{"id":"aecc89f4-a5af-42d6-ae32-3adaa2edb344","arxiv_id":"2607.17773","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FillGauss generates hit sounds for 3D-scanned containers conditioned on strike position, striker material, and internal fill state, using a new 5,000+ sample dataset.","lead":"This paper introduces FillImpact, a dataset of over 5,000 impact recordings from 88 real containers with different amounts of water or rice, and FillGauss, a model that generates impact sounds from a 3D scan, a hit location, and a text description of the fill state. The work targets applications like VR and robotics where objects should sound different depending on what is inside them.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 80/20 split is only described as a sample partition; without object-disjoint separation, the model can memorize per-object resonances, so the claimed generalization to unseen containers is unestablished.","rationale":"The paper's central claim is that FillGauss enables filling-aware, position-aware impact sound generation that generalizes from 3DGS geometry to physical audio. The dataset is built around 88 real objects, so the key generalization test is whether the model can handle a container it has never seen. The reader's weakest assumption identifies exactly this vulnerability: the 80/20 split is described only as a sample partition, not as an object-disjoint partition. If the same object appears in both splits, the 3DGS geometry stream can memorize object-specific resonances, making the otherwise plausible cross-modal conditioning mechanism appear to generalize when it has only retrieved a training-object fingerprint. This concern is load-bearing because it directly affects all objective and subjective evidence for the central claim. I considered the alternative concern that FillGauss's FAD (1.3500) is worse than fine-tuned TangoFlux (1.0627), which weakens the literal 'state-of-the-art' claim, but the split issue is more fundamental: if leakage exists, even the in-domain FAD advantage over SonicGauss-FT is not evidence of generalization. The paper provides no code or data release, so the split cannot be audited. A simple object-disjoint retraining experiment would settle the question. Until then, the conditional verdict is appropriate; no change from the reader's verdict is needed.","tokens_in":13948,"tokens_out":5549,"duration_ms":67681,"concrete_test":"Release the split metadata and re-run the full evaluation with a strict object-disjoint split (e.g., 70 objects train / 18 objects test), holding all hyperparameters fixed. Compare FillGauss's FAD and Fine-grained Attribute Matching Rate on the unseen-object test set with the values in Tables I and III. If FAD rises substantially toward the SonicGauss-FT level (1.57) or attribute matching drops below 50%, the reported gains are at least partly due to object-identity memorization. Also inspect the original split file for any object ID present in both splits; if so, the current numbers cannot support the generalization claim without retraining.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V.A states only that 'we partition FillImpact Dataset into an 80% training split and a 20% testing split,' with no statement that the 88 physical objects are split disjointly. FillImpact contains over 5,000 recordings from 88 objects, with each object appearing under multiple fill materials, fill levels, strikers, and strike positions. A random sample-level 80/20 split will likely place different recordings of the same container in both train and test. Because FillGauss conditions on dense 3DGS geometry (Section IV.B-C), the geometry encoder can act as a near-unique object fingerprint: at test time, the model can retrieve object-specific resonances rather than predict them from geometry plus fill state. This would inflate every reported metric in Tables I-III, including FAD, KL, MOS, and attribute matching, and would invalidate the claim that FillGauss generalizes to new 3D-scanned containers. The concern is sharpened by the ablation in Table IV: removing the geometry cross-attention causes the largest FAD degradation, consistent with the geometry stream carrying object-identifying information. No code or split metadata is released, so the reader cannot verify disjointness. The FAD conflict with TangoFlux (1.35 vs 1.06) further undermines the literal 'state-of-the-art' phrasing, but the leakage risk is the more load-bearing issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a new task, fine-grained filling-aware impact sound generation, and contributes FillImpact, a dataset of over 5,000 recordings from 88 real containers with varied fill materials, fill levels, strikers, and strike positions, together with 3DGS reconstructions. It then proposes FillGauss, a latent-diffusion framework that fuses 3DGS geometric features, a 3D strike coordinate, and text-described physical state through self- and cross-attention to condition audio generation on the internal fill state. The paper reports physical acoustic analysis of the dataset, quantitative comparisons against SonicGauss and TangoFlux, subjective listening tests, ablations, and