{"id":"ef14945f-3186-478d-8c38-66c714434ff4","arxiv_id":"2510.26800","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"OmniX trains separate LoRA adapters on FLUX.1-dev so one framework handles panorama generation, intrinsic perception (depth, normals, albedo, roughness, metallic), and completion, then feeds the maps into PBR-ready 3D scenes.","lead":"OmniX repurposes the FLUX.1-dev image model with per-task LoRA adapters to generate, perceive, and complete 360° panoramas, predicting distance, normals, albedo, roughness, and metallic maps from a single image. It also introduces PanoX, a synthetic indoor/outdoor panorama dataset, and shows the outputs being turned into PBR-ready 3D scenes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PBR material generalization is the load-bearing risk: Sec E admits metallic prediction generalizes poorly, and the only quantitative out-domain test is two synthetic UE scenes, so 'graphics-ready' transfer to real scenes is unsupported.","rationale":"The reader's weakest assumption — that PanoX's limited synthetic diversity, especially for metallic/roughness, is the load-bearing risk — is confirmed by the manuscript's own limitations section. Sec E explicitly concedes poor metallic generalization and casts doubt on the usefulness of 2D priors for PBR material estimation. Since the headline 'graphics-ready 3D scenes' depends on these material maps, this is a genuine soft spot, not a manufactured one. I do not find a load-bearing derivation error in the flow-matching objective or adapter architecture; the concern is empirical support for real-world transfer. The reader's CONDITIONAL verdict already captures this adequately, so no change is recommended.","tokens_in":15283,"tokens_out":4963,"duration_ms":50770,"concrete_test":"Obtain or re-train the adapters, then evaluate on a real-world panorama benchmark with ground truth material and geometry, e.g., PhyIR real test scans or a set of captured HDRIs with manually verified SVBRDF. Compute PSNR/LPIPS for metallic/roughness and a Blender relighting MSE against a reference render for a known light source. If real-world metallic PSNR drops substantially below the PanoX-OutDomain value (e.g., more than 30% relative) or relighting error is no better than a zero-metallic baseline, the graphics-ready generalization claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that OmniX produces graphics-ready PBR scenes from arbitrary inputs — requires that the material maps (metallic, roughness, albedo) predicted by LoRA adapters trained on PanoX transfer to real-world panoramas and are mutually consistent enough for physically based rendering. The evidence for this is thinner than the claim. Quantitatively, the only held-out evaluation is PanoX-OutDomain, two synthetic scenes rendered from the same Unreal Engine 5 asset pipeline as the training scenes (Sec 3.2); this tests generalization to new UE scenes, not to real captures. Qualitatively, Fig 9 shows in-the-wild predictions, but the authors themselves state in Sec E that 'OmniX-Pano2Metallic ... performs poorly in generalization' and that 'neural rendering and PBR rendering ... may indicate that pre-trained 2D image priors have limited benefits for PBR material estimation.' Since metallic/roughness are precisely the PBR inputs needed for graphics-ready scenes, a failure here directly voids the headline application for most in-the-wild inputs. Additionally, each modality is trained by an independent adapter with no joint consistency supervision (Tables 6-7 show joint modeling doesn't help), so there is no evidence that separate depth/normal/albedo/roughness/metal predictions are coherent enough for relighting or simulation. The downstream demos are qualitative only.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces OmniX, a framework that adapts the pre-trained 2D flow-matching DiT model FLUX.1-dev to panoramic vision tasks via task-specific LoRA adapters and a Separate-Adapter cross-modal attention design. The proposed formulation covers image-to-panorama generation, RGB-to-X perception (distance, normal, albedo, roughness, metallic), and masked completion. The authors also construct PanoX, a synthetic Unreal Engine 5 dataset of 10k instances from 8 indoor/outdoor scenes with dense geometry and material annotations. Experiments on the held-out PanoX-OutDomain show large improvements over several baselines on albedo/roughness/metallic estimation and competitive depth/normal accuracy; qualitative results demonstrate panorama generation, completion, and downstream mesh-based relighting and physical simulation. The paper claims that this pipeline yields graphics-ready 3D scenes for PBR rendering, relighting, and simulation.","tokens_in":15569,"tokens_out":6814,"duration_ms":96555,"significance":"The perception results are promising: Table 2 shows consistent and large PSNR/LPIPS gains over five strong baselines on the held-out synthetic domain, and the ablations in Table 4 and Appendix