{"id":"0d5cc5e9-ea2b-42eb-9f67-44c07052c544","arxiv_id":"2608.05920","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"FlaRe combines per-primitive latent radiance descriptors on planar Gaussians with a shared decoder and hardware ray tracing, making rendering, secondary rays, editing, and mesh extraction work in one scene model.","lead":"Floating Radiance Networks (FlaRe) represents a 3D scene as a set of flat floating Gaussian shapes, each holding a small learned code for its local appearance, with a shared network decoding those codes into color and opacity.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central unification claim rests on untested generalization of per-primitive descriptors to secondary-ray and deformed queries; only qualitative evidence supports the distinctive capabilities.","rationale":"The reader's weakest_assumption precisely identifies the load-bearing premise: the per-primitive descriptor must define a local radiance field that remains valid when queries change (secondary rays, deformation, stylization). My analysis confirms this is the least secure link in the central claim, because the descriptor is trained only against photometric loss on primary camera rays and the entire evidence for the distinctive capabilities is qualitative. The ray-tracing and deformation claims are not supported by any numerical evaluation, and the limited capacity of the directional encoding (4-band positional encoding) makes extrapolation to novel secondary-ray directions particularly risky. I agree with the reader's CONDITIONAL verdict: the paper should be accepted only if the authors add quantitative validation of these capabilities, ideally using synthetic scenes with ground-truth references. A secondary concern, also noted by the reader, is the supplement's statement that hyperparameter selection is guided by validation/test PSNR, which may inflate the reported benchmark numbers; this affects the competitive-rendering-quality claim but is less central than the unvalidated descriptor generalization. The compactness claim based on primitive counts rather than memory or parameter budgets is another supporting issue, not the primary one. The paper is competently written and the rendering pipeline is well-specified, but the central unification claim currently rests on untested extrapolation, so the condition is appropriate.","tokens_in":52983,"tokens_out":10588,"duration_ms":101083,"concrete_test":"Render a synthetic scene with a reflective planar surface and a textured object, train FlaRe only on views that do not observe the object's reflection in the mirror, then render a novel viewpoint where the reflection is visible and compare PSNR/LPIPS in the reflected region against a path-traced ground truth. If the reflected-region PSNR drops by more than about 3 dB relative to directly visible regions, the descriptor does not extrapolate to secondary-ray directions and the central ray-tracing claim is unsupported. The same protocol can be adapted to deformation by bending a trained flat-textured proxy and comparing the render to a ground-truth render of the bent surface.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline contribution is a single learned representation that supports novel-view synthesis, recursive ray tracing, deformation, mesh extraction, and stylization without task-specific models. For this to hold, the per-primitive latent descriptor z_i must encode a local radiance function that remains valid when queried outside the training distribution: secondary rays after reflection or refraction approach primitives from directions that may not appear in any training view, and deformation changes the local frame and the (u,v) coordinates fed to the decoder, while z_i is left unchanged. The view-direction input uses only a 4-band positional encoding (24 dims) and the LUT local-coordinate resolution is capped at 64×64, so the network has limited capacity for high-frequency off-distribution directional variation. The optimization objective is purely a photometric loss on primary camera rays (Eq. 4 in §3.4), providing no constraint that the decoded color/opacity remains physically consistent under these novel queries. The paper offers only qualitative demonstrations for ray tracing (Fig. 2), deformation (Fig. 4), and stylization (Fig. 7), with no numerical error metrics for reflected/refracted regions, deformed-scene fidelity, or stylization accuracy. If those off-distribution queries produce large errors, the unified-scene-model claim collapses into a set of disconnected demos, reducing the contribution to a moderately competitive novel-view synthesis method.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Floating Radiance Networks (FlaRe), a scene representation in which the scene is composed of planar generalized Gaussian primitives, each carrying a 96-dimensional latent descriptor of a local 2D neural radiance field. A lightweight shared auto-decoder maps the descriptor, local surface coordinates, and viewing direction to color and opacity, with hardware-accelerated ray–triangle proxies used for intersection. The authors claim that this single representation supports high-quality novel-view synthesis as well as recursive ray tracing, primitive-level deformation, mesh extraction, and appearance stylization without task-specific scene models. Quantitative results are reported on Mip-NeRF360, Tanks and Temples, and Deep Blending (Table 2), with additional controlled experiments on LUT-Encoding and primitive reduction. The supplement contains detailed gradient derivations, architecture descriptions, and reproducibility notes.","tokens_in":53332,"tokens_out":4112,"duration_ms":42532,"significance":"If