritization story but could also mean the loss is insensitive to the decomposition. None of these are fatal to the whole paper—the framework does improve over CombiNeRF—but they do mean the central mechanism attribution is not established.\n\nWho it's for: people working on few-shot NeRF, especially on INGP/3DGS-adjacent representations, who want to see what transfers and what doesn't. It's not a rethink; it's a useful data point.\n\nRecommendation: send it to peer review, but require a patch-MSE control and multi-seed results before accepting the DW loss claim. The attention module and the negative results are worth publishing regardless.","headline":"Solid incremental few-shot NeRF paper whose headline DW-loss mechanism is confounded with patch sampling and extra rays—worth refereeing, but with a required control.","tokens_in":15060,"tokens_out":2527,"would_cite":false,"duration_ms":23877,"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":"Discrete-wavelet loss upweighting low frequencies, paired with attention-based cross-branch interactions, materially improves few-shot neural radiance fields on hash-encoding backbones.","keywords":["discrete wavelet transform","few-shot neural radiance fields","Instant-NGP","hash encoding","multi-head attention","cross-branch interactions","frequency regularization","novel view synthesis"],"falsifier":"Run the CombiNeRF baseline with patch-based ray sampling and the standard MSE loss (same patch size, same lazy schedule, no DWT) on 3-shot LLFF. If PSNR reaches or exceeds DWTNeRF's 20.25, the wavelet decomposition is not the active ingredient driving the reported gain.","tokens_in":13890,"feed_emoji":"🖼️","tokens_out":7908,"duration_ms":75464,"temperature":0.7,"pith_summary":"DWTNeRF sets out to make few-shot novel view synthesis work on the fast Instant-NGP-style hash-encoding backbone, where common few-shot tricks fail because the multi-resolution encoding neither maps cleanly to frequencies nor tolerates architectural changes. The key idea is a Discrete Wavelet loss that decomposes rendered and ground-truth views into LL, LH, HL and HH sub-bands and upweights the low-frequency LL term so early training focuses on structure before detail. On the 3-shot LLFF benchmark the paper reports 15.07% higher PSNR, 24.45% higher SSIM and 36.30% lower LPIPS than vanilla INGP, and a modest edge over the CombiNeRF baseline (20.38 vs 20.12 PSNR) that carries over to the 4-shot NeRF-Synthetic setting. It also argues that attention-based cross-branch interactions are the model-based modification that does not degrade INGP.","feed_headline":"Low-frequency wavelet loss sharpens few-shot NeRF rendering","feed_subtitle":"DWTNeRF beats vanilla hash-encoding NeRF by 15% in PSNR and edges CombiNeRF on 3-shot LLFF.","key_machinery":"The load-bearing object is the Discrete Wavelet (DW) loss, $L_{DW} = \\sum_{sb} \\lambda_{sb}\\| I_{sb} - I^{gt}_{sb}\\|_2^2$, computed on $192\\times192$ rendered patches with Haar or Daubechies wavelets, with $\\lambda_{LL}$ set higher than the other sub-band weights (0.4 vs 0.2 on LLFF). Because the DWT needs a 2D image, the loss forces patch-based ray sampling; it is applied lazily every 10 (LLFF) or 150 (NS) iterations and stopped at iteration 5K to keep the cost down. The second module is multi-head attention inserted at two points: on the concatenation of hash and view encodings before the color MLP, and again on the concatenated density/color outputs. It does this without adding MLP layers, which the paper argues is why it survives INGP's sensitivity to architectural change.","core_discovery":"The central claim is that frequency-domain supervision can substitute for frequency masking on hash-encoding radiance fields, and that color-density interactions can be learned without touching the MLP architecture. DWTNeRF realizes this by taking the two-dimensional discrete wavelet transform of rendered and ground-truth patches, weighting low-frequency sub-bands more heavily in the training objective, and passing hybrid position/view encodings plus concatenated density-color outputs through multi-head attention. In the paper's experiments this yields results that are competitive with state-of-the-art few-shot methods on 2- to 4-shot LLFF and 4-shot NeRF-Synthetic, with the largest gains in the most extreme few-shot regime and a clear margin over Vanilla INGP.","pith_inferences":["An untested but decisive check: run CombiNeRF with patch-based ray sampling and the ordinary MSE loss under the same lazy schedule; if PSNR reaches 20.25, the reported DW-loss gain is a sampling effect, not a frequency-prioritization effect.","The fact that Haar, db2 and db3 give nearly identical results suggests the specific frequency decomposition is not the active ingredient; a spatially structured loss over patches may be what helps.","The attention modules only handle input and output interactions; lightweight alternatives such as cross-attention or small graph layers might achieve the same INGP-compatible gains.","If low-frequency emphasis reduces hallucinated geometry, a similar wavelet-based supervision on depth or density maps could extend the benefit beyond color."],"forward_implications":["If the central claim holds, few-shot INGP can be improved by an image-space frequency-prioritizing loss that replaces the