D give useful evidence for the Separate-Adapter design and for independent rather than joint prediction. PanoX is a potentially valuable resource, being one of the few panoramic datasets with dense material annotations, and the paper is transparent about its limitations in Sec. E. However, the system-level headline claim—that OmniX enables graphics-ready 3D scenes for arbitrary input panoramas—is not supported by the current evidence: the quantitative evaluation is restricted to two synthetic scenes generated by the same UE5 pipeline as the training data, and the authors themselves concede that metallic prediction generalizes poorly. The strength of the paper is therefore in the method and dataset for panoramic perception, not yet in the demonstrated generality of the PBR/scene-construction application.","major_comments":[{"comment":"The only quantitative out-of-distribution test is PanoX-OutDomain, which consists of two scenes rendered from the same Unreal Engine 5 asset pipeline used to build the training scenes. This tests generalization to novel synthetic environments, not to real panoramic captures. Since the abstract and Sec. 3.4 claim graphics-ready 3D scenes for PBR rendering and relighting, the PBR material maps are load-bearing, yet Sec. E states that OmniX-Pano2Metallic 'performs poorly in generalization' and that 2D priors may have 'limited benefits for PBR material estimation.' Figure 9 provides only qualitative evidence on real images. Please add quantitative evaluation on real panoramic inputs (or a clearly justified proxy), or explicitly restrict the graphics-ready claim to synthetic and same-pipeline inputs.","section":"§3.2, Sec. E, Tables 2–3"},{"comment":"No error bars, confidence intervals, or significance tests are reported; all tables appear to be point estimates from a single training run. This matters especially because PanoX-OutDomain contains only two scenes, so the variance of the reported metric could be large relative to the differences between methods. For example, Table 3 shows OmniX as second-best on distance and the normal columns are not available for two baselines, making the comparison fragile. Please report standard deviations over multiple runs (or bootstrap over the test set) and state the number of seeds for the main comparisons.","section":"Tables 2, 3, 4, 5, 6, 7"},{"comment":"Each modality is trained with an independent adapter, and the ablations show that joint modeling of geometry or materials does not improve accuracy. While this is a defensible design choice, it leaves open the question of cross-modal consistency: nothing ensures that the separately predicted distance, normal, albedo, roughness, and metallic maps correspond to the same underlying scene. For the 'graphics-ready' claim, this consistency is essential, because inconsistent maps will produce incorrect relighting and simulation. The only downstream validation (Sec. 4.5, Fig. 8) is qualitative. Please add a quantitative evaluation of the assembled 3D scene—for example, re-rendering error against PanoX ground truth, or multi-view mesh consistency—or explicitly soften the claim to 'visually plausible' rather than physically consistent.","section":"Sec. 4.1, Tables 6–7, Sec. 4.5"},{"comment":"The qualitative image-to-panorama results in Fig. 10 are obtained from input images that, per the caption, are themselves generated by FLUX.1-dev. This does not demonstrate generalization to real single-view photographs, which is the practical setting for the claimed application. Please provide examples with real images, or state clearly that the generation results are limited to synthetic/generated inputs.","section":"Appendix B, Fig. 10"}],"minor_comments":[{"comment":"Typos: 'two prevalent ways to constructing' should be 'two prevalent ways of constructing'; 'we achieve the highest normal estimation accuracy and and the second highest' contains a doubled 'and'.","section":"Abstract and Sec. 4.3"},{"comment":"PanoX is presented as a contribution and benchmark, but no dataset/code release link or license is provided. Please add an availability statement.","section":"General"},{"comment":"The column header suggests PSNR and LPIPS are reported for all modalities, but the roughness and metallic columns contain only one metric. Align the header with the reported entries or add the missing values.","section":"Table 5"},{"comment":"Raistrick et al. 2023a and 2023b are identical; Yu et al. 2024c and 2024d are also identical. Cite each work once to avoid confusion.","section":"References"},{"comment":"The demonstration videos are said to be on the project page, but no URL is given in the manuscript. Include the project page link in the final version.","section":"Sec. 4.5"}],"recommendation":"major_revision","confidential_remarks":"The central perception method and dataset are solid, and the tables show substantial gains in the evaluated setting. The main risk is overclaiming external validity for the 'graphics-ready' application; the