{"id":"82908a46-1d3e-4160-b2e2-86ae4d4399c7","arxiv_id":"2505.02005","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A mixture-of-experts NeRF with hash-based gating and heterogeneous grid-resolution experts improves large-scale scene rendering accuracy and training efficiency over prior decomposition-based NeRFs.","lead":"Switch-NeRF++ trains large-scale neural radiance fields using a learned mixture of hash-based experts with different grid resolutions, replacing hand-crafted scene decomposition with an end-to-end gating network. It reports faster training and rendering while improving or matching rendered image quality on city-scale benchmarks, including a 6.5 square kilometer urban scene.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unqualified 'state-of-the-art' accuracy claim is contradicted by the paper's own VastGaussian comparison (Table 14), and the headline numbers use homogeneous rather than the proposed heterogeneous experts.","rationale":"The paper's core technical contribution is plausible: replacing MLP experts and gating with hash-based experts and a hash-based gate is a natural and likely effective way to scale Switch-NeRF, and the ablations (Homo vs. Heter on Block-NeRF and UrbanBIS, efficient hash dispatching, training-time comparisons) give real support for the efficiency and for the heterogeneous design. The reader's CONDITIONAL verdict is therefore reasonable. The additional load-bearing concern I identify is narrower but decisive for one part of the central claim: the unqualified 'state-of-the-art' accuracy assertion conflicts with the paper's own Table 14, where VastGaussian beats Switch-NeRF++ on a majority of scenes under a protocol the authors themselves adopt. The paper admits this in the prose but does not carry the caveat into the abstract or contributions. A second, related gap is that the headline Mega-NeRF and MatrixCity results use homogeneous experts, so the proposed heterogeneous mixture is not what generates the main SOTA-style numbers. These are corrigible presentation and evaluation issues rather than fundamental flaws in the method; conditional acceptance with mandatory qualification and ideally additional heterogeneous-expert results on Mega-NeRF is the right outcome. I do not see a reason to reject or to move away from the reader's conditional verdict.","tokens_in":38611,"tokens_out":7538,"duration_ms":76488,"concrete_test":"Take Table 14's VastGaussian-aligned protocol and count the number of Mega-NeRF scenes on which Switch-NeRF++ has the highest PSNR. If VastGaussian wins a majority of scenes, as the current table shows, the abstract's unqualified 'state-of-the-art' claim is false and must be restated; additionally, rerun the Mega-NeRF benchmark with heterogeneous hash experts rather than the same-expert setting to determine whether the core HMoHE contribution preserves the reported accuracy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and contributions claim 'state-of-the-art scene rendering accuracy' without qualification. Under the paper's own VastGaussian-aligned evaluation in Table 14, Switch-NeRF++ has the best PSNR on only Sci-Art; VastGaussian is higher on Building, Rubble, Residence, and Campus. The text concedes this ('While Switch-NeRF++ underperforms Vast-Gaussian on some scenes'), so the unqualified SOTA claim is internally inconsistent with the reported evidence. Compounding this, the implementation details state that the Mega-NeRF and MatrixCity experiments use the same hash experts, i.e., the homogeneous variant, while the heterogeneous HMoHE design is evaluated only on Block-NeRF and UrbanBIS. The headline accuracy results on the main benchmark are therefore not produced by the paper's core contribution. This does not invalidate the architectural idea or the reported efficiency gains, but it means the central claim as stated is not supported by the paper's own numbers and must be substantially qualified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Switch-NeRF++, a mixture-of-experts framework for large-scale NeRF. It replaces the MLP gating network and MLP experts of the authors' previous Switch-NeRF with a hash-based gating network and a set of hash-encoding experts with different resolution ranges ('heterogeneous mixture of hash experts', HMoHE). The gating network is trained end-to-end with a photometric loss and a balance loss, and each 3D point is dispatched to a single expert. Experiments are reported on Mega-NeRF, Block-NeRF, an adapted UrbanBIS scene covering more than 6.5 km^2, and MatrixCity. The paper claims state-of-the-art rendering accuracy, an 8x training acceleration, and a 16x rendering acceleration over Switch-NeRF.","tokens_in":38731,"tokens_out":7525,"duration_ms":69007,"significance":"The proposed architecture addresses a