{"id":"06175ff7-e4c8-4297-a163-15e288b27e08","arxiv_id":"2504.17954","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"iVR-GS reconstructs volume-rendered scenes from multi-view images using composable, editable 3D Gaussian splatting, enabling real-time exploration and editing without the original volume.","lead":"This paper trains multiple small 3D Gaussian splatting models, each on images of one volume feature, then stitches them together to let users explore an entire volume in real time. The splats are made editable so color, opacity, and lighting can be adjusted, or matched to a reference image, without the original dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4.5's concatenation-only composition assumes spatially separable Gaussian opacity fields; close or nested iso-surfaces can over-occlude or duplicate geometry, and no stress test in the paper covers this.","rationale":"The paper's central contribution is not merely a fast renderer but the ability to compose independently trained editable models so that all basic TFs become visible and editable without access to the volume data. Section 4.5 is the only mechanism for this, and it is a pure concatenation of Gaussian attributes. The depth-sorted alpha blend in Eq. (2) is exact only if each model's alpha and color fields are correct at their own depths and effectively zero elsewhere. 3DGS Gaussian kernels cannot satisfy the second condition for close or nested iso-surfaces because their spatial support is unbounded and their opacity is a single per-primitive scalar. The four datasets in Table 1 likely contain structures that are well separated enough to avoid this failure mode, so they do not stress the composition assumption. The appendix explicitly asserts that adding more basic TFs does not affect rendering quality, but Figure 4 there reports only file size versus TF count, not PSNR or LPIPS, leaving the assertion unsupported. The proposed concentric-shell test would settle whether the concern is real: if it shows degradation with shell separation or TF count, the composability claim must be qualified to spatially well-separated basic scenes. This is not a rejection; the code is released, the quantitative tables are internally consistent, and the method plausibly works on the tested datasets. The reader's CONDITIONAL verdict already captures the need for such qualification, so no change to the verdict is required.","tokens_in":21891,"tokens_out":11337,"duration_ms":123308,"concrete_test":"Create a synthetic volume with two concentric shells whose scalar values lie in two disjoint TF ranges, with shell separation set to 1, 2, 4, and 8 voxels. Train one basic iVR-GS per shell using the 162-view protocol of Section 5.1, compose by concatenation as in Section 4.5, and compare the composed model against ground-truth DVR with both TFs active. If PSNR/LPIPS degrades monotonically as separation shrinks, or if difference images show ghost opacity or duplicated specular highlights in the gap, the composition assumption fails for nested geometry. Repeat the composition with 5, 10, and 20 basic TFs on the same volume; if PSNR drops as TF count rises, the appendix's assertion that composition quality is independent of the number of basic scenes is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.5 composes basic iVR-GS models by decompressing and appending Gaussian attribute lists into one parameter list, with no joint optimization or correction. The correctness of this step relies on the opacity fields of the different basic TFs being additive under Eq. (2). But Eq. (2) is a depth-sorted alpha blend, and 3DGS primitives have unbounded Gaussian spatial support. Even when TF ranges are disjoint in voxel-value space, two iso-surfaces from different ranges can be arbitrarily close or nested in object space. A Gaussian trained for range A can therefore have non-negligible opacity at the surface of range B and vice versa; after concatenation, the transmittance product in Eq. (2) multiplies both models' contributions, causing over-occlusion, duplicated specular highlights, or visible seams. The four datasets in Table 1 do not include a deliberate stress case for close or nested structures, and the appendix's claim that increasing the number of basic TFs 'will not influence rendering quality' is asserted without a quality-versus-TF-count experiment. Thus the central 'no extra optimization' composability claim is only demonstrated for favorably separated structures.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces iVR-GS, a 3D Gaussian splatting-based inverse volume rendering framework. For each of several disjoint opacity transfer functions (basic TFs), a basic iVR-GS model is trained on multi-view images; the Gaussians are augmented with normals and Blinn-Phong shading attributes so that color, opacity, and lighting can be edited at inference time. Basic models are compressed with vector quantization and then composed by concatenating their Gaussian attribute lists, with no joint optimization, to make the full volume visible. The paper compares reconstruction quality, speed, and model size against Plenoxels, CCNeRF, and base 3DGS on four volumes and demonstrates relighting, iterative editing, and inverse volume exploration on three additional datasets. The main asserted contribution is that independently trained basic models can be merged without extra optimization and then edited without accessing the original volume data.","tokens_in":22100,"tokens_out":9984,"duration_ms":96987,"significance":"The