{"id":"f36453c8-f3ff-4079-8ea2-b32be2f813e9","arxiv_id":"2510.10726","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A unified feed-forward network that ingests optional geometric priors and jointly predicts point maps, depth, camera poses, normals, and 3D Gaussians, reporting state-of-the-art results on multiple 3D benchmarks.","lead":"WorldMirror is a single feed-forward model that takes images plus any available hints about camera poses, intrinsics, or depth, and outputs point clouds, depths, camera parameters, normals, and 3D Gaussian scenes. If its results hold up, one model could replace several specialized 3D vision systems and make 3D reconstruction more practical for AR and robotics.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"NVS SOTA claim is not pose-free: Table 4 uses GT cameras for back-projection; re-report with predicted cameras.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing concern: the headline NVS results in Table 4 are not produced by a fully pose-free system, despite the paper's explicit claim to enable pose-free novel view synthesis. This directly affects a central component of the abstract's SOTA claim. The paper's own ablation (Table 6) partially addresses the issue by reporting a 'w/o GT Cameras' variant, which is a credit to the authors, and on RealEstate10K the difference is negligible, suggesting the model may be nearly pose-free on that dataset. However, the ablation is only against itself, not against the baselines, and the VR-NeRF drop of ~1 dB shows that GT-camera usage can matter substantially. Since AnySplat and FLARE operate without GT cameras, the Table 4 comparison is not apples-to-apples. The proposed concrete test—recomputing Table 4 with predicted cameras—would settle whether WorldMirror retains its SOTA margin over pose-free baselines. Secondary concerns, such as the abstract overstating depth performance relative to Table 7, are real but less load-bearing than the NVS protocol issue because the depth SOTA claim is not the central novel contribution. Overall, the concern matches the reader's weakest assumption, and the verdict should remain CONDITIONAL pending the regenerated numbers.","tokens_in":18188,"tokens_out":5932,"duration_ms":51498,"concrete_test":"Recompute Table 4 by using predicted camera poses and intrinsics (the 'w/o GT Cameras' setting of Table 6) for Gaussian center back-projection, on the same RealEstate10K, DL3DV, and VR-NeRF test splits used in Table 4, and compare PSNR/SSIM/LPIPS against AnySplat and FLARE under the same protocol. If WorldMirror's advantage over AnySplat shrinks to within one standard error or inverts on any dataset, the pose-free NVS SOTA claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"WorldMirror claims pose-free novel view synthesis that outperforms AnySplat (Sec. 4.1). But the headline NVS results in Table 4 are produced by back-projecting the predicted Gaussian depth maps using ground-truth camera poses [R|t] and intrinsics K (Eq. 4). This is a protocol mismatch: AnySplat is uncalibrated, so the comparison gives WorldMirror access to ground-truth camera geometry at test time. The Table 6 ablation 'w/o GT Cameras' replaces GT with predicted cameras and shows that on RealEstate10K the PSNR is essentially unchanged (20.30 vs 20.29), but on DL3DV it drops ~0.2 dB (20.69 vs 20.91) and on VR-NeRF it drops ~1 dB (24.76 vs 25.75). However, that ablation is only reported for WorldMirror against itself; it does not compare against AnySplat or FLARE under the predicted-camera protocol. If the central claim is universal pose-free NVS, the SOTA claim is only supported by a system that is not fully pose-free in the headline evaluation. The fix is straightforward: re-run the Table 4 benchmark using the camera head's predicted poses/intrinsics for Gaussian placement, and report the comparison to baselines.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"WorldMirror presents a single feed-forward transformer that consumes N multi-view images together with any subset of geometric priors (camera poses, intrinsics, per-view depth) and predicts point maps, multi-view depth, camera parameters, surface normals, and 3D Gaussians. Priors are encoded as tokens (compact tokens for camera/intrinsics, dense additive tokens for depth), randomly dropped during training, and a curriculum schedules tasks, data, and resolution. Experiments cover point maps (7-Scenes/NRGBD/DTU), cameras (RealEstate10K/Sintel/TUM), normals (ScanNet/NYU/iBims), depth (NYU/Sintel/KITTI), and NVS (RealEstate10K/DL3DV/VR-NeRF), with claims of state-of-the-art results over VGGT, π3, AnySplat, FLARE, and others, plus universal gains from any prior injection.","tokens_in":18475,"tokens_out":7384,"duration_ms":63681,"significance":"If the results hold, WorldMirror