{"id":"5b2a32fd-6979-4020-b799-07aa6a605f35","arxiv_id":"2608.07012","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A vision-language framework predicts editable JSON-like 3D scene programs from a few uncalibrated RGB views, and a new 110K-scene dataset supports training.","lead":"Scenix turns a few uncalibrated photos of an indoor room into a structured, editable 3D scene program, with each object identified, described, and placed in metric 3D space. It pairs a newly built dataset of over 110,000 indoor scenes with a vision-language pipeline that could make 3D scene editing from ordinary photos practical.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The metric-scale assumption is unverified: with uncalibrated RGB, absolute scale is not identifiable from image geometry alone, and the paper provides no controlled experiment showing the learned prior transfers; a scale-isolation test is needed.","rationale":"The reader's weakest assumption identifies exactly the load-bearing condition: metric scale and absolute placement must be recoverable from sparse uncalibrated RGB by learned priors. I agree this is the most central issue. If absolute scale is wrong, every downstream metric in the paper (F1@5%, F1@10%, 3D IoU) is corrupted, and the 'executable' program instantiates a dollhouse-sized or aircraft-hangar-sized scene. The paper's own OOD results are consistent with a scale-prior failure: F1@5% drops from 0.331 on in-domain InfiniGen to 0.239 on out-of-domain SpatialGen for the 9B model, but this drop is confounded by category and inventory errors, so it does not isolate scale. The proposed scale-rendering test would isolate the issue cleanly, because pixel-identical images force the model to reveal whether its metric predictions come from the image or from a dataset-specific prior. \n\nI considered other concerns, such as the lack of code/data release and the fairness of the single-view baseline comparison in Table 5. Those are real limitations, but they do not attack the validity of the central claim as directly as the scale ambiguity. The baseline comparison, while imperfect, is a controlled shared-front-end setup; the missing artifacts affect reproducibility, not correctness. The scale concern, by contrast, threatens the core meaning of 'metric spatial annotations' and 'executable scene program.' \n\nThe verdict should remain CONDITIONAL/UNCHANGED: the paper is a substantial empirical contribution, but the central metric claim is conditioned on a scale prior that has not been demonstrated. The reader already assigned CONDITIONAL; my analysis reinforces that condition rather than changing it to ACCEPT or REJECT.","tokens_in":13559,"tokens_out":6772,"duration_ms":83848,"concrete_test":"Select a set of held-out test scenes (e.g., 20 InfiniGen cases and 20 SUN RGB-D cases). For each scene, render it at three global metric scales: 1.0, 1.5, and 0.67, by scaling all room dimensions, object centers and extents, camera positions, and the focal length by the same factor, producing pixel-identical images. Run Scenix on these identical images. Because the inputs are identical, the predicted program will be identical (modulo decoding stochasticity, which can be controlled by fixing the seed or running multiple samples). Compare the predicted absolute centers and extents against the correspondingly scaled ground truth. If the localization F1@5%, F1@10%, and 3D IoU remain roughly constant across the three scales, the model has a scale-invariant prior and the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Scenix recovers an executable scene program with metric room dimensions and object centers/extents from sparse uncalibrated RGB images, without camera, depth, or scale estimation. This requires that absolute metric scale be supplied by the learned prior of LayoutVLM. The paper never isolates this assumption. Equation (2) trains LayoutVLM to regress absolute coordinates directly, and Equation (3) supervises those coordinates; the only mechanism for scale is whatever the model absorbs from the training distribution, which is dominated by InfiniGen synthetic scenes (91% of training cases). No experiment varies scene scale while holding appearance constant, and the OOD evaluation on SpatialGen is still synthetic and uses only 48 scenes. \n\nThe scale ambiguity is not merely theoretical. Given any image, scaling the entire scene (room boundaries, object centers and extents, camera positions) and the focal length by a common factor produces pixel-identical input while changing every ground-truth metric annotation by that factor. A deterministic model must emit the same program for both, so it cannot be correct for both; it can only succeed if its prior over room and object sizes happens to match the true scale. The paper provides no evidence this prior transfers to real-world rooms, which vary systematically