{"id":"396652b7-7901-412e-9343-5208c4bee3ce","arxiv_id":"2411.12089","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"FruitNinja generates 3D interior textures for Gaussian Splatting objects by progressively inpainting cross-sectional views with a diffusion model, enabling real-time arbitrary slicing without further optimization.","lead":"This paper presents a method for teaching 3D digital objects to show realistic insides when sliced, like watermelon flesh and seeds, using Gaussian Splatting and a diffusion model. It allows real-time cutting and viewing in tools like AR/VR and games, without needing to recompute textures after each slice.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quantitative evidence for 'unmatched visual quality' is circular: Table 1's FID/KID reference images are the same 1–4 cross-section images used for DreamBooth fine-tuning and SDS training, and the table actually shows 2D Inpainting (Fine-tuned) with better FID/KID than Ours.","rationale":"I focused on the quantitative evaluation because the central technical contribution—real-time slicing at arbitrary angles—may well be sound and is supported by the figures, but the paper's own numbers do not establish it. The reader's weakest_assumption concerned whether sparse canonical cross-sections and Eq. (6) can infer a full interior; that is a real limitation for non-coherent interiors, but for a generative method plausible hallucination is partly the point and the paper's qualitative results partially address it. The evaluation contamination is more immediately decisive: it invalidates the only quantitative evidence for 'unmatched visual quality', and the table even contradicts the abstract's superiority claim. The correct disposition remains CONDITIONAL because the issue is fixable with held-out metrics, the real-time rendering property is plausible, and the qualitative demonstrations show the pipeline can work. I agree with the reader's overall verdict but not with their identification of the weakest assumption; the reader did flag the contamination in their rationale, hence 'partial' rather than 'disagree'. No code or data release is available beyond 'plan to release', so external verification is currently impossible; a held-out protocol plus released artifacts should be conditions for full acceptance.","tokens_in":11788,"tokens_out":8708,"duration_ms":93556,"concrete_test":"Re-run the quantitative evaluation with a per-object held-out split of real cross-section photographs: use only a training subset for DreamBooth fine-tuning and SDS reference views, and compute FID/KID/CLIP only on held-out cross-section images never seen by any component of the pipeline. If only 1–4 images per object are available, collect additional real sliced-object photographs or use leave-one-out cross-validation, reporting mean and variance over folds. If Ours's held-out FID/KID are not comparable to or better than the fine-tuned 2D inpainting baseline, or if the in-set advantage disappears or reverses, the abstract's 'unmatched visual quality' and 'substantially outperforms' claims must be removed or qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is the quantitative evaluation supporting the abstract's 'substantially outperforms existing approaches' and 'unmatched visual quality'. In Section 4.1 the authors collect 1–4 real cross-section images per object and use them for DreamBooth fine-tuning of Stable Diffusion; Section 3.2 additionally uses those images as SDS reference views during 3DGS training. Table 1 then computes FID and KID 'using the collected real-world canonical cross-sections images mentioned in 4.1'. The reference set is therefore not held out: both the diffusion model and the 3DGS are trained to reproduce the exact images on which fidelity is measured, so low FID/KID can reflect memorization rather than generalization to arbitrary cuts. The problem is compounded by an internal inconsistency: Table 1 shows Ours with FID 209.2/KID 323.7 versus 2D Inpainting (Fine-tuned) with FID 176.2/KID 224.5, so Ours is worse on both fidelity metrics, yet the text calls this 'comparable' and the abstract claims substantial superiority. The only metric where Ours leads (CLIP) is a text-alignment score on canonical cross-sections, not a fidelity measure for arbitrary slicing. Since Table 1 is limited to canonical orientations and Table 2 measures only CLIP self-consistency at 120 random angles, there is no independent quantitative support for the central arbitrary-slicing claim. This is fixable with a held-out protocol, but the claim as stated is not supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FruitNinja addresses the generation of internal textures for 3D Gaussian Splatting (3DGS) objects, enabling real-time rendering when objects are sliced or cut. The pipeline fills the object interior with Gaussian primitives, defines user-specified canonical cross-sections, optimizes reference