{"id":"3a2394cc-e086-4fa3-b2d6-2d9dd73e9a28","arxiv_id":"2607.04540","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Automatic view scheduling via a directed generation graph plus object-level identity and adherence conditioning enables high-quality outdoor 3DGS scenes from arbitrary input geometry without user camera paths.","lead":"SceneFrom3D automatically builds a directed generation graph of anchor views and interpolation paths from outdoor object meshes, then synthesizes multi-view images and optimizes a 3D Gaussian scene with per-object appearance and geometry-adherence controls. It removes the need for hand-designed camera trajectories that previously blocked outdoor geometry-conditioned generation.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The SOTA claim rests on unproven multi-view consistency of independently sampled anchors plus video interpolation for large outdoor geometry.","rationale":"The reader correctly isolates the load-bearing assumption that visibility-optimized anchors + video interpolation yield observations complete and consistent enough for 3DGS on unbounded outdoor geometry. That assumption underpins every contribution bullet and the three-stage pipeline; the acknowledged artifacts (Fig. 12, §5) and the custom-only evaluation already justify CONDITIONAL. No stronger internal inconsistency or missing derivation appears in the full manuscript (including supplemental algorithms and parameter tables). The proposed warp-based consistency check is a direct, low-cost falsifier of the same assumption and would either confirm the pipeline’s robustness or quantify how far the current results rely on favorable layouts and post-hoc 3DGS losses. Hence the reader’s verdict and confidence remain appropriate; no adjustment is required.","tokens_in":21565,"tokens_out":546,"duration_ms":20592,"concrete_test":"On the Village and Tribe Town layouts, warp every generated anchor image into its graph neighbors using the input-mesh depth and camera poses, then report mean LPIPS and scale-invariant depth error on the overlapping valid pixels (excluding sky). If mean LPIPS exceeds ~0.25 or systematic lighting/shadow mismatches appear on >30 % of edges, the multi-view observations are not sufficiently consistent and the SOTA claim weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (first trajectory-free outdoor geometry-conditioned generator that is SOTA with object-level control) requires that the directed generation graph of §3.2 produces a compact set of anchors whose independent diffusion samples (FLUX.2 with parent warping and identity/α conditioning, §3.3.1) plus VACE interpolations (§3.3.2) are sufficiently view-consistent and complete for stable 3DGS optimization (§3.4). Soft visibility (Eq. 5), pose optimization (Eq. 10), and shared-visibility edges do not enforce photometric or illumination consistency across anchors; the paper itself shows residual failures (shadow direction flips in Fig. 12; identity overflow when >8 objects, §5). Table 1 metrics are computed on Bézier trajectories derived from the same anchors and on a 9-layout custom set, so they do not independently stress-test whether the observations remain usable when geometry becomes larger or more unstructured. Without that consistency, both the “no trajectories needed” practicality claim and the quality/controllability claims collapse.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"SceneFrom3D proposes a three-stage geometry-conditioned outdoor 3D scene generation pipeline that removes the need for user-provided camera trajectories. From object meshes with per-object identity images and geometry-adherence parameters α_o, it builds a directed generation graph G_gen whose nodes are visibility-optimized anchor views and whose edges are interpolation trajectories with a generation order. Multi-view images are then synthesized by fine-tuned image diffusion (FLUX.2) at anchors and video diffusion (VACE) along edges, and a 3DGS is optimized with RGB, DSSIM, LPIPS, and mesh-depth losses. The paper claims this is the first trajectory-free outdoor geometry-conditioned generator and reports state-of-the-art quality and object-level controllability on custom outdoor layouts versus UrbanArchitect, YoNoSplat, and a drone path-planning scheduler.","tokens_in":21975,"tokens_out":1382,"duration_ms":25566,"significance":"If the results hold, the main contribution is practical and timely: outdoor geometry is large, unstructured, and unbounded, so automatic view scheduling is a real bottleneck that prior indoor heuristics and trajectory-assuming outdoor methods do not solve. Formulating scheduling as construction of a DAG of anchors and interpolation edges is a clean systems idea, and the object-level identity plus α_o adherence controls go beyond global layout conditioning. The paper supplies detailed algorithms (node init/refinement, soft visibility, pose losses), ablations of tilt/repulsion/interpolation, and qualitative control of α, which are genuine engineering strengths. Residual multi-view failures (shadows, identity overflow) are acknowledged. The work is significant for graphics/VR content creation if evaluation breadth and consistency claims are tightened.","major_comments":[{"comment":"Abstract and §1 claim SOTA outdoor geometry-conditioned generation, but §4.2 evaluates only 9 custom layouts (9–16 objects each) with 10 scenes per baseline and no public outdoor benchmark. Table 1 gains are therefore on an author-defined distribution that also supplies the scheduled cameras used by UrbanArchitect and the multi-views used by YoNoSplat. The SOTA claim should be tempered to this setting, or the evaluation expanded (more layouts, larger/unbounded