is a useful step toward a generalist 3D geometric foundation model: one architecture that handles camera, depth, normal, point-map, and NVS tasks, with a clean token-based mechanism for incorporating optional priors. The breadth of the evaluation is a genuine strength, as is the ablation showing that a single-token prior embedding outperforms dense embeddings. The paper also demonstrates generalization to AI-generated video inputs. The main concerns are evaluative: the headline NVS numbers use ground-truth cameras for Gaussian back-projection, the DL3DV zero-shot status needs clarification, and the 'any-prior' claim lacks a depth-prior NVS ablation. These are fixable, and the Table 6 pose-free ablation suggests the central NVS conclusion may survive a corrected protocol, but the paper as written overstates the pose-free SOTA claim.","major_comments":[{"comment":"The headline NVS comparison is not generated by a fully pose-free system: Eq. (4) back-projects predicted depth maps using ground-truth camera poses and intrinsics. The Table 6 'w/o GT Cameras' ablation substitutes predicted cameras and reports small drops on RealEstate10K (20.30 vs 20.29) but larger drops on DL3DV (20.69 vs 20.91) and VR-NeRF (24.76 vs 25.75), and is not compared against AnySplat/FLARE under the same protocol. Please report Table 4 with predicted cameras and include baselines under that protocol, or explicitly scope the claims to 'given ground-truth cameras'.","section":"Sec. 3.2, Eq. (4); Table 4"},{"comment":"DL3DV is listed in the training data (A.2) and also used for 'zero-shot' NVS evaluation in Table 4. The paper does not state that the FLARE test split is disjoint from the DL3DV training scenes. If any overlap exists, the DL3DV rows are not zero-shot and may favor methods trained on that dataset. Please provide an explicit overlap analysis or re-evaluate on a held-out set.","section":"Sec. 4.1; Sec. A.2"},{"comment":"The paper claims 'any-prior prompting' and universal gains from priors, but Table 4 reports NVS with image-only, intrinsics, camera pose, and intrinsics+camera-pose inputs; no depth-prior NVS row is reported. Since depth is the most directly relevant prior for Gaussian placement, its absence weakens the 'any-prior' claim. Please add a depth-prior (and depth+other) NVS ablation, or explain why depth is not used in this setting.","section":"Sec. 4.1; Table 4; Abstract"},{"comment":"The phrase 'we follow the test-time camera pose alignment introduced by AnySplat' is ambiguous. If this alignment involves fitting predicted cameras to ground-truth cameras (e.g., a similarity transform), then even the 'w/o GT Cameras' variant leaks ground-truth pose information. Please define exactly what is aligned, and state whether any ground-truth geometry is used after the forward pass.","section":"Sec. 4.1, NVS paragraph"}],"minor_comments":[{"comment":"The predicted surface normal in the normal loss is written as \\hat{E}_i; this is the same symbol used for camera parameters and should be \\hat{N}_i to avoid confusion.","section":"Eq. (9)"},{"comment":"Title has a typo: 'Novel View Synthsis' should be 'Novel View Synthesis'.","section":"Table 6"},{"comment":"The four metrics and their aggregation in Fig. 6 are only tersely described. Please spell out how the bars are computed and whether values are averaged over scenes or views.","section":"Sec. 4.2, Fig. 6"},{"comment":"The limitations section mentions dynamic scenes and maximum input resolution, but does not mention the ground-truth-camera NVS protocol or the DL3DV training/evaluation overlap. These should be acknowledged and addressed in the revision.","section":"Appendix C"}],"recommendation":"major_revision","confidential_remarks":"I see no evidence of misconduct or a novelty conflict. The main gap is evaluation protocol rather than methodology. Given that the Table 6 ablation suggests the pose-free NVS version is close on RealEstate10K, a revision that reports predicted-camera comparisons against baselines and clarifies the DL3DV split could make the paper acceptable. I would ask the editor to treat the corrected NVS tables as a condition of acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: WorldMirror is a real step forward for unified 3D geometric prediction — a single VGGT-style model that handles point maps, depth, normals, cameras, and Gaussian-based NVS, with clean integration of optional priors. The prior-injection results are convincing: adding depth or camera pose helps across all tasks, not just the corresponding one. That is a nice empirical finding.