{"id":"da614edb-f596-4b81-8a17-4183cf5dec1e","arxiv_id":"2412.14166","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A 700K-scene procedural, non-semantic synthetic dataset improves large reconstruction models by 1.2 to 1.8 dB PSNR when combined with real data.","lead":"The paper introduces MegaSynth, a set of 700,000 procedurally generated, non-semantic 3D scenes used to train feed-forward scene reconstruction models. Training on MegaSynth plus real data improved novel-view PSNR by 1.2 to 1.8 dB in the paper's experiments, though the reported gains are entangled with an extra geometry-supervision signal.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 1.2–1.8 dB gain is confounded with the geometry loss Lloc: Table 2 row (1) vs (2) shows Lloc adds 3.25 dB, and MegaSynth without Lloc underperforms real-only (21.87 vs 23.89 dB), so the headline improvement cannot be attributed to the synthetic data distribution.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing threat to the central claim. The paper's headline promise is that non-semantic synthesized scenes improve reconstruction, and the supporting experiments compare models trained on DL3DV plus MegaSynth (with Lloc) against models trained on DL3DV alone (without Lloc). This two-variable change leaves the contribution of the data distribution unidentified. The paper's own Table 2 row (1) makes the problem concrete: when controlled MegaSynth data is used without Lloc, the fine-tuned model reaches only 21.87 dB, below the DL3DV-only baseline of 23.89 dB. Adding Lloc raises it to 25.12 dB, which is larger than the entire reported gain. I therefore agree with the reader that the central claim does not follow as worded. I considered whether Table 5 (Kubric and Front3D) already addresses this by showing other synthetic datasets with ground-truth geometry do not help; that would weaken the concern. However, the paper does not state whether Lloc was enabled for those runs, and even if it was, the row (1) versus (2) ablation remains: MegaSynth without Lloc underperforms real data, so the data distribution alone is not shown to be beneficial. The 'comparable' claim for MegaSynth-only training is also contradicted by Table 4 (21.50 vs 23.89 dB). Because the concern is already the basis for the reader's conditional verdict, my assessment does not change the verdict. The requested ablation — disabling Lloc in the full-scale MegaSynth pipeline — would settle the attribution and should be added before publication.","tokens_in":19937,"tokens_out":5730,"duration_ms":46403,"concrete_test":"Train GS-LRM under the full 700K MegaSynth setting with γ=0 in Ltotal = Limg + γ·Lloc (i.e., disable Lloc), then fine-tune on DL3DV exactly as in Table 1, and evaluate on Hypersim and DL3DV at resolution 128. If PSNR does not exceed the DL3DV-only baseline (23.89 dB on Hypersim), the 1.2–1.8 dB gain is attributable to Lloc, not to the MegaSynth data distribution. As a complementary check, run the Table 5 comparison with Lloc enabled on DL3DV + Kubric and DL3DV + Front3D; if those also improve substantially, the effect is generic geometry supervision rather than MegaSynth-specific data design.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the observed improvement is caused by the MegaSynth scene distribution. However, every with-MegaSynth model is trained with an additional geometry loss Lloc (Eq. 6) on the synthetic examples, which the real-only baseline cannot use because real datasets lack ground-truth depth. Table 2 isolates this: the real-data-tuned model improves from 21.87 dB (row 1, controlled data without Lloc) to 25.12 dB (row 2, with Lloc) on Hypersim at resolution 128 — a 3.25 dB jump, larger than the entire headline 1.2–1.8 dB gain. Critically, row (1) without Lloc (21.87 dB) is below the DL3DV-only baseline of 23.89 dB in Table 1, meaning the synthetic data alone actually degrades performance relative to real data. The cross-over is driven by Lloc, not by the data distribution. Table 4's 'MegaSynth-only' row also uses Lloc and yields 21.50 dB versus 23.89 dB for real-only, contradicting the abstract's 'comparable' claim by 2.39 dB. The comparison with Kubric and Front3D in Table 5 could mitigate the concern if those runs used Lloc and still failed to improve, but the paper does not specify whether the same geometry supervision was enabled for them. As written, the abstract attributes the gain to synthesized scenes, while the paper's own ablation attributes it to the auxiliary geometry supervision that only synthesized data can provide.