cross-sectional images via Score Distillation Sampling (SDS) with a depth-conditioned Stable Diffusion model, optionally fine-tunes the diffusion model with DreamBooth on 1–4 real cross-section images per object, and jointly trains cross-section and surface views. It then applies progressive texture refinement and voxel-grid smoothing, and uses an 'OpaqueAtom GS' strategy (atomic clipping and uniform high opacity) to stabilize training and preserve fine details. Experiments on six objects compare against PhysGaussian and 2D inpainting, reporting CLIP, FID, and KID on canonical cross-sections and CLIP cosine similarity on 120 random slicing angles.","tokens_in":12082,"tokens_out":5251,"duration_ms":51325,"significance":"If the method works as demonstrated qualitatively, it fills a genuine gap: existing 3D editing and inpainting methods do not handle interior texture under topological changes, and the 'train once, slice in real time' property is practically useful. The paper's strengths include a clearly described pipeline, a sensible adoption of AtomGS-inspired training for dense small Gaussians, and qualitative results showing multiple coherent cuts on the same object. The central method is not circular: it uses a diffusion prior to synthesize references and fits 3DGS to them. However, the quantitative evidence for the abstract's claims of 'substantially outperforms existing approaches' and 'unmatched visual quality' is not reliable as presented, because the FID/KID reference images are the same images used for DreamBooth fine-tuning and SDS supervision, and the reported numbers in Table 1 actually show the fine-tuned 2D inpainting baseline with better FID and KID than Ours. The evaluation also lacks a held-out fidelity test for the arbitrary-slicing claim that is central to the paper. These issues are fixable with a cleaner evaluation protocol, but they currently block acceptance.","major_comments":[{"comment":"The FID/KID reference set is not held out: the same 1–4 real cross-section images per object used for DreamBooth fine-tuning (Section 3.2, 'Fine-tuning with DreamBooth') and for SDS and reconstruction supervision during 3DGS training are also used to compute FID and KID in Table 1. Low FID/KID can therefore reflect memorization of the training images rather than generalization to novel cuts. Please report results with a held-out set of real cross-section images (for example, leave-one-out or additional real photos per object), or, at minimum, clearly disclose that the metric is computed on training images and interpret the numbers accordingly.","section":"§4.1, §3.2, Table 1"},{"comment":"Table 1 shows Ours with FID 209.2 and KID 323.7, while '2D Inpainting (Fine-tuned)' achieves FID 176.2 and KID 224.5. The text labels Ours as 'comparable' to the 2D inpainting method, but numerically Ours is worse on both fidelity metrics, and the abstract claims FruitNinja 'substantially outperforms existing approaches, showcasing unmatched visual quality.' This is an internal inconsistency in the quantitative support. The claims should be revised to match the reported numbers, or the evaluation should be redone with a protocol that actually demonstrates a fidelity improvement over the baselines.","section":"Table 1 and Abstract"},{"comment":"The core claim of the paper — real-time, arbitrary-angle slicing with plausible internal textures — is not covered by Table 1, which evaluates only canonical cross-sections, nor by Table 2, which measures only CLIP cosine self-consistency among 120 random slices and lacks any fidelity or human-preference comparison at those angles. Please add an evaluation on non-canonical cuts (for example, a user study, per-slice FID against held-out reference images at matching angles, or a comparison with 2D inpainting at random angles), or restrict the abstract's 'unmatched visual quality' claim accordingly.","section":"§4.3, Table 2"},{"comment":"Voxel smoothing assigns colors to untrained Gaussians by inverse-distance-weighted averaging of nearby trained Gaussians within each voxel. This assumes locally coherent textures and is the mechanism responsible for quality on cuts that do not coincide with training views; however, its effect is demonstrated only qualitatively in Figure 10 and no quantitative ablation is provided. Please quantify the impact of Eq. (6) on the metrics in Tables 1 and 2, and state the voxel-grid resolution and smoothing interval used for each object.","section":"§3.3, Eq. (6)"}],"minor_comments":[{"comment":"The text contains a typo: '1-–4 cross-sectional images' should be '1–4 cross-sectional images'.","section":"§4.1"},{"comment":"The notation I^p_label is unclear, and the symbol ϵ is used both for the added noise and as a subscript in ϵ_θ; please rewrite the equation with distinct symbols for added noise and the noise-prediction network.","section":"Eq. (4)"},{"comment":"Section 4.1 says 20 surface views per iteration, while