scenes, or an external proxy mesh suite) so the claim is not over-indexed on a small private set.","section":null},{"comment":"The central practical claim—that a compact visibility-optimized anchor set plus video interpolation yields observations stable enough for 3DGS—rests on soft visibility (Eq. 5) and shared-visibility edges (§3.2.2), which measure geometric coverage, not photometric or illumination consistency. §5 and Fig. 12 already show residual shadow-direction flips across anchors, and the >8-identity overflow limit. Table 1 reports CLIP Aesthetic, MUSIQ, PSNR-D, Chamfer, and F-score on Bézier paths built from the same anchors, but no multi-view consistency metric (e.g., cross-view warping error, identity/appearance agreement, or shadow/lighting consistency). Without such measures, it is hard to know when the pipeline fails as geometry grows more unstructured.","section":null},{"comment":"§4.2 baseline design is only partially fair. Zhang et al. (2021) is an aerial reconstruction planner and produces top-down-biased paths (Fig. 9), so poor eye-level quality is expected and does not fully stress alternative outdoor scheduling objectives. UrbanArchitect receives the authors’ scheduled poses; YoNoSplat receives the authors’ generated images and is known to be limited in view count and domain. The paper should either add a stronger outdoor geometry-conditioned baseline with its own trajectory policy (or a simple heuristic outdoor scheduler) or clearly reframe comparisons as stage-wise ablations rather than end-to-end SOTA.","section":null},{"comment":"View scheduling depends on many free thresholds (Table S1: d0, δ_vis, δ_shared, δ_remove, δ_merge, d_safe, etc.) and preferred distance d0=35. §4.3 ablations cover L_tilt, L_rep, and interpolation, but not sensitivity of coverage/quality to these thresholds or to surface-sample spacing h. Because compactness of V and edge reliability are load-bearing for both cost and reconstruction completeness, a short sensitivity study (or failure cases when thresholds are misspecified) is needed to support the claim of automatic scheduling for arbitrary outdoor geometry.","section":null}],"minor_comments":[{"comment":"Fig. 5 is dense; camera labels A–D and colored identity outlines are hard to parse at print scale. Consider a cleaner layout or larger insets for one scenario in the main paper.","section":null},{"comment":"Notation for the generation graph switches between G_gen, script G, and Ggen; unify early in §3.2.","section":null},{"comment":"§3.3.1 states fine-tuning of FLUX.2-klein-9B on a synthetic multi-reference dataset (~17K pairs in the supplement). A one-sentence main-text note on domain of identity categories (22 categories, architecture grid bias fix) would help readers assess generalization without opening the supplement.","section":null},{"comment":"Table 1: the “W/o interpolation” row has the best MUSIQ but worse structural metrics; the text attributes this to high-frequency artifacts. Adding a short note or a perceptual failure example would make that interpretation less hand-wavy.","section":null},{"comment":"Related work on NBV and aerial path planning is appropriately distinguished; a brief pointer to any concurrent outdoor layout-to-3D systems (if any) would strengthen the “first trajectory-free” claim.","section":null},{"comment":"Minor typography: occasional double-struck or duplicated characters in figure captions (e.g., αα, L script variants) should be cleaned for production.","section":null}],"recommendation":"major_revision","confidential_remarks":"Solid systems paper with a genuine outdoor scheduling gap and careful engineering; main risk is overstated SOTA on a small private suite plus unquantified multi-view consistency. Suitable for a graphics venue after the authors either broaden evaluation or narrow claims. No integrity concerns; self-use of VideoFrom3D is appropriately scoped as a generation strategy, not circular proof of the new scheduler."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The real news is the automatic directed generation graph: visibility densification + pose opt + shared-visibility edges that turn an arbitrary outdoor mesh into anchors, interpolation trajectories, and generation order. Prior outdoor work either handed you trajectories (VideoFrom3D) or stayed indoor/heuristic. That is a genuine practical advance for content pipelines, and the object-level identity + α-adherence knobs are cleanly integrated and visibly useful (Fig. 8).\n\nWhat they do well: the three-stage pipeline is carefully engineered, the soft-visibility terms and tilt/repulsion losses are ablated with clear failure modes (Figs. 7, 10, 11), and the qualitative results on village/desert/etc. look better than the three baselines they actually ran (UrbanArchitect, YoNoSplat, Zhang drone planner). Table 1 numbers move in the right direction on both aesthetics and geometry metrics. Supplemental algorithms and parameter tables make the method reproducible once code drops.\n\nSoft spots, in proportion. The multi-view consistency assumption is the load-bearing one and is only partially held: independent FLUX anchors + VACE interpolation still produce shadow flips (Fig. 12) and hit an 8-identity wall. Metrics are computed on Bézier paths derived from the same anchors on 9 custom layouts × 10 scenes; that is enough to show the method works on their distribution, not enough to underwrite a broad SOTA claim for large unstructured outdoor geometry. Free parameters (d0, several δ thresholds, loss weights) are numerous but not hidden. Citation pattern is normal; self-cite of VideoFrom3D is just the generation backbone.