sparsity sensitivity experiments, claiming a new state-of-the-art for physically grounded cross-modal audio generation. The central novelty is conditioning on internal fill state rather than only surface geometry, and this is a promising direction.","tokens_in":14290,"tokens_out":7169,"duration_ms":85528,"significance":"If the empirical claims are sustained, the work would be a valuable contribution: it defines a concrete task, provides a controlled multimodal dataset with physical annotations, and proposes a principled architecture that combines 3DGS geometry with internal-state conditioning. The physical analysis of the dataset (e.g., frequency shift with fill level, distinct decay for granular vs. liquid fill, striker hardness effects) is a useful prior for future work, and the ablations show the contribution of each component. However, the central generalization claim to unseen 3D-scanned containers is currently unsupported because the train/test split is not stated to be object-disjoint. In addition, the headline state-of-the-art claim conflicts with the paper's own primary metric, and the subjective evaluation has internal inconsistencies. With a corrected evaluation protocol and appropriately qualified claims, the contribution would be solid and publishable.","major_comments":[{"comment":"The 80/20 split is described only as a 'partition of FillImpact Dataset into an 80% training split and a 20% testing split,' with no statement that the 88 physical objects are split disjointly. Because each object appears in multiple recordings across fill levels, fill materials, strikers, and strike positions, a sample-level split will likely place the same object in both training and test sets. The 3DGS geometry stream in the cross-attention fusion (Eq. 4) can then act as a near-unique object fingerprint, letting the model retrieve object-specific resonances rather than predict them from geometry plus fill state. This would inflate every metric in Tables I–IV and invalidates the claim of generalization to unseen containers. Please report an object-disjoint split with per-object counts, and/or add a held-out-object evaluation.","section":"Section V.A (Implementation Details)"},{"comment":"FAD is declared the primary objective metric (Section V.A), but TangoFlux FT achieves a better FAD (1.0627) than FillGauss (1.3500), and FillGauss is also worse than SonicGauss FT on KL sig (0.6977 vs. 0.5750). The unqualified 'state-of-the-art' claims in the Abstract and Section V.B are therefore not supported by the paper's own primary metric. Either restrict the claim to the best 3D-aware method, or provide a principled argument for why a worse FAD is acceptable (e.g., TangoFlux's lack of spatial grounding) and support it with a metric that privileges 3D conditioning.","section":"Table I vs. Abstract"},{"comment":"There is an inconsistency in the reported number of participants: Section V.A states 8 participants, while Appendix C states 20. Additionally, Table II shows FillGauss MOS (4.43) slightly below SonicGauss ZS (4.50), yet FillGauss has a 69.6% win rate; this is not explained. Because the subjective results are central to the physical-grounding claims, the participant count must be corrected and the MOS/win-rate relationship interpreted. Statistical significance testing for the subjective comparisons should also be reported.","section":"Section V.A / Appendix C / Table II"},{"comment":"The task is described as requiring a 'continuous fill level' in the Abstract and IV.A, but Section III states the fill level is discretized into five states (level 0 = hollow). 'Fine-grained' conditioning and interpolation across arbitrary fill levels is not demonstrated with five discrete levels. Clarify the actual conditioning variable and, if continuous interpolation is claimed, add an experiment on unseen intermediate fill levels.","section":"Abstract / Section III / Section IV.A"}],"minor_comments":[{"comment":"Table III lists only four attributes (Striker Mat., Pos., Fill Mat., Level), but the text says five attributes including container material are evaluated. Add the missing column and report its matching rate.","section":"Table III"},{"comment":"The metric 'KL sig' is not defined; specify how the KL divergence is computed and what 'sig' denotes. Also, FAD is described as 'reference-free,' which is misleading; FAD compares generated and reference feature distributions via VGGish.","section":"Section V.A"},{"comment":"No code, checkpoints, or split metadata are released. If the object-disjoint split is adopted, release the object IDs and random seed for reproducibility.","section":"Reproducibility"},{"comment":"The exact normalization used for 'normalized dominant frequency' is not stated. Clarify whether it is per-object or per-level normalization, so the claimed physical trend can be reproduced.","section":"Figure 4 / Section III"}],"recommendation":"major_revision","confidential_remarks":"The object-disjoint split issue is the deciding factor. If the authors cannot demonstrate object-disjoint evaluation, the empirical contribution is severely undermined regardless of the architecture. The FAD/SOTA inconsistency and the participant-count mismatch also need correction. The dataset itself is a useful resource and the framework is reasonable, so I recommend major revision rather than rejection at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. The FillImpact dataset is a genuinely useful resource, and the fill-aware generation task is new. But the evaluation has a load-bearing ambiguity: the train/test split is not stated to be object-disjoint, and the paper calls itself SOTA while its own primary metric shows a baseline beating it.