in size. The in-domain SUN RGB-D evaluation is drawn from the same real-image source used for training, so it cannot test cross-dataset scale transfer. Thus the metric component of the central claim is currently unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Scenix, a system that reconstructs editable 3D indoor scenes from one to ten uncalibrated RGB images by predicting an 'executable scene program'—a JSON representation containing room boundary, wall height, and per-object category, description, 3D center, extent, rotation, and support—and then instantiating it via an asset grounder and a Critic-Editor-Verify loop. The authors introduce XScene, a dataset of roughly 110K synthetic and real scenes with multiview imagery and metric program annotations, and an observation-consistent supervision protocol. Experiments evaluate raw program prediction on held-out InfiniGen and SUN RGB-D splits, zero-shot transfer to 48 OOD SpatialGen scenes, ablation of the refinement stages, and comparison with three single-view baselines on a shared detection front end. The central claim is that direct autoregressive scene-program generation gives more accurate scene layouts than existing single-view reconstruction baselines while preserving object-level editability.","tokens_in":13867,"tokens_out":8477,"duration_ms":82898,"significance":"The executable-scene-program representation is a genuinely useful idea: it makes the reconstruction object-addressable and editable while decoupling structure from asset realization. The XScene dataset is a substantial resource for training and evaluating such models, and the observation-consistent supervision protocol is a sensible way to avoid supervising hidden objects. The paper's claims are evaluated against held-out scenes rather than fitted quantities, so I see no circularity. The stage-wise ablation on SpatialGen clearly attributes most of the OOD gain to the Asset Grounder, which is informative. That said, the headline claim of metric layout recovery from uncalibrated RGB rests on a learned scale prior that is never isolated, and the external comparison is too narrow to support the broad 'more accurate than existing approaches' statement. These are fixable with additional experiments.","major_comments":[{"comment":"The central claim that Scenix recovers metric room dimensions and object centers from uncalibrated RGB is not supported by a scale-isolation experiment. A global scaling of scene geometry and camera focal length produces pixel-identical input, so a deterministic model must rely on a learned prior to output the correct metric program. Eq. (2) and the supervision in Eq. (3) train on absolute annotations, and roughly 91% of training cases are InfiniGen synthetic; the paper provides no controlled experiment that varies scene scale while holding appearance constant. In addition, the localization F1 metric in §5.2 normalizes center distance by room diagonal; please state whether the denominator uses the predicted or ground-truth room diagonal, because this determines whether global scale errors are penalized at all. A concrete test would render XScene or SpatialGen scenes at, e.g., 0.8× and 1.2× scale with adjusted camera intrinsics and report absolute room-dimension error, 3D IoU, and F1@5% under both normalization choices.","section":"§3.2, Eqs. (2)–(3); §5.2"},{"comment":"The contribution claim of 'more accurate scene layout reconstruction than existing approaches' is broader than the evidence. Table 3 reports only the 4B and 9B variants on the in-domain splits; no prior method is evaluated on the XScene test sets or on multi-view inputs. Table 5 is the only external comparison, and it is restricted to a single-view subset of the synthetic OOD SpatialGen data with a shared detection front end. Please add at least one representative prior method on the in-domain splits, or qualify the headline claims to the single-view OOD setting.","section":"§5.3, Table 3; §5.5, Table 5"},{"comment":"Table 5 appears incomplete: the Gen3DSR and 3D-Fixer rows are missing several metric cells (e.g., Gen3DSR has no reported 3D IoU, and 3D-Fixer lacks most class-level metrics), and no standard deviations or per-scene counts are reported for the baseline columns. Without these values and a statement of how many scenes were used, the claimed margins over SAM3D, Gen3DSR, and 3D-Fixer cannot be fully verified.","section":"Table 5"},{"comment":"The SUN RGB-D test split is drawn from the same real-image source used for training, so it does not test cross-dataset transfer of the metric-scale prior to real-world rooms. The only OOD evaluation is on 48 synthetic SpatialGen scenes. As a consequence, the paper does not yet demonstrate that the learned size prior transfers across real-world distributions with systematically different room and object dimensions. A small real-world cross-dataset