real limitation of prior large-scale NeRFs: hand-crafted decomposition rules and homogeneous expert capacity. If the heterogeneous design is validated on the main benchmarks, the paper will make a useful contribution. The hash-based gating and the efficient hash-dispatching implementation are thoughtfully engineered, and the ablations (gating design, expert count, Top-1 vs. Top-2, hash capacity, Gumbel selection) are extensive. The adaptation of a 6.5 km^2 UrbanBIS scene is also valuable for benchmarking scalability. However, as written, the manuscript does not support its central claim: the heterogeneous experts are evaluated only on Block-NeRF and UrbanBIS, while the headline Mega-NeRF and MatrixCity results are obtained with homogeneous experts. In addition, the unqualified 'state-of-the-art' accuracy claim is contradicted by the paper's own VastGaussian comparison in Table 14. These issues are fixable with re-evaluation or careful qualification, so the paper merits a major revision rather than rejection.","major_comments":[{"comment":"The implementation details state that 'For the Mega-NeRF and Matrixcity datasets, we use the same hash experts.' The headline accuracy and efficiency results in Tables 1 and 3 therefore come from the homogeneous variant, not from the proposed HMoHE. The central claim that Switch-NeRF++ 'learns heterogeneous decomposition and heterogeneous NeRFs' is not tested on the primary Mega-NeRF benchmark; the heterogeneous results appear only in Table 2 (Block-NeRF and UrbanBIS). Either the experiments need to be rerun with heterogeneous experts on Mega-NeRF and MatrixCity, or the paper must explicitly state that the headline benchmark uses a homogeneous variant and reframe the core claim accordingly.","section":"Section 4.2, Tables 1 and 3"},{"comment":"The unqualified claim of 'state-of-the-art scene rendering accuracy' is inconsistent with the paper's own VastGaussian-aligned results. In Table 14, VastGaussian has higher PSNR than Switch-NeRF++ on Building (23.50 vs. 22.41), Rubble (26.92 vs. 26.02), Residence (24.25 vs. 24.00), and Campus (26.00 vs. 24.37), with Switch-NeRF++ higher only on Sci-Art (27.40 vs. 26.81). The text correctly concedes that 'Switch-NeRF++ underperforms Vast-Gaussian on some scenes,' contradicting the abstract and contribution statements. The accuracy claim should be qualified, for example as state-of-the-art among NeRF/ray-based large-scale methods or as competitive with large-scale Gaussian-splatting methods.","section":"Abstract, Section 4.4, Table 14"},{"comment":"The claimed 8x training acceleration is based on Switch-NeRF++* trained for 50k steps (5.5h) compared with Switch-NeRF trained for 500k steps (43h). This is not a matched-quality comparison: on PSNR, the 50k model is lower than Switch-NeRF on Building (21.27 vs. 21.54), Sci-Art (25.65 vs. 26.52), and Campus (23.54 vs. 23.62). The 'state-of-the-art results with 5.5h' statement in the table caption is therefore not supported on all scenes and metrics. To substantiate the acceleration claim, the authors should report iso-quality or matched-budget comparisons, such as Switch-NeRF at the same number of steps or the time required to match Switch-NeRF's final PSNR.","section":"Table 1, Section 4.3"}],"minor_comments":[{"comment":"The balance loss is written as 'Lb = nPfigi' with undefined symbols and an unexplained 'minimized as 1' statement; please state the standard form n * sum(f_i * p_i) and define f_i and p_i clearly.","section":"Section 3.3, Eq. (5)"},{"comment":"The pseudocode writes 'G = Softmax(fg)', which is inconsistent with Eq. (2), where the gate values are obtained by applying Softmax to MLP(Hash(x)). Please fix the pseudocode to match the main text.","section":"Algorithm 1, line 18"},{"comment":"The table title contains the typo 'PNSR'; it should be 'PSNR'.","section":"Table 3"},{"comment":"In the pyramid heterogeneous expert design paragraph, 'Swhtch-NeRF++' should be 'Switch-NeRF++'.","section":"Section 3.2"},{"comment":"The abbreviations 'Homo.' and 'Heter.' are used in the table but are not defined in the caption or table header; please define them.","section":"Table 2"},{"comment":"The text attributes the SSIM/LPIPS gap to the SSIM loss used by Gaussian-splatting methods and to rasterization differences. This is plausible, but it should be supported by an ablation in which an SSIM loss is added to Switch-NeRF++, or explicitly stated as a hypothesis rather than as an explanation.","section":"Section 4.4, Table 14 discussion"},{"comment":"The point-cloud decomposition visualizations are compelling but qualitative. A numeric summary of the correlation between density gradients and expert resolutions would strengthen the claim that the heterogeneous experts learn scale-adaptive decompositions.","section":"Figures 