\n\nFillGauss is worth a look because of the dataset, not the architecture. FillImpact — 88 objects, over 5,000 recordings, water/rice fills at multiple levels with three strikers — is exactly the kind of resource that's missing when people treat impact sounds as a function of outer geometry alone. The physical consistency checks (frequency dropping with added mass, granular vs liquid damping, striker brightness) are simple but useful, and they give the dataset face validity. If the authors release it, it will get used.\n\nThe model itself is a reasonable composition of existing pieces: a SplatFormer-style 3DGS encoder, a position encoder, a text encoder, and a TangoFlux latent diffusion backbone with cross-attention fusion. Nothing architecturally exotic, but the ablations show the components matter and the human attribute-matching results — particularly for fill material and level — are what you'd want to see if the conditioning is working. The design choices are defensible.\n\nThe soft spots are in the evidence. First, Table I shows TangoFlux fine-tuned has FAD 1.06, FillGauss has 1.35. The paper's caveat that TangoFlux isn't 3D-aware is fine in a comparison table, but the abstract and conclusion call FillGauss SOTA without that caveat. That's overclaiming. Second, and more serious: the implementation section says only that the dataset is split 80/20, with no statement that the 88 objects are split disjointly. Since each object appears in many configurations, a sample-level split will put the same geometry in train and test. The 3DGS encoder can then act as an object fingerprint, and the model can retrieve object-specific resonances instead of predicting from geometry plus fill state. This would inflate FAD, MOS, and attribute matching, and it would invalidate the generalization claim for unseen containers. The ablation in Table IV, where dropping geometry cross-attention hurts the most, is consistent with the geometry stream carrying identity. This needs to be resolved — either by stating the split is object-disjoint or by redoing the experiment. Third, the subjective section says 8 participants in the main text and 20 in the appendix. Minor, but sloppy. No code or data is released, which makes verification harder.\n\nThe central idea is credible and the dataset is a real contribution. The leakage concern is a missing detail rather than a demonstrated fraud, but it's load-bearing. I'd send this to peer review with a clear request: fix the split description or the experiment, fix the SOTA claim, and release at least the split metadata.","headline":"Useful new dataset and a plausible fill-aware conditioning scheme, but the evaluation overreaches: the split may leak object identity and the SOTA claim trips over the paper's own FAD table.","tokens_in":14774,"tokens_out":3945,"would_cite":true,"duration_ms":43839,"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":"Containers sound different when filled, and a 3D-aware diffusion model can synthesize that difference.","keywords":["Impact sound generation","3D Gaussian Splatting","latent diffusion","filling-aware audio","cross-modal synthesis","FillImpact dataset","acoustic resonance","audio-visual generation"],"falsifier":"An object-disjoint evaluation: hold out all recordings of a set of containers and measure FAD on impacts of those unseen containers. If FillGauss's FAD degrades substantially compared to the random-split result, the claimed physical generalization is not established. A second check is to synthesize audio for the same 3D scan with two different fill descriptions and verify the dominant-frequency shift quantitatively.","tokens_in":13848,"feed_emoji":"🔊","tokens_out":4255,"duration_ms":50261,"temperature":0.7,"pith_summary":"The paper argues that the sound of a struck container depends on its internal fill state, not just its outer geometry, and that existing 3D-aware audio generators fail because they treat objects as hollow shells. To fix this, it introduces FillImpact, a dataset of over 5,000 real impact recordings from 88 containers with controlled fill materials (water and rice), fill levels, and striker materials, plus 3D Gaussian Splatting scans and exact 3D strike coordinates. It then proposes FillGauss, a latent-diffusion framework that fuses 3D geometry, strike position, and a textual description of striker and fill state. Experiments report that FillGauss produces audio whose dominant frequency drops with fill level, whose damping differs between water and rice, and whose brightness differs by striker hardness, claiming state-of-the-art physical grounding for cross-modal audio generation. A sympathetic