{"id":"ad6ce4c5-011f-4c3d-8767-e0ab713b730e","arxiv_id":"2606.31556","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"AugSplat augments Gaussian Splatting with synthetic views from a pre-trained radiance field to improve sparse-view reconstruction quality on mip-NeRF 360 scenes while keeping real-time rendering.","lead":"AugSplat uses a radiance field trained on sparse input views to synthesize extra images that supervise Gaussian Splatting optimization. A smart generalist might read it to see a practical way to improve 3D reconstruction quality when only a few photos are available.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Synthetic views from sparse-view radiance field may contain artifacts that degrade rather than improve Gaussian Splatting optimization","rationale":"The reader's weakest_assumption is precisely the load-bearing condition for the method; the abstract alone supplies no evidence that this condition is satisfied, so the UNVERDICTED verdict remains appropriate even after the full text is consulted.","tokens_in":1786,"tokens_out":321,"duration_ms":15341,"concrete_test":"Report the radiance field's PSNR/SSIM on a held-out real test view for each scene; then ablate by replacing the synthetic views with either (a) ground-truth novel views or (b) the same views corrupted by the observed radiance-field error level. If AugSplat's gain disappears under (b) or is smaller than under (a), the auxiliary-supervision assumption does not hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that a radiance field trained solely on the sparse real views produces novel-view images whose geometry and appearance are sufficiently accurate to serve as auxiliary supervision. In sparse-view mip-NeRF 360 regimes the radiance field itself is known to produce floaters, incorrect densities, and view-dependent artifacts; if these appear in the synthesized images, the staged variant (initial optimization on synthetic data) can lock the Gaussians into incorrect positions before real-image supervision begins, while the dual variant can still inject noisy gradients throughout training. The abstract provides no quantitative check that the radiance-field quality on held-out real views correlates with the reported GS improvement.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces AugSplat, a framework that trains a radiance field on sparse input views to synthesize additional images from novel viewpoints, which are then used as auxiliary supervision for Gaussian Splatting optimization. Two variants are studied: Staged AugSplat (initial optimization on synthetic views followed by real-image supervision) and Dual AugSplat (joint optimization with a decaying weight on the synthetic loss). The central claim, based on experiments with sparse-view mip-NeRF 360 scenes, is that both variants improve reconstruction quality over standard Gaussian Splatting, with the staged variant achieving the strongest average performance, while preserving real-time rasterization at inference.","tokens_in":1897,"tokens_out":511,"duration_ms":21170,"significance":"If the empirical results hold and the synthetic views prove beneficial without introducing harmful artifacts, the work provides a simple, practical bridge between radiance fields and Gaussian Splatting for sparse-view regimes. The explicit comparison of staged versus dual supervision schedules is a useful design contribution, and the emphasis on maintaining real-time inference is aligned with application needs.","major_comments":[{"comment":"Abstract: The statement that 'Experiments on sparse-view mip-NeRF 360 scenes show that AugSplat improves reconstruction quality over standard Gaussian Splatting' and that 'Staged AugSplat achieves the strongest average performance' is presented without any quantitative metrics (PSNR, SSIM, LPIPS), baseline tables, or error analysis. This absence makes the central empirical claim impossible to evaluate from the provided information.","section":"Abstract"},{"comment":"Method description (and Experiments section): No quantitative validation is supplied that the radiance field, trained only on the sparse real views, produces synthetic images whose geometry and appearance are sufficiently accurate to serve as auxiliary supervision. In particular, there is no reported correlation between radiance-field quality on held-out real views and the observed GS improvement, leaving the key assumption that synthetic views are net beneficial (rather than artifact-inducing) untested.","section":"Method / Experiments"}],"minor_comments":[{"comment":"The abstract would be clearer if it briefly named the evaluation metrics and the number of scenes or views used in the mip-NeRF 360 experiments.