\n\nThis is for people building geometry-conditioned outdoor generators or content tools who currently hand-craft camera paths. It is not a foundational theory paper. I would send it to peer review: the scheduling algorithm is concrete, the evidence is honest about its limits, and the engineering is above the usual bar. Engage if you care about outdoor scene generation pipelines; skip if you only track pure generative-model advances.","headline":"Solid systems paper that actually removes the outdoor camera-path bottleneck with a usable visibility graph; SOTA claim is overstated relative to the custom 9-layout eval and residual multi-view inconsistency.","tokens_in":22506,"tokens_out":524,"would_cite":true,"duration_ms":6641,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Outdoor 3D scenes can be generated from coarse or fine geometry alone by automatically building a directed view graph and conditioning each object with an identity image and a geometry-adherence strength.","keywords":["geometry-conditioned 3D generation","outdoor scene synthesis","view scheduling","generation graph","3D Gaussian Splatting","object-level conditioning","video diffusion interpolation"],"falsifier":"On a held-out outdoor mesh with known ground-truth appearance, run the full pipeline and measure whether novel-view renders remain free of floaters, missing regions, and shadow-direction flips while matching the input layout; systematic failure on scenes with more than eight distinct object identities or with strong global illumination variation would refute the claim.","tokens_in":22487,"feed_emoji":"🏞️","tokens_out":668,"duration_ms":6208,"temperature":0.7,"pith_summary":"Most geometry-conditioned 3D scene generators break the problem into three stages—choose cameras, synthesize multi-view images, then reconstruct a 3D representation—but outdoor geometry is large, unstructured, and unbounded, so choosing good cameras is a bottleneck. SceneFrom3D removes that bottleneck by turning the input meshes into a directed generation graph whose nodes are a compact set of anchor views and whose edges are video-interpolation trajectories. The graph tells the system which views to synthesize first, which pairs to interpolate, and in what order generation should proceed. On top of that, every object is given its own identity image (for appearance) and a scalar geometry-adherence parameter (for how tightly the output must follow the input mesh). The result is a freely navigable 3D Gaussian Splatting scene whose layout, object looks, and geometric fidelity are under direct user control, without any hand-specified camera path.","feed_headline":"Outdoor 3D scenes from geometry alone, no camera path needed","feed_subtitle":"A directed view graph plus per-object identity and adherence controls drive free-navigation 3DGS scenes","key_machinery":"The directed generation graph G_gen = (V, E→): nodes V are a compact set of anchor views obtained by visibility-guided densification and pose optimization; directed edges E→ are interpolation trajectories that also encode generation order. This graph, together with object-level conditioning (identity image + α-adherence), organizes both multi-view synthesis and subsequent 3DGS optimization.","core_discovery":"A directed generation graph constructed automatically from outdoor input geometry—nodes as visibility-optimized anchor views, edges as collision-free interpolation trajectories—plus per-object identity images and geometry-adherence parameters, is sufficient to drive a three-stage pipeline that produces high-quality, controllable outdoor 3D Gaussian scenes without requiring explicit camera trajectories as input.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Auto view graph turns outdoor geometry into free-nav 3DGS scenes","Directed anchors and trajectories generate outdoor scenes from meshes","Per-object identity and adherence control outdoor geometry-to-3D","No camera path: scheduled views reconstruct controllable outdoor 3D","Visibility-optimized anchors drive outdoor 3D scene generation"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That a sparse set of automatically chosen anchor views, once densified by video diffusion along graph edges, will yield multi-view observations consistent and complete enough for stable 3D Gaussian reconstruction of large unstructured outdoor scenes.","fun_headline_variants_meta":{"raw":{"variants":["Auto view graph turns outdoor geometry into free-nav 3DGS scenes","Directed anchors and trajectories generate outdoor scenes from meshes","Per-object identity and adherence control outdoor geometry-to-3D","No camera path: scheduled views reconstruct controllable outdoor 3D","Visibility-optimized anchors drive outdoor 3D scene generation"]},"model":"grok-4.5","effort":"low","cost_usd":0.006218,"raw_usage":{"total_tokens":1567,"prompt_tokens":796,"num_sources_used":0,"completion_tokens":90,"cost_in_usd_ticks":62180000,"prompt_tokens_details":{"text_tokens":796,"audio_tokens":0,"image_tokens":0,"cached_tokens":0},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":681,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":796,"tokens_out":90,"duration_ms":6120,"temperature":1.0,"reasoning_tokens":681,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T17:41:36.185555+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a held-out outdoor mesh with known ground-truth appearance, run the full pipeline and measure whether novel-view renders remain free of floaters, missing regions, and shadow-direction flips while matching the input layout; systematic failure on scenes with more than eight distinct object identities or with strong global illumination variation would refute the claim.","supporting_citations":[],"review_version":1}