pipeline is well motivated: it targets the real bottleneck of interactive exploration of large volumes on low-end devices and combines standard components (3DGS, Blinn-Phong, VQ) in a reproducible way. Concrete strengths include a released implementation, an ablation of VQ compression, a comparison of one-stage versus two-stage training, and consistent quantitative gains over three baselines on four datasets. If the composition assumption holds for general volumes, the work would be a solid contribution to visualization-oriented novel view synthesis. The main unresolved question is whether composition by concatenation is valid beyond the favorable cases in the evaluation; the paper's own appendix also makes a scalability claim that is not backed by a quality experiment.","major_comments":[{"comment":"Section 4.5 composes basic models by appending Gaussian attribute lists and assumes that the opacity fields of independently trained basic scenes combine correctly under the depth-sorted alpha blend in Eq. (2). This is not a consequence of the disjointness of the TF ranges: alpha blending is a transmittance product, and Gaussians have unbounded spatial support, so a primitive trained for one value range can have non-negligible opacity at another value range's surface. For close or nested iso-surfaces, which the authors themselves identify as typical of volume scenes in Section 5.1, the concatenated model can produce over-occlusion, duplicated geometry, or seams, and no joint optimization exists to correct for this. The four datasets in Table 1 do not include a stress case of this kind. I request an experiment with synthetic or real volumes containing nested or near-adjacent iso-surfaces, reporting per-basic and composed PSNR, or a clear limitation statement restricting the composability claim to spatially well-separated basic scenes.","section":"Section 4.5, Eq. (2)"},{"comment":"Appendix 2 asserts that increasing the number of basic TFs 'will not influence the rendering quality of the VolVis scene' because each basic model is trained independently. The accompanying experiment (Appendix Figure 4) reports only model size as a function of the number of TFs, with no PSNR or LPIPS measurement. Since the scalability of the composed model is part of the paper's central contribution, this quality-versus-TF-count claim should either be supported by an experiment or explicitly downgraded to an untested expectation.","section":"Appendix 2, 'Number of TFs'"},{"comment":"Sections 5.1 and Table 2 report a single optimization run per method and dataset, and on three of the four datasets the composed-scene PSNR advantage of iVR-GS over base 3DGS is small (27.19 vs 26.45 on chameleon, 23.04 vs 22.95 on rotstrat, 29.20 vs 28.73 on supernova). Without repeated runs or error bars, the claim that iVR-GS achieves the best reconstruction quality for all datasets is not statistically established. In addition, Section 5.1 says baselines are optimized 'in the same way as iVR-GS,' but Section 4.2 describes an alpha-channel L1+SSIM loss used for base 3DGS and iVR-GS; the manuscript does not state whether Plenoxels and CCNeRF received the same alpha supervision. If they did not, the comparison is not controlled, because the authors credit alpha supervision with improving reconstruction quality. Please specify the exact losses used for each baseline or re-run with matched supervision.","section":"Section 5.1, Table 2"}],"minor_comments":[{"comment":"The phrase 'editable Guassians' in the overview paragraph is a typo and should read 'Gaussians'.","section":"Section 4 (overview)"},{"comment":"The Table 1 header contains a duplicated 'volume volume' and does not clearly separate the CPU/GPU memory column from the rendering-time column; please reformat the header.","section":"Table 1"},{"comment":"The text in Section 4.5 says scaling ka, kd, ks, and beta adjusts the magnitudes of the shading terms, but in Eq. (3c) beta is an exponent; scaling it changes the specular falloff rather than the magnitude. Please rephrase, or adjust beta in log space if a magnitude-like control is intended.","section":"Section 4.5"},{"comment":"The iterative editing results in Figure 7 are presented qualitatively; since editing is a claimed contribution, a quantitative metric (e.g., per-edit PSNR against a re-rendered reference for a controlled edit) or a small user study would strengthen the evaluation.","section":"Section 5.2, Figure 7"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the manuscript is a reasonable fit for this venue and the core building blocks are sound. My recommendation rests on the need for a targeted composition stress test and a controlled baseline comparison; neither issue, in my view, requires rejection. The paper's own appendix contains an unsupported scalability claim that should be fixed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, iVR-GS makes a real advance: it brings editable shading and transfer-function control to 3D Gaussian splatting for volume visualization, so you can explore a scene in real time on a consumer GPU without touching the raw volume. Second, the central composability claim—that you can just concatenate the Gaussians from independently trained TF-specific models and get a faithful full scene—is plausible but not yet proven for difficult cases.