\n\nThe main soft spot is the NVS evaluation. The method as described in Eq. 4 back-projects the predicted Gaussian depth using ground-truth camera poses and intrinsics. The headline Table 4 is generated that way, while AnySplat and FLARE are pose-free. The 'w/o GT Cameras' ablation in Table 6 shows the model degrades slightly on DL3DV (20.69 vs 20.91) and significantly on VR-NeRF (24.76 vs 25.75), so the drop is non-trivial on at least one benchmark. That doesn't kill the system — the pose-free variant is still competitive — but the SOTA claim for pose-free NVS is not supported by the reported numbers. The fix is simple: re-run Table 4 with predicted cameras, or clearly mark the GT-pose protocol. I'd want that before trusting the NVS comparisons.\n\nA couple of smaller things. The abstract claims SOTA for depth, but the paper's own Table 7 shows π3 clearly better on KITTI (Abs Rel 0.038 vs 0.063); the text acknowledges the gap, so the abstract should be tempered. There are no error bars anywhere, and the gains are often small, so I'd like to see variance estimates. Also, the model is initialized from VGGT, which makes it hard to attribute the improvements to the new heads versus the larger training scheme; that's fine for a system paper, but worth stating more clearly.\n\nThe point map, camera, and normal results look solid and well-benchmarked. The paper honestly lists limitations and gives a thorough description of training data and losses. The citation pattern is appropriate — Pow3R, VGGT, and AnySplat are all cited. This is not a circular argument or a fabricated claim; it's a solid engineering contribution with a genuine evaluation bug.\n\nVerdict: this deserves a serious referee, and I'd accept it with major revision. The authors need to re-generate the NVS comparison under the predicted-camera protocol and fix the abstract's overclaims. If they do, it will be a useful reference point for unified 3D perception.","headline":"Solid multi-task 3D reconstruction system with convincing prior-injection results, but the headline NVS numbers use GT cameras for back-projection, undercutting the pose-free SOTA claim.","tokens_in":19011,"tokens_out":3278,"would_cite":true,"duration_ms":27287,"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":"WorldMirror proposes a single feed-forward model that accepts any available geometric priors—camera poses, intrinsics, depth maps—and simultaneously outputs point clouds, depth, normals, camera parameters, and renderable 3D Gaussians, with","keywords":["universal 3D reconstruction","feed-forward geometry prediction","multi-modal prior prompting","camera pose estimation","depth estimation","surface normal estimation","novel view synthesis","3D Gaussian splatting"],"falsifier":"Re-run the novel-view benchmark with only the model's predicted cameras and intrinsics used to back-project Gaussian centers, keeping all other settings fixed. If the margin over the previous best feed-forward splatting method collapses, the pose-free SOTA claim fails. Separately, to test the mutual-reinforcement thesis, ablate each prior one at a time on a held-out zero-shot benchmark and check that each prior improves every output task; if a depth prior improves only depth, the 'universal gains' claim is falsified.","tokens_in":18070,"feed_emoji":"🌐","tokens_out":5722,"duration_ms":53316,"temperature":0.7,"pith_summary":"The paper tries to establish that one architecture can be a universal 3D geometry engine: given a handful of images and whatever geometric hints are available (calibrated intrinsics, camera poses, depth maps, or none), it predicts dense point maps, multi-view depth, camera parameters, surface normals, and 3D Gaussians in a single forward pass. The central claim is that input flexibility and multi-task prediction are mutually reinforcing—injecting any single prior improves not just the corresponding output but all geometric outputs, and injecting all priors gives the largest gains. If true, a single model could replace several specialized pipelines (depth estimators, pose estimators, normal predictors, novel-view renderers) and would naturally exploit sensor data that real systems often have. The paper reports state-of-the-art numbers on point-map, camera, depth, normal, and novel-view benchmarks while keeping feed-forward speed, and it shows the predicted geometry also accelerates downstream per-scene optimization.","feed_headline":"One network turns images plus any priors into full 3D geometry","feed_subtitle":"Camera poses, intrinsics, or depth maps each improve every output, from point clouds to novel views.","key_machinery":"Multi-Modal Prior Prompting: camera poses (rotation as quaternion plus normalized translation) and intrinsics (normalized focal lengths and principal point) are each compressed into a single token and concatenated with image tokens; depth maps are normalized and turned into dense tokens that are added directly to spatially aligned image tokens. During training a dynamic