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The manuscript as presented supplies no experimental results or figures, which raises a question about whether the full results section was included in the version under review."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments on our manuscript. Below we provide point-by-point responses to the major comments and indicate the revisions we will make.","responses":[{"response":"We agree that the abstract would be strengthened by including concrete quantitative results. In the revised version we will update the abstract to report the key average improvements (e.g., PSNR, SSIM, LPIPS) achieved by Staged and Dual AugSplat relative to standard Gaussian Splatting on the sparse-view mip-NeRF 360 scenes, with explicit references to the corresponding tables.","revision_made":"yes","referee_comment":"[Abstract] Abstract: The statement that 'Experiments on sparse-view mip-NeRF 360 scenes show that AugSplat improves reconstruction quality over standard Gaussian Splatting' and that 'Staged AugSplat achieves the strongest average performance' is presented without any quantitative metrics (PSNR, SSIM, LPIPS), baseline tables, or error analysis. This absence makes the central empirical claim impossible to evaluate from the provided information."},{"response":"The observation is correct: the current manuscript does not provide a direct quantitative assessment of radiance-field accuracy on held-out views or a correlation analysis with the observed Gaussian Splatting gains. While the end-to-end improvements serve as indirect support, we acknowledge that an explicit validation would be valuable. We will add a short analysis in the Experiments section that reports radiance-field metrics on held-out real views and examines their relationship to the final reconstruction quality.","revision_made":"yes","referee_comment":"[Method / Experiments] Method description (and Experiments section): No quantitative validation is supplied that the radiance field, trained only on the sparse real views, produces synthetic images whose geometry and appearance are sufficiently accurate to serve as auxiliary supervision. In particular, there is no reported correlation between radiance-field quality on held-out real views and the observed GS improvement, leaving the key assumption that synthetic views are net beneficial (rather than artifact-inducing) untested."}],"tokens_in":1481,"tokens_out":435,"duration_ms":21409,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core move is to train a radiance field on the limited real views, render extra images from nearby angles, and feed those into Gaussian Splatting as extra supervision. Staged AugSplat runs an initial phase on the synthetic data before switching to real images; Dual keeps both losses active with a decaying weight on the synthetic term. This is a direct response to the weak initial point cloud that usually hurts sparse-view GS.\n\nThe approach is new as a targeted use of radiance-field view synthesis for GS augmentation rather than a rehash of prior work. It keeps the real-time rasterization advantage of GS and tries to improve quality without extra real captures. The claim that staged performs best on average across sparse mip-NeRF 360 scenes is at least a concrete experimental direction.\n\nThe obvious soft spot is the quality of the synthetic views themselves. Radiance fields trained on sparse inputs commonly produce floaters, wrong densities, and view-dependent errors. If those artifacts reach the GS stage they can lock in bad geometry early (staged) or keep injecting noise (dual). The abstract asserts improvement but supplies no PSNR, SSIM, LPIPS, baseline tables, or correlation between radiance-field held-out error and final GS quality, so the central assumption cannot be checked. The stress-test concern therefore stands on the information given.\n\nThis is for people already working on real-time novel-view synthesis from few images. A reader who needs a practical tweak to existing GS pipelines might extract value once the numbers appear; it is not aimed at shifting the broader field.\n\nI would send it for peer review. The idea is straightforward and the two variants are clearly motivated, but the paper needs the missing quantitative checks and artifact analysis to be taken seriously.","headline":"AugSplat adds synthetic views from a sparse-trained radiance field to supervise Gaussian Splatting, with staged and dual variants, but the abstract gives no numbers so the gains are unverified.","tokens_in":2419,"tokens_out":434,"would_cite":false,"duration_ms":24810,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"AugSplat uses a radiance field trained on sparse inputs to synthesize auxiliary views that improve Gaussian Splatting optimization.","keywords":["Gaussian Splatting","Radiance Fields","Sparse-view Reconstruction","View Augmentation","Novel View Synthesis","3D