the unified-scene-model claim is substantiated, FlaRe would be a valuable step toward bridging neural radiance fields and conventional graphics pipelines, because the same explicit primitives would serve rendering, ray tracing, editing, and geometry extraction. The paper has notable strengths: it ships source code, provides thorough analytic gradient derivations in the supplement, includes a controlled, seeded comparison of LUT-Encoding against hash grids, and evaluates primitive reduction against two pruning baselines. However, the distinctive capabilities beyond novel-view synthesis are currently supported only by qualitative figures, and the main quantitative claim is weakened by a hyperparameter-selection procedure that appears to use test performance. The significance of the work therefore depends on additional validation that is not yet present in the manuscript.","major_comments":[{"comment":"The central claim that FlaRe supports recursive ray tracing, deformation, stylization, and mesh extraction within one untrained representation is supported only by qualitative images (Figs. 2, 4, 7, 8). The per-primitive descriptors z_i are optimized solely against a photometric loss on primary camera rays (main-text loss in §3.4 and supplement Eq. for L), so the validity of the decoded radiance under secondary-ray directions, deformed local frames, and stylized descriptors is an empirical claim that is never measured. I request numerical evaluations: for example, PSNR/SSIM or relative error in reflected/refracted regions against reference renderings, Chamfer distance or F-score for deformed geometry, CLIP score and structure-preservation metrics for stylization, and Chamfer/F-score for mesh extraction on a standard dataset. Without such measurements, the headline unification claim is not established and the contribution reduces to a competitive novel-view-synthesis method with disconnected demos.","section":"Graphics Capabilities; Ray-Tracing; Mesh modification"},{"comment":"The supplement states: 'Hyperparameter selection during tuning is guided by validation/test PSNR.' Because Table 2 reports test-set PSNR as the primary evidence of competitive NVS quality, the reported rankings may be optimistically biased toward the test set. This is load-bearing for the paper's second main claim. The authors should clarify whether the final numbers were selected on a held-out validation split or on the test split, report the selection protocol per scene, and ideally provide sensitivity of Table 2 to hyperparameter choices or standard errors over multiple training runs.","section":"Supplementary Material, Reproducibility, Hyperparameter search"},{"comment":"The recursive ray-tracing capability is not described at the algorithmic level. The main text and supplement detail only primary-ray rendering (ordered intersections, front-to-back compositing, loss function). There is no description of how secondary rays are generated, how material properties (reflection/refraction coefficients, roughness) are assigned or computed, how many bounces are supported, or how the shared decoder is evaluated at secondary intersections (e.g., whether the same z_i is used and how the viewing direction is expressed in the primitive frame). This missing specification prevents reproducibility of the claimed ray-tracing results and makes it impossible to assess the off-distribution query premise.","section":"Ray-Tracing; Color aggregation along the ray"},{"comment":"The quantitative comparison in Table 2 reports a single value per metric without error bars or significance tests. Several margins are small (e.g., Mip-NeRF360 PSNR 28.93 for FlaRe versus 28.89 for Nexels and 28.84 for Scaffold-GS), and the reported SSIM/LPIPS for FlaRe trail several baselines. Given the hyperparameter-selection concern, the claim of 'competitive' quality would be strengthened by reporting variance across seeds or at least a clear statement of the number of runs and the selection rule for the reported result.","section":"Table 2; Novel-View Synthesis"}],"minor_comments":[{"comment":"The abstract contains a duplicated sentence: 'Source code is available online. Source code can be found at: https://github.com/KByrski/FlaRe'. One of the two sentences should be removed.","section":"Abstract"},{"comment":"The notation for the loss weights is inconsistent: the main text uses λ1 and λ2 in the total loss on the left-hand side, while the supplement defines L = λ1·((1−λ)·LRGB,base + λ·LRGB,FlaRe) + λ2·Ls. The main text additionally writes Ls with a subscript λ_s. These should be unified.","section":"Loss function (§3.4) and Supplementary Loss function"},{"comment":"The procedure for computing the texture-coordinate transformation A when 4∤k is described with 'one can express ... and then compute ... using these same coefficients', which is difficult to follow. A closed-form expression or pseudocode for this case would improve reproducibility.","section":"Mesh modification (Supplementary)"},{"comment":"The supplement honestly reports that the equal-budget hash grid outperforms the proposed LUT on the localized-detail signal (51.72 dB versus 43.05 dB). This caveat should also appear in the main-text discussion of Table 3, since it qualifies the claim that LUT-Encoding is uniformly superior.","section":"LUT-Encoding experiment"},{"comment":"The compression study in Figure 9 reports PSNR versus primitive count, but the curves for PUP-3DGS and LP-3DGS are not tabulated in the main text. The supplement Table 4 helps, but a short statement of the operating points (e.g., the exact remaining primitive ratios for PUP-3DGS) in the