10 and 11"}],"recommendation":"major_revision","confidential_remarks":"This manuscript is an extension of the authors' ICLR paper, Switch-NeRF. The new material (hash-based gating, heterogeneous hash experts, efficient hash dispatching, and the UrbanBIS benchmark) is sufficient for a journal paper if the evaluation is completed. However, the present submission's central claim is not supported by the reported experiments, because the heterogeneous contribution is not evaluated on the main Mega-NeRF or MatrixCity benchmarks and because the unqualified state-of-the-art claim is contradicted by the paper's own VastGaussian comparison. I recommend major revision, with the expectation that either the experiments are completed with heterogeneous experts or the claims are substantially reworded."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something real: it replaces Switch-NeRF's MLP gating with a hash-based gating network and swaps homogeneous MLP experts for heterogeneous multi-resolution hash experts with different resolution ranges. That combination is new, and the efficiency story holds up—8x training and 16x rendering speedups over Switch-NeRF are plausible given the architecture and are reported on the same hardware. The UrbanBIS adaptation gives a genuinely large 6.5 km² testbed, which is a useful contribution in itself.\n\nThe core idea is well supported where it is actually tested. On Block-NeRF and UrbanBIS, the heterogeneous version consistently beats the homogeneous version across PSNR, SSIM, and LPIPS, and the density-gradient analysis in Fig. 11 shows the gating is learning the intended scale-based decomposition. The authors also deserve credit for conceding in the text that they underperform VastGaussian on some scenes.\n\nThe soft spots are real but not fatal. The abstract and contributions claim \"state-of-the-art scene rendering accuracy\" without qualification, which is internally inconsistent with Table 14: VastGaussian wins on four of five scenes under the aligned evaluation. That needs to be qualified, e.g., \"among NeRF-style methods\" or with the GS numbers clearly separated. Second, the Mega-NeRF and MatrixCity experiments use homogeneous experts, so the headline benchmark does not test the paper's central contribution—the heterogeneity. The heterogeneous design is only validated on Block-NeRF and UrbanBIS. That is not a flaw in the architecture, but it means the paper's own numbers on the main benchmark do not demonstrate the claimed benefit. Third, there are no error bars and the code is not released, which is minor for a systems paper but worth noting.\n\nThe central argument holds: learnable heterogeneous decomposition improves quality and efficiency over both hand-crafted decomposition and homogeneous alternatives. The issues are in the claims and the experimental framing, not in the method. This deserves peer review, not rejection. A serious referee should ask for a qualified SOTA statement, an explicit note that heterogeneity is only evaluated on two datasets, and ideally code release and error bars.\n\nI would bring this to a reading group and cite it in work on large-scale scene representation. Send it to review.","headline":"Well-engineered extension that delivers real efficiency gains and a useful heterogeneous hash-expert design, but the unqualified SOTA claim and the homogeneous-only main benchmark overstate the accuracy case.","tokens_in":39376,"tokens_out":2469,"would_cite":true,"duration_ms":26425,"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":"A learnable hash gating network that splits a city-scale scene into scale-appropriate hash experts improves rendering quality while training 8× faster and rendering 16× faster than predecessor methods.","keywords":["neural radiance fields","large-scale scene modeling","mixture of experts","learnable scene decomposition","hash grid encoding","heterogeneous scene representation","novel view synthesis","scalable rendering"],"falsifier":"Train the heterogeneous and homogeneous versions on a large scene whose detail is spread uniformly across scales, or whose finest structures sit below the smallest expert resolution; if the heterogeneous advantage vanishes or the gate shows no resolution-gradient correlation, the fixed resolution pyramid, not learnable decomposition, is carrying the result.","tokens_in":38313,"feed_emoji":"🏙️","tokens_out":9422,"duration_ms":91033,"temperature":0.7,"pith_summary":"This paper tries to establish that scene decomposition for large-scale neural radiance fields should be learned, not hand-designed, and that the learned experts should differ from one another in grid resolution. It proposes a single end-to-end framework in which a hash-based gating network routes each 3D point to one of several hash-grid experts, each