Section 3.2 says 10–20 surface views; please make the numbers consistent and specify the exact value used in the experiments.","section":"§4.1 and §3.2"},{"comment":"Table 2 reports no error bars, standard deviations, or per-object breakdowns; please include them, since the number of random angles and objects is small.","section":"Table 2"},{"comment":"The terminology for the proposed strategy varies among 'OpaqueAtom GS', 'Opaque-Atomic Gaussians', and 'OpaqueAtomGS'; please unify the spelling in the text, figures, and captions.","section":"Throughout"},{"comment":"The convergence criterion for progressive texture refinement is stated as 'below a predefined threshold ϵ', but the threshold value is never given; please report the actual threshold or describe how it was determined.","section":"§3.3"}],"recommendation":"major_revision","confidential_remarks":"The evaluation protocol is the main obstacle to publication. The method itself appears plausible and the qualitative demonstrations are promising, so I recommend major revision rather than rejection: the circularity in the FID/KID protocol is fixable with held-out data, and the claim language can be aligned with the actual metrics. Please ask the authors to release code and data to facilitate verification of the real-time slicing claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine new task and the system plausibly works, but the quantitative evidence for the headline claim is circular and the abstract overstates what the table shows. I think the paper deserves a serious referee, but the evaluation section needs to be redone before it can be accepted.\n\nWhat's new: the paper is the first to target interior texture generation for 3DGS objects so that arbitrary slicing reveals plausible content without per-edit optimization. The pipeline is a coherent assembly of known pieces — PhysGaussian interior filling, SDS with depth-conditioned Stable Diffusion, DreamBooth fine-tuning, progressive refinement, voxel smoothing, and an atomized, opaque Gaussian variant. The qualitative results, especially the watermelon and pomegranate, look believable. The ablations are informative and the failure modes they illustrate (spatial conflicts, untrained Gaussians) are honest.\n\nSoft spots: the stress-test note lands. The FID/KID reference set is the same 1–4 cross-section images used for DreamBooth fine-tuning and SDS training, so low scores partly measure memorization. Worse, the paper's own Table 1 shows 2D Inpainting (Fine-tuned) with FID 176.2/KID 224.5 vs Ours 209.2/323.7 — Ours is worse on both fidelity metrics, yet the text calls it \"comparable\" and the abstract claims \"substantially outperforms\" and \"unmatched visual quality.\" That claim is not supported. The only metric where Ours leads is CLIP, which measures text alignment on canonical cross-sections, not fidelity for arbitrary cuts. Table 2's cosine similarity at 120 random angles is a self-consistency measure, not a ground-truth check. So the central arbitrary-slicing claim currently rests on qualitative demos. That is fixable with a held-out protocol, but it's a real gap.\n\nMinor: no error bars, key hyperparameters (voxel grid resolution, smoothing interval, opacity threshold) are empirical and not analyzed, and there is no code or data release yet.\n\nOverall: the core approach is not circular — using a diffusion prior to synthesize references and fitting 3DGS to them is a reasonable design. The task is new and the engineering is thoughtful. This is a conditional accept with major evaluation revisions, not a desk reject. I'd bring it to reading group mainly to talk about evaluation pitfalls. I wouldn't cite it until the evaluation is cleaned up and artifacts are out.","headline":"A plausible new task and a coherent pipeline, but the quantitative evaluation is circular and the abstract's superiority claim is contradicted by the paper's own Table 1.","tokens_in":12701,"tokens_out":2315,"would_cite":false,"duration_ms":22953,"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":"FruitNinja trains a 3D Gaussian object once so any cut reveals plausible interior texture in real time, the authors claim.","keywords":["3D Gaussian Splatting","interior texture generation","cross-section inpainting","score distillation sampling","diffusion prior","real-time slicing","voxel smoothing","OpaqueAtom GS"],"falsifier":"Take an object whose interior is not locally coherent and is absent from the diffusion prior — for instance, a hollow chocolate bunny with a single toy inside, or a block of Swiss cheese with random holes — train FruitNinja using canonical cross-sections, then cut it at an angle not in the training set and photograph the actual cut. If the rendered interior reproduces the hidden toy or the random holes, the claim survives; if it shows only smooth, averaged flesh-like texture, the claim that arbitrary cuts reveal plausible internal textures is settled