Reconstruction","Real-time Rendering"],"falsifier":"A direct comparison showing that Gaussian Splatting models trained with the added synthetic views achieve lower PSNR, SSIM, or LPIPS scores than models trained only on the original sparse real views would falsify the benefit of the augmentation.","tokens_in":2674,"feed_emoji":"🖼️","tokens_out":623,"duration_ms":22588,"temperature":0.7,"pith_summary":"The paper shows that a radiance field trained on limited real views can generate synthetic images from nearby novel viewpoints, and these images serve as extra supervision during Gaussian Splatting training. This matters because standard Gaussian Splatting optimization degrades when initial geometry from sparse views is incomplete or noisy, while radiance fields handle sparsity more robustly but render too slowly for real-time use. Combining the two through staged or dual supervision produces higher-quality reconstructions that still run at real-time speeds at inference. Experiments on sparse-view mip-NeRF 360 scenes confirm gains over plain Gaussian Splatting, with the staged variant performing best on average.","feed_headline":"Radiance fields boost sparse-view Gaussian Splatting quality","feed_subtitle":"Synthetic images generated from a radiance field trained on limited inputs raise reconstruction metrics while real-time rendering is retaine","key_machinery":"Radiance field-informed view augmentation: a radiance field trained on sparse real views generates synthetic images used as auxiliary supervision for Gaussian Splatting optimization.","core_discovery":"A radiance field is first trained on the given sparse real views. It then synthesizes additional images from nearby novel viewpoints that increase view-space coverage. These synthetic images supply auxiliary supervision to Gaussian Splatting: Staged AugSplat optimizes first on the synthetic set before switching to real images, while Dual AugSplat trains jointly on both sets with a decaying weight on the synthetic loss. Both variants raise reconstruction quality on sparse-view mip-NeRF 360 scenes relative to standard Gaussian Splatting while preserving real-time rasterization at test time.","pith_inferences":["The same augmentation pattern could be tested on other fast-rendering methods that also depend on good initial geometry.","Applying the staged schedule to progressively sparser input sets would reveal how much view density the radiance field can compensate for.","The method implicitly suggests that any slow but robust sparse-view reconstructor could bootstrap a faster one through synthetic data.","Real-world captures with uneven camera distributions would be a direct test of whether the synthesized views remain helpful outside controlled benchmarks."],"forward_implications":["Staged AugSplat reaches the highest average performance across the tested sparse-view scenes.","Dual AugSplat delivers nearly matching quality while keeping real-image supervision active throughout training.","Both variants outperform standard Gaussian Splatting without sacrificing real-time rendering speed.","The approach increases effective view coverage available for supervision in sparse regimes."],"fun_headline_variants":["Radiance fields guide Gaussian Splatting for sparse views","Synthetic novel views aid Gaussian Splatting optimization","AugSplat applies radiance field augmentation to splatting","Staged and dual training on synthetic views for splatting"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The radiance field trained on the sparse real views produces synthetic images of high enough quality to supply beneficial auxiliary supervision without introducing artifacts that degrade the final Gaussian Splatting model.","fun_headline_variants_meta":{"raw":{"variants":["Radiance fields guide Gaussian Splatting for sparse views","Synthetic novel views aid Gaussian Splatting optimization","AugSplat applies radiance field augmentation to splatting","Staged and dual training on synthetic views for splatting"]},"model":"grok-4.3","cost_usd":0.00911,"raw_usage":{"total_tokens":4126,"prompt_tokens":749,"num_sources_used":0,"completion_tokens":53,"cost_in_usd_ticks":91099500,"prompt_tokens_details":{"text_tokens":749,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3324,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":749,"tokens_out":53,"duration_ms":28757,"temperature":1.0,"reasoning_tokens":3324,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-01T05:48:25.634187+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A direct comparison showing that Gaussian Splatting models trained with the added synthetic views achieve lower PSNR, SSIM, or LPIPS scores than models trained only on the original sparse real views would falsify the benefit of the augmentation.","supporting_citations":[],"review_version":1}