evaluation, or an explicit limitation statement, is needed to scope the claims.","section":"§4.1; §5.3"}],"minor_comments":[{"comment":"The class-pattern filter is described only by examples (lights, tableware, books, wall ornaments, hardware); please provide the exact category list and a sensitivity analysis, since all reported metrics depend on this filter.","section":"§5.2"},{"comment":"The visibility threshold κ is fixed; an ablation over κ would clarify how observation-consistent supervision affects downstream results.","section":"§3.2"},{"comment":"The coordinate transformation for SUN RGB-D boxes is described only briefly; specify the world-frame convention and how the camera-to-world transform is obtained for single images.","section":"§4.1"},{"comment":"The text mentions 'Hunyuan-turbo' but does not cite it in the reference list; please add the appropriate citation.","section":"§4.1"},{"comment":"Figure 4(a) shows qualitative successes but the mentioned ambiguities (reflection, rotation, diagonal symmetry, category errors) are not quantified; consider reporting per-case metrics for the displayed examples.","section":"Figure 4"},{"comment":"The paper does not report inference time or compute for LayoutVLM and the Critic-Editor-Verify loop; given the agentic nature of the loop, this information would help readers judge practicality.","section":"§5.4"}],"recommendation":"major_revision","confidential_remarks":"The metric-scale issue is the main risk and should be addressed with a scale-isolation experiment; the evaluation gaps are also important. The XScene dataset and the executable-program formulation are genuine contributions, and the stage-wise ablation is well designed. If the authors can supply the missing controls and broaden the comparison, the paper would be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one. Scenix is a genuine new formulation: instead of lifting sparse RGB views into geometry, it predicts an executable scene program—blueprint plus per-object category, description, center, extent, rotation, support—and then instantiates that into an editable scene. The XScene dataset is a real contribution: 110K scenes with multiview images and metric annotations, observation-consistent supervision so hidden objects aren't penalized, and a split that separates synthetic and real data. The pipeline is well thought out; the Asset Grounder and the bounded Critic–Editor–Verify loop are clearly motivated, and the staged ablation on SpatialGen shows where the gains come from. The single-view comparison with SAM3D, Gen3DSR, and 3D-Fixer uses a shared detection front-end, which is fair and informative. Related-work coverage is honest and current.\n\nNow the soft spots. The big one is metric scale. With uncalibrated RGB, absolute scale is not identifiable from image geometry; the model has to learn it from the training distribution, which is 91% InfiniGen synthetic. The paper never isolates this. A deterministic model that sees pixel-identical inputs from scenes scaled by different factors, with the focal length adjusted to match, must emit the same program, so it can only be correct for one scale unless its prior is genuinely calibrated. There is no controlled experiment varying scene scale while holding appearance constant, and the OOD test is 48 synthetic SpatialGen scenes. That is the load-bearing unresolved question. I don't think it kills the paper—learned priors can supply scale in practice—but the claim of an executable metric scene program requires that evidence.\n\nLesser issues: the in-domain XScene table compares only model sizes, not against prior methods. Tables 4 and 5 are small and omit error bars. No code or data release, so the dataset contribution isn't yet independently usable. The in-domain standard deviations look implausibly tiny (e.g., .003), but that is probably an artifact of macro averaging over many scenes; I would not call it cheating.\n\nBottom line: this deserves a serious referee. The representation and the dataset will be useful for people working on indoor scene understanding, AR/VR, and embodied AI regardless of whether the scale prior transfers to general real rooms. Send it to review, and ask for a scale-isolation test, at least one external baseline on the in-domain split, and a code/data release commitment.","headline":"A genuinely new task and dataset for sparse-view editable 3D scenes, but the load-bearing metric-scale assumption is unverified; send to review with a demand for a scale-isolation experiment.","tokens_in":14427,"tokens_out":2232,"would_cite":true,"duration_ms":24357,"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":"A fine-tuned vision-language model can turn one to ten uncalibrated photos into an executable and editable 3D scene program.","keywords":["sparse-view 3D reconstruction","executable scene program","scene layout prediction","vision-language