figure caption would improve readability.","section":"Figure 9"}],"recommendation":"major_revision","confidential_remarks":"The paper is positioned as a unified representation, but the quantitative evidence is concentrated in novel-view synthesis, while the distinctive capabilities (ray tracing, deformation, stylization, mesh extraction) are qualitative. The hyperparameter-selection-on-test issue in the supplement is a serious concern for the reported NVS numbers and should be resolved before publication. There is also a heavy cluster of self-citations to closely related works by the same group; this is acceptable if they are genuinely prior building blocks, but the novelty relative to Nexels and 3DGRT should be stated more crisply. I would advise the editor to require quantitative evaluations of the graphics capabilities and a clarification of the hyperparameter protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FlaRe is a competently assembled integration of known pieces, and its genuinely new contribution is the specific combination: per-primitive 96-dimensional latent descriptors decoded by a shared MLP, collision-free LUT encoding, and planar generalized Gaussian primitives. The capability table is fair: no single prior method has local neural appearance plus hardware ray tracing with secondary rays, editing, mesh extraction, and primitive reduction all in one representation. The novel-view synthesis results are competitive, with second PSNR on Mip-NeRF360, and the supplement is unusually thorough, with full gradient derivations, a controlled LUT ablation, and an honest statement that hyperparameter tuning was guided by validation/test PSNR. The citation pattern is appropriate; the self-citations are prior building blocks, not circular evidence.\n\nThe main soft spot is precisely what the stress-test note flags: the unified-representation claim depends on per-primitive descriptors remaining valid under off-distribution queries - secondary rays after reflection or refraction, deformed geometry, and stylized descriptors - and the paper gives only qualitative visuals for those capabilities. No error metrics for reflected/refracted regions, deformed-scene fidelity, or stylization accuracy. That is a real gap. It is not fatal to the core novel-view synthesis contribution, which Table 2 directly supports, but it means the headline claims for ray tracing, editing, and stylization are not yet demonstrated at the same standard as the rendering numbers.\n\nTwo smaller concerns: the compactness claim is count-based rather than a matched parameter/memory comparison, and the test-set-guided tuning could inflate reported numbers. Neither is disqualifying, but error bars or validation-based tuning would make the comparison more convincing. The far-field outdoor limitation is acknowledged and reasonable.\n\nOverall, this paper deserves a serious peer review. I would send it out and ask the authors to quantify the non-NVS capabilities and to report validation-based tuning or error bars. The core representation is useful for anyone working on neural scene graphs or graphics-CV integration, and the supplements will save future builders real time.","headline":"FlaRe combines known building blocks into a genuinely new single representation; core rendering numbers are solid, but ray tracing, deformation, and stylization claims rest only on qualitative demos.","tokens_in":652,"tokens_out":1404,"would_cite":true,"duration_ms":31912,"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":"FlaRe ties neural radiance fields to explicit planar primitives, so one learned scene model supports rendering, ray tracing, deformation, mesh extraction, and stylization.","keywords":["floating radiance networks","neural scene representation","planar generalized Gaussian primitives","recursive ray tracing","novel-view synthesis","descriptor-space stylization","mesh extraction","primitive reduction"],"falsifier":"Train FlaRe on a scene containing a planar mirror, then render a reflection from a novel viewpoint and compare it against a ground-truth photograph taken from the mirrored camera position; if the reflected image's error is substantially larger than the primary-view error, the descriptors do not generalize to secondary rays.","tokens_in":52762,"feed_emoji":"🖼️","tokens_out":5535,"duration_ms":51165,"temperature":0.7,"pith_summary":"The paper argues that the usual trade-off between continuous neural radiance fields and explicit, addressable geometry is not necessary. It introduces Floating Radiance Networks (FlaRe), where a scene is a set of floating planar generalized Gaussian primitives, each carrying a 96-dimensional latent descriptor of a local radiance field, and one lightweight decoder shared across the scene turns that descriptor, local coordinates, and viewing direction into color and opacity. Because the primitives are explicit and ray-traceable, the same model supports hardware-accelerated primary and recursive ray tracing, primitive-level deformation, mesh extraction, and appearance stylization without retraining for each effect. On standard novel-view-synthesis benchmarks FlaRe reaches competitive quality with a compact set of primitives, which is what makes the unification claim worth taking seriously.","feed_headline":"FlaRe unifies rendering, ray tracing, and editing in one scene model","feed_subtitle":"Planar Gaussian primitives carry compact radiance descriptors decoded by a shared network.","key_machinery":"The load-bearing object is the isotropic planar generalized Gaussian primitive: a flattened, elliptical Gaussian with a shape parameter κ, represented by a shared triangulated polygon