reader should care because this is a step toward interactive and robotic settings where the internal state of a container is visually hidden but acoustically audible.","feed_headline":"Fill-aware impact sounds emerge from 3D scans and a text prompt","feed_subtitle":"A diffusion model fuses container geometry, strike point, and internal fill state to synthesize plausible impact audio.","key_machinery":"The load-bearing mechanism is the cross-modal attention fusion pair: context self-attention combines the positional embedding of the strike point with the text embedding of the physical state into a unified excitation context; then geometry cross-attention lets the 3D Gaussian features query that context, so each surface region selectively absorbs positional and damping semantics. This fused embedding conditions a diffusion-transformer denoiser, trained with Huber loss, that predicts the audio latent. The dataset's verified physical trends (dominant frequency decreasing with fill level, higher decay for rice than water, higher spectral centroid for steel) provide the prior that the model is","core_discovery":"On its own terms, the central discovery is that internal filling state is a first-class acoustic condition: adding mass lowers resonance frequency according to the harmonic-oscillator relation f ∝ sqrt(k/m), granular fill suppresses resonance through high damping, and harder strikers shift energy to higher frequencies. FillGauss represents an object as 3D Gaussians, encodes the exact strike point with positional encoding, encodes fill and striker as natural-language text, and uses a self-attention-plus-cross-attention fusion to make the 3D surface query the localized physical context. The fused condition guides a pretrained latent audio diffusion model. The paper shows qualitative spectrogra","pith_inferences":["If fill state is truly acoustically discriminable, then the audio itself could be used to infer the fill state of an unseen container, a task the paper does not attempt but its dataset enables.","The paper's argument implies a failure mode for purely geometric 3D-aware models: they will average over fill states and produce a generic container sound; one could test this by comparing a hollow-only model on filled containers.","The five-level discretization of fill could be extended to continuous regression or to dynamic processes like sloshing, which the paper lists as future work.","A practical extension would be to train on object-disjoint splits to verify generalization beyond the 88 recorded containers."],"forward_implications":["If FillGauss works as claimed, interactive 3D scenes can add plausible impact sounds for containers whose fill state changes, without re-recording audio for every state.","The dependence on a textual fill description means fill state can be controlled at inference time: the same 3D scan can sound hollow, water-filled, or rice-filled on demand.","Because the model is conditioned on exact 3D strike coordinates, sounds can vary continuously with strike position across the surface, enabling positional audio feedback in virtual reality and robotic manipulation.","The finding that fine-grained text prompts outperform hard-coded categorical embeddings suggests pretrained language models are an effective carrier for physical condition semantics.","The sparsity experiment indicates that a compressed 3D Gaussian representation still yields acceptable audio, pointing to a practical route for real-time synthesis."],"fun_headline_variants":["Impact sounds learn what's inside a 3D scan","Fill levels change how a 3D object sounds when hit","3D splats plus text: impact audio that knows its fill","Internal fill shapes impact sound—now modeled in 3D"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reported 80/20 train/test split is described only over samples, not over objects; if the same container appears in both training and testing at different fill levels, the model could memorize object-specific resonances rather than learn transferable physics.","fun_headline_variants_meta":{"raw":{"variants":["Impact sounds learn what's inside a 3D scan","Fill levels change how a 3D object sounds when hit","3D splats plus text: impact audio that knows its fill","Internal fill shapes impact sound—now modeled in 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000828,"raw_usage":{"total_tokens":3480,"prompt_tokens":798,"completion_tokens":2682,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":2611}},"tokens_in":542,"tokens_out":2682,"duration_ms":20733,"temperature":1.0,"reasoning_tokens":2611,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T17:02:45.953581+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An object-disjoint evaluation: hold out all recordings of a set of containers and measure FAD on impacts of those unseen containers. If FillGauss's FAD degrades substantially compared to the random-split result, the claimed physical generalization is not established. A second check is to synthesize audio for the same 3D scan with two different fill descriptions and verify the dominant-frequency shift quantitatively.","supporting_citations":[],"review_version":1}