model","indoor scene dataset","object grounding","scene editing","uncalibrated RGB views"],"falsifier":"Render the same known room from different camera heights, focal lengths, and capture distances, feed each rendering set through Scenix, and measure the predicted wall height and object extents: if the absolute metric outputs shift systematically with camera intrinsics or capture distance, the learned metric prior is not actually recovering scale from images. A second check is to compare predicted centers and extents against ground truth in meters without any best-fit global alignment; if errors concentrate in a single scale factor, the evaluation's alignment step is hiding the failure.","tokens_in":13386,"feed_emoji":"🛋️","tokens_out":9733,"duration_ms":83258,"temperature":0.7,"pith_summary":"The paper tries to establish that sparse-view 3D indoor scene reconstruction can be solved by treating it as direct prediction of an executable scene program, a structured text description of the room boundary and each object's category, appearance, position, extent, rotation, and support, rather than as staged detection, depth lifting, and association. It claims that a fine-tuned vision-language model, given one to ten uncalibrated RGB images, can autoregressively output such a program in one pass and that this representation reconstructs layouts more accurately than single-view baselines while remaining editable at the object level. To supply the missing supervision, the paper constructs XScene, a dataset of about 110,000 synthetic and real indoor scenes with observation-consistent targets, so a model learns to reconstruct only entities visible in the selected views. A reader should care because success would mean a casual photo burst of a room, with no camera poses, depth, masks, or floor plans, could become a structured and editable 3D environment.","feed_headline":"A scene program turns a few photos into editable 3D rooms","feed_subtitle":"Sparse uncalibrated RGB images are enough to predict room geometry, object poses, and descriptions directly.","key_machinery":"The carrying object is the executable scene program, a constrained JSON schema that separates a blueprint (room boundary and wall height) from object records that couple semantic descriptions with metric transformations. The pipeline has three mechanisms: LayoutVLM, a fine-tuned vision-language model that autoregressively predicts the program from a variable number of views in a single pass; the Asset Grounder, which builds an evidence-backed multiview inventory, reconciles it with the predicted seeds through retention, replacement, suppression, and cloning decisions, and uses selected visual evidence to condition open-vocabulary asset generation; and a bounded Residual Critic-Editor-Verify loop that renders a metric bird's-eye proxy, converts discrepancies into object-specific testable clauses, proposes minimal translation, rotation, or resize edits, and verifies each edit on a fresh rendering until all clauses pass or the iteration budget ends. Together these separate structural induction from asset realization, so where objects are and what they look like can be corrected independently.","core_discovery":"The central claim is that reconstruction from sparse uncalibrated RGB views reduces to conditional generation of a scene program: LayoutVLM models a distribution over serialized JSON programs given the input views and autoregressively emits a blueprint (polygonal boundary and wall height) plus a list of individually addressable object records (category, long-form description, 3D center, extent, rotation, and support relation). The program is executable because it can be rendered as a canonical coarse 3D layout, populated with generated assets, verified through canonical projections, and edited object-by-object. The paper reports that on single-view out-of-distribution SpatialGen scenes this direct program prediction reaches F1@5% of 0.355 for the 9B model and 0.287 for the 4B model, against 0.192 to 0.207 for detection-based reconstruction baselines sharing the same 2D front end, and that it recovers objects that are fully occluded and therefore undetectable. On multiview out-of-distribution scenes, the full pipeline raises F1@5% over raw predictions by 34.0% (9B) and 48.3% (4B), with the evidence-based Asset Grounder reconciliation contributing most of the gain.","pith_inferences":["The same scene-program interface could support incremental editing: a user or language model could modify one object record and re-run only the verification loop instead of regenerating the whole scene, which the paper does not evaluate.","A controlled experiment that re-renders one scene under different camera heights, focal lengths, and capture distances would isolate whether the learned metric prior transfers across camera intrinsics; the paper reports