\n\nWhat's genuinely new is the combination: Blinn-Phong shading attributes on each Gaussian, disjoint TF models trained separately, composition by parameter concatenation, vector quantization for storage, and inverse exploration from a reference image. The paper is clearly written, the code is available, and the quantitative comparison against Plenoxels, CCNeRF, and base 3DGS on four volumes is consistent and favorable. The two-stage training trick (geometry first, then shading) is a nice practical win.\n\nThe soft spots are real but not fatal. The stress-test concern about composition holds up: Gaussians have infinite spatial support, so two models trained on disjoint TF ranges can still have primitives that overlap in space. If the iso-surfaces are close or nested, the transmittance product in alpha blending can cause over-occlusion or duplicated highlights. The paper's four datasets don't include such a stress case, and the appendix claim that increasing the number of TFs 'will not influence rendering quality' is not backed by an experiment. That is a genuine gap. The evaluation also lacks error bars, editing quality is mostly qualitative, and there's no head-to-head with StyleRF-VolVis or a clear statement on whether baselines used alpha-channel supervision. These are all addressable.\n\nOverall, the central argument holds up: given the evidence, the method works on the tested data and the editing capabilities are new. The composition failure mode is a limitation to probe, not a refutation. This paper deserves serious refereeing, with requests for a composition stress test (e.g., nested or close iso-surfaces) and more experimental rigor. I'd certainly cite it, and I'd bring it to our reading group for the discussion.","headline":"A solid, citable step toward explorable 3DGS volume visualization; the composition-by-concatenation claim needs stress-testing, but the paper deserves peer review.","tokens_in":22625,"tokens_out":2371,"would_cite":true,"duration_ms":24783,"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":"Independently trained Gaussian splat models can be stitched into a single editable volume scene.","keywords":["volume visualization","novel view synthesis","3D Gaussian splatting","transfer function","scene composition","inverse rendering","vector quantization","interactive editing"],"falsifier":"Train two basic iVR-GS models on transfer-function bands whose opacity ranges overlap for the same voxel values, compose them by concatenation, and compare against a single model trained on the union of the bands. If rendering shows duplicated surfaces, incorrect transparency ordering, or seams at the overlap, the disjoint-transfer-function composability claim fails.","tokens_in":21687,"feed_emoji":"🧊","tokens_out":8137,"duration_ms":75814,"temperature":0.7,"pith_summary":"iVR-GS sets out to remove the biggest limitation of using novel view synthesis for volume visualization: once a scene is reconstructed, the transfer function and lighting are baked in, so users cannot explore other structures. The paper's claim is that this can be fixed by representing each basic scene as a set of editable 3D Gaussians, training those basic models independently on disjoint opacity ranges of a transfer function, and then composing them into one model by simply concatenating their Gaussian primitives. The composed model then renders the entire volume in real time, supports edits to color, opacity, and light direction and magnitude, and can even infer the rendering settings of a reference image through inverse exploration. The paper reports that this approach beats Plenoxels, CCNeRF, and base 3DGS in reconstruction quality on the tested datasets while keeping the compressed model small.","feed_headline":"Composed Gaussian splats render whole volumes without retraining","feed_subtitle":"Merging independently trained 3D Gaussian models lets users recolor, relight, and explore a volume in real time.","key_machinery":"The central object is the editable Gaussian primitive, a 3D Gaussian point augmented with a normal and with Blinn-Phong shading attributes (offset color, ambient, diffuse, and specular coefficients, and shininess) in addition to the standard mean, rotation, scale, and opacity. There is also a per-scene palette color shared by all primitives. The composability mechanism is concatenation: because each primitive is spatially explicit and independent, appending the decompressed attribute lists of basic models forms the composed model, whereas grid-based NeRF representations interfere when summed. The Blinn-Phong decomposition is what makes editing possible: changing the palette color, scaling opacity, or adjusting the shading coefficients and light direction re-renders the scene.","core_discovery":"The central claim is that a volume-rendered scene does not need to be reconstructed as a single monolithic model under one transfer function. Instead, each visible part of the volume, defined by one basic transfer-function opacity bump, can be learned as its own iVR-GS model made of editable Gaussian primitives; concatenating the attribute lists of these independently trained models produces a composed model that makes every part visible without joint optimization. Each editable Gaussian carries the usual geometry attributes plus a normal and Blinn-Phong shading attributes, so color and lighting are explicit and editable rather than baked into spherical harmonic coefficients. The paper further claims that this representation supports inverse exploration: freezing all primitives and optimizing a small set of global color, opacity, and light parameters lets the