paper's own Sec. E acknowledges poor metallic generalization. I would support acceptance after either adding real-world quantitative evidence or carefully scoping the claim. The self-citations do not appear to create circularity, and I see no reason to doubt the integrity of the experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid systems paper whose headline claim is stronger than its evaluation. The genuinely new pieces are PanoX — the first panoramic dataset I know covering both indoor and outdoor scenes with dense geometry and PBR material annotations — and the Separate-Adapter LoRA design for FLUX.1-dev, which is a concrete, sensible way to reuse a 2D flow-matching prior for multiple spatially aligned tasks. The unified formulation in Eq. 3–5 is not deep, but it is clean and covers perception, generation, and completion without moving the pretrained weights.\n\nOn evidence: the quantitative gains on PanoX-OutDomain are consistent, and the ablations (Shared-Branch vs Shared-Adapter vs Separate-Adapter, camera-ray, joint vs independent) are informative. I do not see a load-bearing math error. The flow-matching objective with extra conditions is standard, and nothing is circular.\n\nSoft spots. The out-domain test is two synthetic Unreal Engine scenes rendered from the same asset pipeline as the training scenes, so it tests generalization to new UE layouts, not to real captures. No error bars or significance tests are reported, so the margin over DiffusionRenderer on metallic (18.9 vs 14.5 PSNR) is hard to interpret. The in-the-wild results are qualitative only. More important, the authors state in Sec. E that OmniX-Pano2Metallic \"performs poorly in generalization\" and that 2D priors may have limited benefits for PBR material estimation. That is exactly the load-bearing component for the \"graphics-ready 3D scenes\" claim: metallic and roughness are what make a mesh relightable in a PBR pipeline. The joint-modeling ablations also show independent adapters perform as well, so there is no demonstrated consistency between the depth, normal, albedo, and material predictions. The relighting and simulation demos are qualitative. None of this kills the core contribution, but it means the paper overstates what is actually supported.\n\nPractical concern: no code, data, or checkpoints are released, which makes the synthetic benchmark claims hard to verify. The self-citations to DreamCube and LayerPano3D are for standard components, so no circularity issue.\n\nWho this is for: panorama/3D scene generation and inverse rendering researchers. The dataset and adapter recipe are worth knowing even if the PBR transfer claim needs more work. It deserves a serious referee. I would send it out, with instructions to push on real-world PBR generalization and artifact release.","headline":"A useful unified-adapter recipe and a new synthetic panorama dataset, wrapped in a 'graphics-ready' claim that outruns the evidence; worth refereeing, but the authors should release artifacts and test on real scenes.","tokens_in":16099,"tokens_out":2526,"would_cite":false,"duration_ms":27690,"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":"OmniX claims that one frozen 2D flow-matching model, fitted with lightweight LoRA adapters, can generate and perceive panoramas and assemble them into relightable, graphics-ready 3D scenes.","keywords":["panorama perception","inverse rendering","PBR material estimation","3D scene generation","flow matching","LoRA adapters","panoramic dataset","graphics-ready scenes"],"falsifier":"Run the OmniX metallic and roughness adapters on a diverse set of real indoor and outdoor panoramas with known PBR ground truth (or close-up material checks) and compare against the synthetic-only PanoX results: if metallic maps degrade sharply on non-synthetic scenes, as the paper's own limitation section suggests, the 'graphics-ready for arbitrary scenes' claim collapses to the synthetic domain.","tokens_in":15116,"feed_emoji":"🌍","tokens_out":3199,"duration_ms":29277,"temperature":0.7,"pith_summary":"OmniX claims that a pre-trained 2D flow-matching image generator can be converted, without retraining its core weights, into a unified engine for panoramic image generation, intrinsic perception (distance, normals, albedo, roughness, metallic), and masked completion. The key mechanism is a Separate-Adapter cross-modal structure in which each input and output modality gets its own lightweight LoRA adapter while sharing the frozen diffusion transformer. If correct, a single-model chain from one perspective image to full panorama, to multimodal property maps, to a textured, relightable 3D mesh becomes practical, closing the gap between 2D lifting and graphics pipelines. The authors train twelve adapters on a new synthetic panorama dataset (PanoX) and report consistent gains over prior intrinsic-decomposition and geometry estimators on out-of-domain panoramas, then demonstrate PBR relighting and physics simulation in