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MegaSynth, a procedurally generated dataset of 700K non-semantic 3D scenes for training feed-forward large reconstruction models (LRMs). The generation pipeline composes primitive shapes with random textures and lighting within a controlled floor plan, then samples camera poses and renders RGB and depth supervision. The authors train GS-LRM and Long-LRM on MegaSynth combined with the real DL3DV dataset, using either joint training or pre-training/fine-tuning, and report consistent PSNR improvements of 1.2–1.8 dB over real-data-only baselines across in-domain and out-of-domain benchmarks. They further claim that models trained solely on MegaSynth perform comparably to real-data-trained models, and demonstrate an application to monocular depth estimation.","tokens_in":20377,"tokens_out":8725,"duration_ms":70986,"significance":"If the claims are substantiated, the work would show that non-semantic synthetic scenes can provide scalable and effective training data for scene-level reconstruction, potentially reducing dependence on expensive real-world capture. The dataset itself (700K scenes), the detailed ablations of data controllability and scale, and the released code/data are valuable contributions. However, the central attribution of the performance gains to the synthetic scene distribution is currently confounded with an auxiliary geometry loss, so the significance of the data-distribution claim is not yet established.","major_comments":[{"comment":"The headline 1.2–1.8 dB gain is confounded with the geometry loss Lloc. In Table 2, adding Lloc (row 2 vs row 1) raises the real-data-tuned Hypersim PSNR from 21.87 to 25.12 dB, a 3.25 dB gain larger than the entire reported improvement. More importantly, row (1) (MegaSynth without Lloc) reaches only 21.87 dB, below the DL3DV-only baseline of 23.89 dB in Table 4; that run is also marked as unstable (Fail. Iter. 57k). Thus, MegaSynth data alone does not improve over real data, and the improvement in Table 1 should be attributed to the combination of synthetic data and the geometry supervision it enables, not to the synthetic scene distribution per se. Please provide an ablation that isolates the data distribution from the supervision signal, e.g., MegaSynth with and without Lloc alongside a real-data baseline with a comparable auxiliary loss, and adjust the wording of the central claim accordingly.","section":"Sec. 5.3, Eq. (6), Table 2"},{"comment":"The abstract claim that models trained solely on MegaSynth 'perform comparably' to real-data-trained models is not supported by the paper's own numbers: MegaSynth-only achieves 21.50 dB PSNR on Hypersim versus 23.89 dB for DL3DV-only, a 2.39 dB gap larger than the headline gains. Since the paper elsewhere treats differences on the order of 1 dB as meaningful, this gap is substantial. Please either report additional benchmarks where the gap is smaller, provide statistical significance, or soften the claim to reflect that synthetic-only training substantially underperforms real-data training.","section":"Sec. 6.4, Table 4"},{"comment":"The comparison with other synthetic datasets (Kubric and Front3D) does not state whether the geometry loss Lloc was enabled in those runs. Because Table 2 shows Lloc to be the dominant factor in the MegaSynth gains, the conclusion that 'realistic 3D assets or scene composition is not the guarantee for improving reconstruction quality' is not established unless the auxiliary supervision is held constant across all synthetic data conditions. Please clarify the loss configurations for each row, and ideally re-run the comparison with the same geometry loss enabled for Kubric and Front3D.","section":"Sec. 6.5, Table 5"}],"minor_comments":[{"comment":"The dataset name is misspelled as 'Kurbic'; it should be 'Kubric'.","section":"Table 5 caption"},{"comment":"The phrase 'trained sorely with MegaSynth' should read 'trained solely with MegaSynth'.","section":"Sec. 7"},{"comment":"The column headers for the three boolean factors (Control, LSloc, Scale) are not clearly labeled in the table as printed; please make the factors explicit in the caption.","section":"Table 2"},{"comment":"The term 'non-semantic' is used in several senses (no semantic classes, no affordances, no object relationships); please provide an operational definition at first use.","section":"Sec. 4"},{"comment":"The user study reports only averaged rankings; please include the number of participants, the rating instructions, and inter-rater agreement.","section":"Table 7"},{"comment":"The fine-tuning protocol for Depth Anything V2 on MegaSynth is not described; please provide details in the appendix.","section":"Table 6"},{"comment":"The generation time of '700K scenes in 3 days' should be accompanied by the compute environment used for rendering.","section":"Figure 1 caption"}],"recommendation":"major_revision","confidential_remarks":"The dataset and pipeline are likely a useful contribution, and the authors already have the ablation structure (Table 2) that could resolve the main confound with a few additional runs. I would support a major revision rather than rejection, provided the authors either disentangle the data-distribution effect from Lloc or re-frame the central claim to cover the full synthetic pipeline including its metadata."