covering a different resolution range. The paper claims this heterogeneous mixture captures the varied geometry and appearance of real scenes, and that it reaches higher rendering accuracy than prior large-scale NeRFs on datasets up to 6.5 km² while cutting training time by 8× and rendering time by 16× relative to its predecessor.","feed_headline":"Learnable scene splits make city-scale 3D rendering 16× faster","feed_subtitle":"A hash-based gate assigns each region its own grid resolution, beating hand-crafted splits in accuracy and speed.","key_machinery":"The load-bearing object is the resolution pyramid of hash experts: eight multi-resolution hash encodings whose base resolutions are logarithmically spaced from 16 to 512 and whose desired resolutions from 2048 to 16384, so experts range from coarse to fine. The mechanism that makes routing learnable is the gate-value multiplication $\\hat E(x)=G(x)_i E_i(x)$, which sends gradient through the discrete selection, supported by a gating balance loss $L_b = n\\sum_i f_i g_i$ that encourages balanced use of experts. Around this core sit a hash-based gating network that computes gate values by looking up a hash feature and passing it through a small MLP, and a fused hash-dispatching implementation that avoids explicit input reorganization and capacity padding.","core_discovery":"The central claim, stated in the paper's own terms, is that a large-scale scene is best represented by a Heterogeneous Mixture of Hash Experts (HMoHE): a set of hash-based NeRF experts whose grid resolutions form a pyramid from coarse to fine, together with a trainable hash gate that learns which resolution each scene region needs. The paper argues that the discrete Top-1 expert choice becomes optimizable because the selected expert's output is multiplied by the gate value, so gradients flow back into the gating network, and a balance loss keeps experts from collapsing. It presents evidence that on urban scenes the gate assignments correlate with local density gradients: low-resolution experts receive low-gradient regions and high-resolution experts receive high-gradient regions, which the homogeneous baseline does not show. The paper's quantitative claim is state-of-the-art rendering accuracy on Mega-NeRF, Block-NeRF, UrbanBIS, and MatrixCity scenes, with 8× faster training and 16× faster rendering than the previous Switch-NeRF.","pith_inferences":["The fixed resolution ladder is a strong prior rather than a learned quantity; scenes whose feature-scale distribution falls outside the chosen 16-to-16384 range would likely need the ladder adapted per scene, a testable extension the paper does not explore.","Since the gate learns only from final rendered color, the decomposition may track appearance as much as geometry; adding depth or normal supervision could produce partitions that are more geometrically coherent and might help in texture-poor regions.","The same sparsely gated design with heterogeneous hash resolutions could be carried over to other grid-based 3D representations such as Gaussian splatting, offering those methods a learnable alternative to manual cell division.","The reported correlation between expert resolution and local density gradient suggests a 'scale field' could be predicted directly, possibly reducing the number of experts or improving initialization."],"forward_implications":["Large-scale scenes can be reconstructed end-to-end with no manual partitioning, so the same training pipeline transfers to new cities or districts without hand-tuned decomposition rules.","Because only one expert is active per 3D point, adding experts raises capacity without proportionally raising per-point computation, making the method more scalable by expert count.","The resolution pyramid lets flat ground and detailed facades be represented by the same network at different grid scales, which is why the heterogeneous version beats the homogeneous one on every reported metric.","At 6.65 s per 1,152 × 864 image with less than half the memory of prior methods, the model becomes practical for interactive or repeated rendering of very large scenes.","Training for 50k steps (5.5 h) already reaches state-of-the-art accuracy on most Mega-NeRF metrics, so the method offers a direct compute-accuracy trade-off for users with limited GPU budgets."],"supporting_citations":[{"why":"It supplies the volume-rendering model and rendering loss that the whole framework optimizes.","marker":"[1]"},{"why":"It is the hand-crafted distance-clustering baseline whose accuracy and efficiency the paper compares against.","marker":"[2]"},{"why":"It is the street-block decomposition baseline that motivates the need for learnable partitioning.","marker":"[3]"},{"why":"It is the predecessor method whose homogeneous MLP experts are replaced