Blender.","feed_headline":"One model turns panoramas into relightable 3D scenes","feed_subtitle":"LoRA adapters on a frozen 2D generator predict depth, normals, and PBR maps that assemble into graphics-ready meshes.","key_machinery":"Separate-Adapter cross-modal structure built on a pre-trained DiT-based flow matching model: each spatially aligned condition (RGB panorama, camera ray, mask) and each target modality gets its own LoRA adapter, and the adapters interact through cross-modal attention inside the MMDiT blocks. This lets one frozen backbone handle image-to-panorama generation, RGB-to-X perception for five modalities, and completion/guided perception, with a unified multiple-input/multiple-output flow-matching loss. Horizontal blending is used to enforce equirectangular seam continuity.","core_discovery":"At its core, the paper tries to establish that perception of scene intrinsics—geometry and PBR materials—can be formulated as just another conditioning/generation task for a flow-matching model. Treating RGB-to-X as generation of X from an RGB condition (and optionally camera rays), and masked completion as generation from masked conditions, the same pre-trained 2D generative prior can be reused across all tasks. The load-bearing design choice is Separate-Adapter: distinct LoRA adapters for each condition and target, joined through cross-modal attention, which preserves the original weight distribution and generalizes better than channel-wise or token-wise sharing. The paper's evidence is qu","pith_inferences":["The Separate-Adapter design suggests a scaling route: adding a new modality (e.g., semantics, lighting, or HDRI) may require training only one more LoRA rather than re-tuning the generator—an extensibility the paper does not explicitly test.","A direct test is to train or fine-tune the metallic adapter on procedurally varied synthetic materials or real panoramic PBR captures; the paper's limitation section implies the bottleneck is data scarcity, not architecture.","Because the method avoids panorama-specific inductive biases apart from horizontal blending, the same adapter scheme may transfer to narrow field-of-view images; the paper notes this possibility but does not demonstrate it.","The bumpy surfaces caused by imperfect distance predictions could likely be smoothed by post-hoc mesh refinement or a depth-refinement network before assigning PBR materials—an integration the paper leaves implicit."],"forward_implications":["A single frozen 2D generator plus lightweight adapters can replace separate depth, normal, albedo, roughness, and metallic estimators for panoramic inputs.","Image-to-panorama generation can be extended to output geometry and PBR material maps in the same framework, enabling textured 3D meshes without manual UV work.","Mask-guided completion allows progressive, interactive scene extension from a single viewpoint, moving toward explorable and even city-scale 3D worlds.","The PanoX dataset provides a new benchmark for panoramic intrinsic decomposition covering both indoor and outdoor scenes.","If the pipeline generalizes, artists and robotics simulators could go from one photo to a relightable, physics-ready scene asset automatically."],"fun_headline_variants":["Panorama to relightable 3D scenes with LoRA adapters","Unified framework for panorama generation and perception","2D flow priors repurposed for graphics-ready 3D worlds","One model predicts depth, normals, PBR from panoramas","Separate LoRA adapters unify panorama tasks into 3D"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central claim depends on training on synthetic panoramas rendered from only eight Unreal Engine scenes transferring to arbitrary real-world scenes—especially for metallic and roughness maps, where the authors concede the metallic adapter 'performs poorly in generalization.'","fun_headline_variants_meta":{"raw":{"variants":["Panorama to relightable 3D scenes with LoRA adapters","Unified framework for panorama generation and perception","2D flow priors repurposed for graphics-ready 3D worlds","One model predicts depth, normals, PBR from panoramas","Separate LoRA adapters unify panorama tasks into 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1204,"prompt_tokens":771,"completion_tokens":433,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":355}},"tokens_in":515,"tokens_out":433,"duration_ms":4930,"temperature":1.0,"reasoning_tokens":355,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T07:03:32.190525+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the OmniX metallic and roughness adapters on a diverse set of real indoor and outdoor panoramas with known PBR ground truth (or close-up material checks) and compare against the synthetic-only PanoX results: if metallic maps degrade sharply on non-synthetic scenes, as the paper's own limitation section suggests, the 'graphics-ready for arbitrary scenes' claim collapses to the synthetic domain.","supporting_citations":[],"review_version":1}