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things up front. First, this is a serious dataset paper: 700K procedurally generated non-semantic scenes, 50x larger than DL3DV, with thoughtful controls on geometry complexity, camera sampling, materials, and lighting, plus code/data links. Second, the central headline—that MegaSynth improves reconstruction by 1.2–1.8 dB—is confounded with the geometry loss Lloc. Every with-MegaSynth model uses Eq. 6 on synthetic depth; the real-only baselines cannot. Table 2 shows Lloc alone accounts for 21.87→25.12 dB (a 3.25 dB jump), and the same table shows MegaSynth without Lloc leaves the tuned model at 21.87, below the DL3DV-only baseline of 23.89. So the cross-over is driven by the auxiliary geometry supervision, not by the non-semantic scene distribution as claimed. The stress-test note gets this right.\n\nCredit where due: the scene-level extension of primitive-based synthesis (beyond LRM-Zero's objects) is a real step, and the ablation on data control, scale, and metadata in Table 2 is more informative than most dataset papers. The result that synthetic-only training reaches 21.50 dB against 23.89 for real-only is not 'comparable,' but it is closer than one might expect for zero semantics, and it supports the low-level reconstruction thesis better than the 'comparable' wording does.\n\nSoft spots: the abstract overstates. A 2.4 dB gap is not comparable. There are no error bars or multiple seeds on the 1.2–1.8 dB claim—for comparisons decided by ~1 dB, that matters. And Table 5's comparison to Kubric/Front3D doesn't state whether those runs also used Lloc, which is exactly the confound that needs controlling. The fix is straightforward: ablate Lloc on the full 700K setting, and ideally run a real-data baseline with a geometry proxy (COLMAP depth or a mono-depth pseudo-GT) to isolate the data distribution from the supervision signal.\n\nBottom line: the dataset will likely be useful to people training scene-level LRMs, and the non-semantic thesis is testable and interesting. But as written the experiments do not support the abstract's attribution of the gain to the synthesized scene distribution. I'd send this to review and ask for those ablations. It deserves a serious referee.","headline":"The MegaSynth dataset is a serious engineering contribution, but the headline PSNR gain is confounded with the geometry loss Lloc, and the 'comparable to real data' claim is overstated.","tokens_in":20887,"tokens_out":3781,"would_cite":false,"duration_ms":32837,"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":"MegaSynth, a 700K-scene procedurally generated dataset, improves feed-forward 3D scene reconstruction by 1.2–1.8 dB PSNR and matches real-data training when used alone.","keywords":["MegaSynth","procedural generation","large reconstruction model","3D scene reconstruction","non-semantic primitives","Gaussian splatting","novel view synthesis","synthetic training data"],"falsifier":"Train GS-LRM on MegaSynth renderings without the geometry loss $L_{\\text{loc}}$ and compare to training on real data alone; if the PSNR advantage vanishes or reverses, the claim that the synthetic data itself drives reconstruction quality is falsified.","tokens_in":19798,"feed_emoji":"🏗️","tokens_out":5008,"duration_ms":41447,"temperature":0.7,"pith_summary":"Feed-forward 3D scene reconstruction models improve when trained on procedurally generated scenes that contain no semantic information. The paper introduces MegaSynth, a 700K-scene synthetic dataset built from shape primitives and controlled lighting, and shows that adding it to real data raises novel-view PSNR by 1.2–1.8 dB across in-domain and out-of-domain benchmarks. Models trained only on MegaSynth perform comparably to ones trained on real scenes, supporting the claim that multi-view reconstruction depends mainly on low-level geometry and appearance rather than on scene semantics. The synthetic pipeline also supplies exact depth and camera metadata, enabling a geometry loss that stabilizes training and strengthens reconstruction.","feed_headline":"700K synthetic scenes boost 3D reconstruction by up to 1.8 dB","feed_subtitle":"Non-semantic procedurally generated data matches real-scene training and sharpens generalization.","key_machinery":"The central mechanism is a procedural generator that builds scenes from non-semantic primitives and provides precise per-pixel depth and camera metadata, which in turn enables a Gaussian-location loss $L_{\\text{loc}}$ that stabilizes training and improves geometry. The generator's controllability over complexity, camera baselines, lighting, and textures allows it to loosely match real-world distributions while eliminating the need for semantically valid scenes.","core_discovery":"MegaSynth is a procedurally generated dataset of 700K scenes, over 50 times larger than the real DL3DV dataset, in which scenes are composed from non-semantic shape primitives (cubes, spheres, cylinders, cones) placed within simple floor plans and lit by randomized ambient, sunlight, and luminous light sources. By removing semantic priors such as object affordances and scene composition, the