false.","tokens_in":11497,"feed_emoji":"🍉","tokens_out":11726,"duration_ms":112039,"temperature":0.7,"pith_summary":"The paper introduces FruitNinja, a method for generating the interior textures of 3D Gaussian Splatting (3DGS) objects — 3D scenes rendered as clouds of colored Gaussian kernels — so that slicing, cutting, or tearing the object exposes a plausible internal structure in real time. The authors claim this is the first such method: a model trained once needs no extra optimization at edit time, unlike per-view 2D inpainting or inherited-color interior filling. Training is driven by a small set of user-specified cross-section views (for example, 30 evenly spaced vertical and 40 horizontal slices for a watermelon), a pre-trained text-to-image diffusion model, and a voxel-smoothing step that colors every untrained interior particle from nearby trained ones. The interior is represented as densely packed, small, fully opaque Gaussian particles, which the paper shows is necessary for stable training and sharp skin-to-flesh color transitions. If the central claim is right, interactive 3D content such as virtual fruit cutting, game objects, and AR/VR editing can show believable interiors in real time rather than after a slow per-edit diffusion pass.","feed_headline":"First real-time slicing of 3D object interiors","feed_subtitle":"Train once with a few cross-sections, then cut anywhere and see plausible internal textures on the fly.","key_machinery":"The mechanism that carries the argument is a cross-section-conditioned training loop over an OpaqueAtom Gaussian volume. For each user-defined cutting plane $ax + by + cz + d = 0$, a 3D mask keeps only Gaussians near the plane; rendering under that mask exposes the slice, and a depth-conditioned text-to-image diffusion model combined with score distillation sampling (SDS) generates a reference image of what that slice should look like. Those reference views are then used as reconstruction targets while 10–20 random exterior views are trained alongside them to preserve the object's surface. Progressive refinement re-applies SDS to the current renderings to reconcile conflicting seeds, and voxel smoothing assigns colors to untrained Gaussians by inverse-distance weighting, $$C = \\frac{\\sum_i w_i C_i}{\\sum_i w_i},$$ with weights based on Euclidean distance within a voxel. OpaqueAtom GS makes this stable by capping each Gaussian's scale at roughly one three-thousandth of the object's dimensions and giving every particle full opacity, avoiding the large-Gaussian drift and front-back blending that blur abrupt interior boundaries.","core_discovery":"The central claim is that a single 3D Gaussian Splatting object can be trained with both its outside and inside so that arbitrary geometric and topological edits — cuts, slices, removals — reveal a convincing interior while rendering in real time. The authors state this as a first: prior work either inpaints each newly exposed region per edit (slow and inconsistent) or fills the interior by copying surface colors (unrealistic for fruit and other objects whose interiors differ from their exteriors). FruitNinja instead seeds the interior with raw Gaussian particles, renders and optimizes a small set of user-defined cutting planes using a depth-conditioned diffusion model with score distillation sampling, jointly trains those reference cross-sections with randomly selected surface views, and progressively refines them so conflicting seed textures settle into a coherent volume. Untrained interior Gaussians are then colored by inverse-distance-weighted averaging of nearby trained ones on a voxel grid. With the OpaqueAtom GS strategy — capping Gaussian scale and using uniform full opacity — the paper reports that slicing at 120 arbitrary angles yields the best CLIP scores and cross-slice consistency among the compared baselines.","pith_inferences":["[Editorial inference] The same pipeline could extend to any object whose internal structure is locally predictable — layered rock, anatomical tissue, or foam — provided a diffusion model can be prompted or fine-tuned to imagine its cross-sections, which is a testable extension of the paper's symmetry assumption.","[Editorial inference] For objects with hidden, non-local, or stochastic interiors, such as a box with random contents or a machine part with a single internal void, the smooth voxel averaging will produce a plausible-looking but fictitious interior; slicing such an object would expose that fiction, so the method is a texture generator rather than a true interior scanner.","[Editorial inference] The voxel-smoothing prior is local and isotropic, so replacing it with a learned 3D generative prior over Gaussian colors could capture elongated or nested structures such as fibers, shells, and cavities without sacrificing real-time rendering."],"forward_implications":["Once trained, a FruitNinja object can be sliced, torn, or partially