no such experiment, so scale robustness remains the open risk.","Because evaluation removes global translation and rotation before computing distances, systematic scale errors could be masked; reporting absolute metric errors in meters would sharpen comparison against pointmap- and depth-based reconstruction.","If the representation generalizes, it could act as a compact bridge between vision-language models and 3D simulators, since the program is a text description a simulator can instantiate directly."],"forward_implications":["Reconstruction no longer needs external geometric priors: no camera calibration, depth, point clouds, floor plans, or annotated masks are required at inference time.","Direct program prediction avoids the compounding error chain of segmentation-then-lifting and can recover entities that are fully occluded or absent from 2D detections.","The executable program gives a natural editing interface: changing a JSON field moves, resizes, rotates, or replaces an object while the scene remains a valid structured layout.","Model scale in the LayoutVLM primarily improves spatial localization rather than class-level inventory, suggesting layout accuracy is the capacity-limited part of the task.","Observation-consistent supervision forces the model to reconstruct only what is visible, keeping reconstruction separate from unconstrained scene hallucination."],"supporting_citations":[{"why":"InfiniGen's procedural scene generators provide the adjustable synthetic multiview supervision and readable object parameters used to build the majority of XScene training cases.","marker":"(Raistrick et al. 2023)"},{"why":"SUN RGB-D supplies real single-view images with 3D annotations that, after geometric correction and generated descriptions, form the real-image portion of XScene.","marker":"(Song, Lichtenberg, and Xiao 2015)"},{"why":"SpatialGen scenes provide the out-of-distribution testbed with continuous views from which three to ten are subsampled for zero-shot transfer and refinement evaluation.","marker":"(Fang et al. 2026)"},{"why":"Gen3DSR serves as a single-view reconstruction baseline and provides the shared 2D front end so that all compared methods receive identical masks and labels.","marker":"(Ardelean, Özer, and Egger 2025)"},{"why":"SAM3D is the detection-and-completion baseline whose inability to recover fully occluded objects motivates direct scene-program prediction.","marker":"(Chen et al. 2026)"},{"why":"3D-Fixer is the coarse-to-fine in-place completion baseline used in the single-view comparison.","marker":"(Yin et al. 2026)"}],"fun_headline_variants":["Scene programs make sparse photos editable 3D rooms","Editable 3D scenes from a few uncalibrated photos","Few photos to structured 3D scenes via executable programs","Sparse-view reconstruction defaults to editable scene programs","Programs from photos: editable 3D rooms with sparse views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that metric 3D scale and absolute placement can be recovered from a handful of uncalibrated RGB images by learned priors alone, with no camera, depth, or scale estimation anywhere in the pipeline.","fun_headline_variants_meta":{"raw":{"variants":["Scene programs make sparse photos editable 3D rooms","Editable 3D scenes from a few uncalibrated photos","Few photos to structured 3D scenes via executable programs","Sparse-view reconstruction defaults to editable scene programs","Programs from photos: editable 3D rooms with sparse views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00068,"raw_usage":{"total_tokens":3156,"prompt_tokens":1079,"completion_tokens":2077,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":695,"completion_tokens_details":{"reasoning_tokens":1994}},"tokens_in":695,"tokens_out":2077,"duration_ms":15012,"temperature":1.0,"reasoning_tokens":1994,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:37:11.383036+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render the same known room from different camera heights, focal lengths, and capture distances, feed each rendering set through Scenix, and measure the predicted wall height and object extents: if the absolute metric outputs shift systematically with camera intrinsics or capture distance, the learned metric prior is not actually recovering scale from images. A second check is to compare predicted centers and extents against ground truth in meters without any best-fit global alignment; if errors concentrate in a single scale factor, the evaluation's alignment step is hiding the failure.","supporting_citations":[{"cited_title":"2026 International Conference on 3D Vision (3DV) , pages =","cited_arxiv_id":null,"evidence_quote":"SpatialGen scenes provide the out-of-distribution testbed with continuous views from which three to ten are subsampled for zero-shot transfer and refinement evaluation."}],"review_version":1}