authors make generation highly scalable and controllable. Training large reconstruction models such as GS-LRM and Long-LRM jointly with MegaSynth and real data, or pre-training on MegaSynth before fine-tuning on real data, improves reconstruction quality by 1.2–1.8 dB PSNR relative to real-data-only training across DL3DV, Hypersim, MipNeRF360, and Tanks & Temples. When trained exclusively on MegaSynth, the model performs comparably to the real-data-trained model, indicating that scene semantics are largely unnecessary for learning multi-view reconstruction.","pith_inferences":["If low-level geometry and appearance are all that matter for multi-view reconstruction, synthetic data generation could expand to unbounded scale with simple procedural rules, making data collection for scene-level models nearly free.","The geometry loss enabled by synthetic depth might be the real driver of gains; a controlled study on real data with pseudo-depth supervision would separate the effects of data distribution from supervision.","Extending the generator to include outdoor-specific structures, such as open skies and distant terrain, could close the remaining indoor-bias gap the paper observes."],"forward_implications":["Joint training and pre-training with MegaSynth improve accuracy for two different LRM architectures, at multiple resolutions, and on both in-domain and out-of-domain test sets.","Models pre-trained on MegaSynth generalize better to out-of-domain scenes than models trained only on real data.","MegaSynth-only training yields geometry and image quality comparable to real-data-only training, suggesting that semantic priors are not necessary for scene-level reconstruction.","MegaSynth also benefits monocular depth estimation when fine-tuning a pretrained model."],"supporting_citations":[{"why":"Supplies the real scene dataset used for joint training and as the main comparison baseline.","marker":"[47]"},{"why":"GS-LRM, one of the two feed-forward reconstruction models trained on MegaSynth.","marker":"[88]"},{"why":"Long-LRM, the other reconstruction model, used to confirm gains generalize across architectures.","marker":"[93]"},{"why":"Introduces the large reconstruction model framework that this work scales up with synthesized data.","marker":"[29]"},{"why":"Prior work on primitive-based synthetic data for object-level reconstruction, extended here to scenes.","marker":"[79]"},{"why":"Kubric dataset used as a comparison synthetic dataset that fails to improve reconstruction.","marker":"[23]"},{"why":"Front3D dataset used as an artist-designed synthetic comparison that does not help reconstruction.","marker":"[20]"},{"why":"Hypersim dataset provides an out-of-domain evaluation set with ground-truth depth.","marker":"[56]"},{"why":"Optimization-based 3DGS baseline for overall performance comparison.","marker":"[37]"}],"fun_headline_variants":["Synthetic scenes rival real data for 3D reconstruction","MegaSynth: 700K synthetic scenes sharpen 3D reconstruction","No semantics needed: synthetic data boosts 3D by 1.8 dB","Synthetic-only training matches real data for 3D","Scaling 3D with procedural scenes, no semantics required"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gains attributed to the synthetic data could actually come from the extra depth supervision that only synthetic data can provide; if that supervision is removed, the improvement may disappear.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic scenes rival real data for 3D reconstruction","MegaSynth: 700K synthetic scenes sharpen 3D reconstruction","No semantics needed: synthetic data boosts 3D by 1.8 dB","Synthetic-only training matches real data for 3D","Scaling 3D with procedural scenes, no semantics required"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1301,"prompt_tokens":967,"completion_tokens":334,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":243}},"tokens_in":583,"tokens_out":334,"duration_ms":3407,"temperature":1.0,"reasoning_tokens":243,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:24:39.147747+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train GS-LRM on MegaSynth renderings without the geometry loss $L_{\\text{loc}}$ and compare to training on real data alone; if the PSNR advantage vanishes or reverses, the claim that the synthetic data itself drives reconstruction quality is falsified.","supporting_citations":[{"cited_title":"Dl3dv-10k: A large-scale scene dataset for deep learning- based 3d vision","cited_arxiv_id":null,"evidence_quote":"Supplies the real scene dataset used for joint training and as the main comparison baseline."},{"cited_title":"Kubric: A scalable dataset generator","cited_arxiv_id":null,"evidence_quote":"Kubric dataset used as a comparison synthetic dataset that fails to improve reconstruction."},{"cited_title":"Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding","cited_arxiv_id":null,"evidence_quote":"Hypersim dataset provides an out-of-domain evaluation set with ground-truth depth."}],"review_version":1}