proxy so rays can be intersected on the GPU. Each primitive stores a trainable latent descriptor $z_i \\in \\mathbb{R}^{96}$ that encodes the local radiance field. A single three-layer auto-decoder MLP maps the descriptor, LUT-encoded local coordinates $(u,v)$, and positionally encoded view direction to color and opacity; the final opacity is the MLP prediction modulated by the generalized Gaussian kernel. A collision-free multiresolution lookup table replaces hashed encodings, and warmup with a constant-color base renderer stabilizes geometry before the neural appearance is enabled.","core_discovery":"The central claim is that a scene can be represented as a set of independently addressable planar primitives, each with its own compact latent vector, and that this representation is expressive enough to serve simultaneously as the renderer, the ray-tracing engine, and the editable object model. Color and opacity are not stored per primitive as fixed coefficients; they are decoded at every ray–primitive intersection from the primitive descriptor, local surface coordinates, and viewing direction by a network shared by all primitives. The paper's experiments show that this design preserves reconstruction quality comparable to dedicated renderers on standard reconstruction benchmarks, while the explicit structure gives recursive reflection and refraction, deformation, TSDF mesh extraction, and style transfer in one model.","pith_inferences":["If descriptor stability holds under non-rigid deformation, the same representation could plug into animation or physics pipelines, since the primitives are already mesh proxies.","The finite planar primitives likely make the unified representation most tractable for object-centric and indoor scenes; the paper itself notes far-field outdoor accuracy is limited.","The LUT-encoding result suggests collision-free lookup may benefit other coordinate-based neural fields beyond planar Gaussian primitives."],"forward_implications":["The same trained scene can be rendered with hardware-accelerated primary rays and recursive secondary rays for reflections, refractions, transparency, and shadows, without converting to another representation.","Deforming the proxy polygon recovers updated Gaussian parameters directly, so primitive-level geometry edits take effect without retraining.","Appearance stylization runs in descriptor space while fine-tuning the shared decoder, preserving structure and shading.","The representation retains reconstruction quality under aggressive primitive reduction, so a compact primitive set suffices for many scenes.","The collision-free LUT encoding outperforms hash grids at equal parameter budgets on controlled 2D signal reconstruction."],"supporting_citations":[{"why":"Establishes the continuous volumetric radiance-field paradigm that FlaRe contrasts with and extends.","marker":"(Mildenhall et al. 2020)"},{"why":"Provides the explicit primitive Gaussian-splatting baseline with fixed per-primitive appearance that FlaRe competes against.","marker":"(Kerbl et al. 2023)"},{"why":"Supplies the mesh-proxy technique and hardware-accelerated ray tracing that FlaRe adopts for primitive intersection.","marker":"(Moenne-Loccoz et al. 2024)"},{"why":"Introduces the generalized Gaussian kernel that FlaRe modifies into its isotropic planar form.","marker":"(Rong et al. 2025)"},{"why":"Defines the multiresolution hash encoding that FlaRe's collision-free LUT encoding is compared against and replaces.","marker":"(Müller et al. 2022)"},{"why":"Contributes the planar-surfel representation and the mesh-extraction regularizers that FlaRe directly adopts.","marker":"(Huang et al. 2024)"}],"fun_headline_variants":["FlaRe merges neural rendering, ray tracing, and editing in one model","Planar Gaussian primitives carry latent radiance fields for unified FlaRe scene model","FlaRe: one scene representation for rendering, ray tracing, and edits","Neural radiance fields meet explicit geometry: FlaRe's unified scene model","FlaRe makes neural scenes ray-traceable and editable with planar primitives"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The per-primitive descriptor is learned only from photometric loss on training views, yet operations like reflections, refractions, deformation, and stylization require it to stay valid for new ray directions and new geometry; the paper shows these work qualitatively but gives no numerical evaluation of that extrapolation.","fun_headline_variants_meta":{"raw":{"variants":["FlaRe merges neural rendering, ray tracing, and editing in one model","Planar Gaussian primitives carry latent radiance fields for unified FlaRe scene model","FlaRe: one scene representation for rendering, ray tracing, and edits","Neural radiance fields meet explicit geometry: FlaRe's unified scene model","FlaRe makes neural scenes ray-traceable and editable with planar primitives"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000237,"raw_usage":{"total_tokens":1489,"prompt_tokens":911,"completion_tokens":578,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":474}},"tokens_in":527,"tokens_out":578,"duration_ms":6005,"temperature":1.0,"reasoning_tokens":474,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T21:07:14.603330+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train FlaRe on a scene containing a planar mirror, then render a reflection from a novel viewpoint and compare it against a ground-truth photograph taken from the mirrored camera position; if the reflected image's error is substantially larger than the primary-view error, the descriptors do not generalize to secondary rays.","supporting_citations":[],"review_version":1}