positional-encoding masking used by FreeNeRF.","The gains are concentrated at 2–4 training views and largely fade at 6–9 views, so the method targets extreme few-shot capture rather than general view synthesis.","Removing frequency masking from CombiNeRF already improves PSNR from 20.12 to 20.19, so part of the benefit comes from undoing an INGP-incompatible regularizer rather than from the wavelet loss alone.","Because the attention modules sit outside the MLP, the framework could be transplanted onto other hash-encoding renderers, and the paper suggests 3DGS and SDF representations as future targets."],"supporting_citations":[{"why":"Instant-NGP's multi-resolution hash encoding is the backbone whose few-shot brittleness motivates the method; everything is built on top of it.","marker":"[22]"},{"why":"CombiNeRF is the regularization baseline and base implementation that DWTNeRF extends with the DW loss and attention modules.","marker":"[2]"},{"why":"FreeNeRF introduced the frequency-masking idea that the DW loss is designed to replicate in an INGP-compatible way.","marker":"[39]"},{"why":"Supplies the multi-head attention operation used for input- and output-level cross-branch interactions.","marker":"[34]"},{"why":"Provides the orthogonal wavelet and filter-bank construction that defines the sub-band decomposition in the DW loss.","marker":"[30]"},{"why":"mi-MLP's residual and element-wise cross-branch architectures are the model-based baselines that the attention approach must beat.","marker":"[46]"},{"why":"LLFF is the primary few-shot benchmark dataset with 2/3/6/9-view protocols and the source of the headline 3-shot numbers.","marker":"[20]"},{"why":"NeRF-Synthetic provides the 4-shot benchmark and the definition of the NeRF rendering task the method addresses.","marker":"[21]"}],"fun_headline_variants":["Wavelet loss boosts few-shot NeRF by 15% PSNR","DWTNeRF: Wavelet domain wins in sparse-view rendering","Low-frequency wavelet loss lifts few-shot NeRF quality","DWTNeRF uses wavelets to cut NeRF overfitting in few-shot","Wavelet transform gives hash-encoding NeRF a boost in few-shot"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper never compares its wavelet loss against the ordinary pixel-difference loss computed on the same image patches, so the reported improvement could come from patch-based ray sampling itself rather than from the wavelet decomposition or the low-frequency weighting.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet loss boosts few-shot NeRF by 15% PSNR","DWTNeRF: Wavelet domain wins in sparse-view rendering","Low-frequency wavelet loss lifts few-shot NeRF quality","DWTNeRF uses wavelets to cut NeRF overfitting in few-shot","Wavelet transform gives hash-encoding NeRF a boost in few-shot"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000916,"raw_usage":{"total_tokens":3912,"prompt_tokens":905,"completion_tokens":3007,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":2913}},"tokens_in":521,"tokens_out":3007,"duration_ms":22381,"temperature":1.0,"reasoning_tokens":2913,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:57:56.694842+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the CombiNeRF baseline with patch-based ray sampling and the standard MSE loss (same patch size, same lazy schedule, no DWT) on 3-shot LLFF. If PSNR reaches or exceeds DWTNeRF's 20.25, the wavelet decomposition is not the active ingredient driving the reported gain.","supporting_citations":[{"cited_title":"Instant neural graphics primitives with a multires- olution hash encoding","cited_arxiv_id":null,"evidence_quote":"Instant-NGP's multi-resolution hash encoding is the backbone whose few-shot brittleness motivates the method; everything is built on top of it."},{"cited_title":"CombiNeRF: A Combination of Regularization Techniques for Few-Shot Neural Radiance Field View Synthesis","cited_arxiv_id":null,"evidence_quote":"CombiNeRF is the regularization baseline and base implementation that DWTNeRF extends with the DW loss and attention modules."},{"cited_title":"Freenerf: Im- proving few-shot neural rendering with free frequency reg- ularization","cited_arxiv_id":null,"evidence_quote":"FreeNeRF introduced the frequency-masking idea that the DW loss is designed to replicate in an INGP-compatible way."},{"cited_title":"Attention is all you need","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-head attention operation used for input- and output-level cross-branch interactions."},{"cited_title":"Wavelets and filter banks","cited_arxiv_id":null,"evidence_quote":"Provides the orthogonal wavelet and filter-bank construction that defines the sub-band decomposition in the DW loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"mi-MLP's residual and element-wise cross-branch architectures are the model-based baselines that the attention approach must beat."},{"cited_title":"Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar","cited_arxiv_id":null,"evidence_quote":"LLFF is the primary few-shot benchmark dataset with 2/3/6/9-view protocols and the source of the headline 3-shot numbers."},{"cited_title":"Srinivasan, Matthew Tancik, Jonathan T","cited_arxiv_id":null,"evidence_quote":"NeRF-Synthetic provides the 4-shot benchmark and the definition of the NeRF rendering task the method addresses."}],"review_version":1}