composed model match a user-provided reference image and then generalize to novel views. Vector quantization compresses the composed model by roughly four times, and the paper reports a file size up to 7.2 times smaller than base 3DGS on one dataset.","pith_inferences":["A natural extension the paper leaves implicit is to treat time steps or ensemble members as basic scenes, so composition would let users flip between states in real time while keeping the whole volume visible.","The disjoint-transfer-function assumption suggests a stress test the paper does not run: if two basic transfer functions overlap in value range, the concatenated primitive list may contain duplicate geometry for the same structures; a fusion rule for overlapping bands would generalize the method.","The paper's stated limitations, composed models reaching GB scale for hundreds of basic scenes and Blinn-Phong being the assumed lighting model, define where the approach stops; replacing Blinn-Phong with a learned per-primitive BRDF while keeping concatenation-based composition is a testable next step.","Inverse exploration currently tunes global color, opacity, and light parameters; extending it to per-region or per-primitive parameters would let a reference image drive local transfer-function changes, which the paper does not explore."],"forward_implications":["Large volumes that normally need a cluster can be explored on a consumer GPU, since rendering cost at inference depends on Gaussian primitives, not volume resolution.","A scene can be revealed progressively: each newly trained basic model for another transfer-function band can be appended to the composed model without retraining existing parts.","Users can relight a reconstructed scene from new light directions because normals and Blinn-Phong parameters are stored per primitive, not baked into view-dependent colors.","Given one reference image, the composed model can recover the color, opacity, and light settings that produced it and then render novel views under those settings.","With vector quantization, adding more transfer-function bands increases model size more slowly, so storage scales more gracefully than storing each basic model at full precision."],"supporting_citations":[{"why":"Supplies the base 3D Gaussian splatting representation and rasterization that iVR-GS extends.","marker":"[22]"},{"why":"Establishes the application of 3D Gaussian splatting to volume-rendered anatomy on consumer devices, the scenario iVR-GS builds on.","marker":"[33]"},{"why":"Plenoxels is the explicit-grid baseline whose parameter-summing composability is compared and outperformed.","marker":"[7]"},{"why":"CCNeRF is the rank-concatenation baseline whose composability is compared and outperformed.","marker":"[39]"},{"why":"Provides the normal consistency loss and BRDF decomposition ideas adopted for editable Gaussians.","marker":"[8]"},{"why":"Supplies the vector-quantization compression scheme for 3DGS attributes that iVR-GS adapts.","marker":"[34]"},{"why":"Defines the LPIPS perceptual metric used to report reconstruction quality.","marker":"[58]"},{"why":"The renderer used to generate the multi-view training images with Blinn-Phong lighting.","marker":"[1]"}],"fun_headline_variants":["Composed Gaussians open entire volumes for real-time edits","iVR-GS: every volume part visible with editable splats","Editable Gaussians let you explore any volume in real time","Composite Gaussian models render full volumes, editable on the fly","Inverse volume rendering meets editable 3D Gaussian splats"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that independently trained Gaussian models, each covering a different non-overlapping visible range of the volume, can be merged by simply concatenating their primitive lists, with no joint optimization, and still render the combined scene correctly.","fun_headline_variants_meta":{"raw":{"variants":["Composed Gaussians open entire volumes for real-time edits","iVR-GS: every volume part visible with editable splats","Editable Gaussians let you explore any volume in real time","Composite Gaussian models render full volumes, editable on the fly","Inverse volume rendering meets editable 3D Gaussian splats"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001003,"raw_usage":{"total_tokens":4264,"prompt_tokens":984,"completion_tokens":3280,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":3195}},"tokens_in":600,"tokens_out":3280,"duration_ms":23499,"temperature":1.0,"reasoning_tokens":3195,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:28:08.348671+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train two basic iVR-GS models on transfer-function bands whose opacity ranges overlap for the same voxel values, compose them by concatenation, and compare against a single model trained on the union of the bands. If rendering shows duplicated surfaces, incorrect transparency ordering, or seams at the overlap, the disjoint-transfer-function composability claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CCNeRF is the rank-concatenation baseline whose composability is compared and outperformed."},{"cited_title":"Niedermayr, J","cited_arxiv_id":null,"evidence_quote":"Supplies the vector-quantization compression scheme for 3DGS attributes that iVR-GS adapts."},{"cited_title":"https://gitlab.kitware.com/ paraview/paraview","cited_arxiv_id":null,"evidence_quote":"The renderer used to generate the multi-view training images with Blinn-Phong lighting."}],"review_version":1}