removed interactively, and the exposed interior renders immediately with no diffusion sampling or per-edit optimization.","Cuts at angles never seen during training still show plausible textures because the entire interior volume is populated, trained, and voxel-smoothed, not just the canonical slice planes.","The pipeline handles objects whose interiors differ strongly from their surfaces — fruit flesh versus rind, cake crumb versus frosting — where an interior-filling baseline that inherits surface colors visibly fails.","Evaluated at 120 random cut angles on the paper's six objects, the method reports the highest CLIP score and cross-slice cosine similarity, so consecutive edits stay visually consistent."],"supporting_citations":[{"why":"Supplies the 3D Gaussian Splatting representation and alpha-composited rendering that FruitNinja fills and trains.","marker":"[18]"},{"why":"Provides the Score Distillation Sampling loss used to synthesize each reference cross-section view from a diffusion prior.","marker":"[27]"},{"why":"Provides the depth-conditioned latent diffusion model that generates the cross-section images and supports optional fine-tuning.","marker":"[31]"},{"why":"Provides the relative depth estimator whose depth maps condition the diffusion model on the shape of each slice.","marker":"[1]"},{"why":"Inspires the atomic clipping strategy that caps Gaussian sizes to keep interior particles small and stable.","marker":"[22]"},{"why":"Supplies the internal Gaussian filling procedure based on opacity-field ray casting that FruitNinja reuses, and is its main baseline.","marker":"[39]"},{"why":"Supplies the DreamBooth fine-tuning procedure that adapts the diffusion model to object-specific cross-sections when needed.","marker":"[32]"},{"why":"Supplies the joint render-and-refine loop used in progressive texture refinement to reconcile inconsistent cross-sections.","marker":"[35]"}],"fun_headline_variants":["First 3D interior generation for real-time slicing","Cut any 3D object and see inside instantly","FruitNinja: first to synthesize 3D interiors on cut","First real-time interior textures for 3D slicing","Slicing 3D objects reveals lifelike interiors instantly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a handful of user-chosen cross-section photos and a pre-trained text-to-image model can stand in for the object's entire unseen interior, and that every point not covered by those views can be plausibly colored by averaging nearby trained colors.","fun_headline_variants_meta":{"raw":{"variants":["First 3D interior generation for real-time slicing","Cut any 3D object and see inside instantly","FruitNinja: first to synthesize 3D interiors on cut","First real-time interior textures for 3D slicing","Slicing 3D objects reveals lifelike interiors instantly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000534,"raw_usage":{"total_tokens":2598,"prompt_tokens":1002,"completion_tokens":1596,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":1514}},"tokens_in":618,"tokens_out":1596,"duration_ms":11406,"temperature":1.0,"reasoning_tokens":1514,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:55:01.580742+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an object whose interior is not locally coherent and is absent from the diffusion prior — for instance, a hollow chocolate bunny with a single toy inside, or a block of Swiss cheese with random holes — train FruitNinja using canonical cross-sections, then cut it at an angle not in the training set and photograph the actual cut. If the rendered interior reproduces the hidden toy or the random holes, the claim survives; if it shows only smooth, averaged flesh-like texture, the claim that arbitrary cuts reveal plausible internal textures is settled false.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Supplies the 3D Gaussian Splatting representation and alpha-composited rendering that FruitNinja fills and trains."},{"cited_title":"Barron, and Ben Milden- hall","cited_arxiv_id":null,"evidence_quote":"Provides the Score Distillation Sampling loss used to synthesize each reference cross-section view from a diffusion prior."},{"cited_title":"Physgaussian: Physics- integrated 3d gaussians for generative dynamics","cited_arxiv_id":null,"evidence_quote":"Supplies the internal Gaussian filling procedure based on opacity-field ray casting that FruitNinja reuses, and is its main baseline."},{"cited_title":"Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation","cited_arxiv_id":null,"evidence_quote":"Supplies the DreamBooth fine-tuning procedure that adapts the diffusion model to object-specific cross-sections when needed."},{"cited_title":"Dreamgaussian: Generative gaussian splatting for ef- ficient 3d content creation","cited_arxiv_id":null,"evidence_quote":"Supplies the joint render-and-refine loop used in progressive texture refinement to reconcile inconsistent cross-sections."}],"review_version":1}