and whose training and rendering times are used for the 8× and 16× speedup claims.","marker":"[4]"},{"why":"It provides the multi-resolution hash encoding used both in the gating network and in each hash expert.","marker":"[5]"},{"why":"It supplies the adapted 6.5 km² urban scene used to demonstrate very-large-scale scalability.","marker":"[10]"},{"why":"It is the source of the sparsely gated mixture-of-experts design, including the gate-value multiplication that makes routing trainable.","marker":"[11]"},{"why":"It provides the gating balance loss that stabilizes expert utilization in the joint optimization.","marker":"[64]"},{"why":"It provides the optimized MoE dispatching kernel that the paper's fused hash dispatching is compared against for time and memory.","marker":"[73]"},{"why":"It supplies the MatrixCity aerial dataset used to test generalization to a different type of large-scale scene.","marker":"[76]"}],"fun_headline_variants":["Hash experts learn scene splits for 16x faster city-scale rendering","Trainable gates pick grid resolutions for scalable NeRF scenes","Heterogeneous hash experts speed up large-scale NeRF by 8-16x","Switch-NeRF++ trains 8x faster with learnable scene decomposition","City-scale NeRF gets 16x faster rendering via learned experts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a hand-set ladder of coarse-to-fine hash-grid resolutions is the right axis for separating a scene's parts, and that image-reconstruction error plus one balance term can teach the gate that separation without any geometric guide.","fun_headline_variants_meta":{"raw":{"variants":["Hash experts learn scene splits for 16x faster city-scale rendering","Trainable gates pick grid resolutions for scalable NeRF scenes","Heterogeneous hash experts speed up large-scale NeRF by 8-16x","Switch-NeRF++ trains 8x faster with learnable scene decomposition","City-scale NeRF gets 16x faster rendering via learned experts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000411,"raw_usage":{"total_tokens":2201,"prompt_tokens":1091,"completion_tokens":1110,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":707,"completion_tokens_details":{"reasoning_tokens":1015}},"tokens_in":707,"tokens_out":1110,"duration_ms":6838,"temperature":1.0,"reasoning_tokens":1015,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:03:58.141658+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the heterogeneous and homogeneous versions on a large scene whose detail is spread uniformly across scales, or whose finest structures sit below the smallest expert resolution; if the heterogeneous advantage vanishes or the gate shows no resolution-gradient correlation, the fixed resolution pyramid, not learnable decomposition, is carrying the result.","supporting_citations":[{"cited_title":"Mega-nerf: Scal- able construction of large-scale nerfs for virtual fly-throughs,","cited_arxiv_id":null,"evidence_quote":"It is the hand-crafted distance-clustering baseline whose accuracy and efficiency the paper compares against."},{"cited_title":"Block-nerf: Scalable large scene neural view synthesis,","cited_arxiv_id":null,"evidence_quote":"It is the street-block decomposition baseline that motivates the need for learnable partitioning."},{"cited_title":"Switch-neRF: Learning scene decomposition with mixture of experts for large-scale neural radiance fields,","cited_arxiv_id":null,"evidence_quote":"It is the predecessor method whose homogeneous MLP experts are replaced and whose training and rendering times are used for the 8× and 16× speedup claims."},{"cited_title":"Instant neural graphics primitives with a multiresolution hash encoding,","cited_arxiv_id":null,"evidence_quote":"It provides the multi-resolution hash encoding used both in the gating network and in each hash expert."},{"cited_title":"Ur- banbis: a large-scale benchmark for fine-grained urban building instance segmentation,","cited_arxiv_id":null,"evidence_quote":"It supplies the adapted 6.5 km² urban scene used to demonstrate very-large-scale scalability."},{"cited_title":"Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,","cited_arxiv_id":null,"evidence_quote":"It is the source of the sparsely gated mixture-of-experts design, including the gate-value multiplication that makes routing trainable."},{"cited_title":"Gshard: Scaling giant models with conditional computation and automatic sharding,","cited_arxiv_id":null,"evidence_quote":"It provides the gating balance loss that stabilizes expert utilization in the joint optimization."},{"cited_title":"Ma- trixcity: A large-scale city dataset for city-scale neural rendering and beyond,","cited_arxiv_id":null,"evidence_quote":"It supplies the MatrixCity aerial dataset used to test generalization to a different type of large-scale scene."}],"review_version":1}