injection scheme randomly zeros out each prior with probability 0.5, so the model learns to exploit whichever priors are present and degrade gracefully when none are. A shared transformer backbone feeds several dense prediction heads that output point maps, depth, surface normals, and 3D Gaussians, plus a camera head; 3D Gaus","core_discovery":"The core discovery, stated on the paper's own terms, is that a model which can consume any subset of geometric priors and simultaneously predict many 3D representations learns to use prior information as a shared scaffold: a depth hint sharpens camera estimates, a pose hint clarifies depth, and so on. The authors show that this transfer is systematic—single-modality priors give dual benefits (improving the corresponding task and the rest), and the full prior set yields the best point maps, depths, normals, and renderings. They also find that compact global embeddings for poses and intrinsics (one token per view) outperform dense per-pixel conditioning, and that a curriculum from simple to co","pith_inferences":["If the universal-gain result is genuine, it implies the network is forming a shared internal scene model rather than task-specific shortcuts; a direct test would be to hold out a prior type never seen in training (e.g., sparse LiDAR depth) and check that it still lifts all tasks.","The pose-free novel-view claim should be read through the 'w/o GT Cameras' ablation, not the headline Table 4 numbers, because those are produced with ground-truth cameras for back-projection; re-running the full benchmark with only predicted cameras would settle how much of the reported margin is actually pose-free.","The finding that compact single-token embeddings beat dense ray embeddings suggests the priors act as global conditioning rather than per-pixel constraints; one consequence is that noisy or approximate priors (e.g., from an onboard SLAM) may be nearly as effective as exact ones—testable by injecting perturbed poses.","The measured marginal gains per prior could be turned into an active-sensing policy: a robot could query the cheapest prior (e.g., intrinsics) first and only request depth or poses when the expected geometric gain justifies the sensor cost."],"forward_implications":["A single model can serve as a common geometry backbone for AR, robotics, and content creation, since it outputs point clouds, camera poses, depth, normals, and renderable splats in one pass without per-scene optimization.","Any available prior—even one modality—should improve all downstream geometric tasks, so systems with RGB-D sensors, SLAM poses, or known calibration can expect a uniform quality boost rather than a task-specific one.","Providing all three priors (intrinsics, poses, depth) yields the largest measured gains, implying that sensor-rich pipelines obtain the biggest benefit.","The predicted point cloud and Gaussians serve as a strong initialization for per-scene 3D Gaussian optimization, reaching better rendering quality in fewer iterations than random initialization.","Because the model degrades gracefully when no priors are given, it can be deployed in settings where sensor information is sometimes unavailable."],"fun_headline_variants":["Any geometric prior input boosts every 3D output from one network","One model takes any priors, outputs point clouds, depths, normals, and more","Universal gains: every prior type helps every 3D task in one model","One network: any prior in, all 3D geometry out"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the headline novel-view numbers represent the pose-free system; in fact those numbers are produced with ground-truth camera poses and intrinsics, so the pose-free claim rests entirely on the 'w/o GT Cameras' ablation row in Table 6.","fun_headline_variants_meta":{"raw":{"variants":["Any geometric prior input boosts every 3D output from one network","One model takes any priors, outputs point clouds, depths, normals, and more","Universal gains: every prior type helps every 3D task in one model","One network: any prior in, all 3D geometry out"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000644,"raw_usage":{"total_tokens":2752,"prompt_tokens":653,"completion_tokens":2099,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":397,"completion_tokens_details":{"reasoning_tokens":2017}},"tokens_in":397,"tokens_out":2099,"duration_ms":14462,"temperature":1.0,"reasoning_tokens":2017,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T10:15:40.139029+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the novel-view benchmark with only the model's predicted cameras and intrinsics used to back-project Gaussian centers, keeping all other settings fixed. If the margin over the previous best feed-forward splatting method collapses, the pose-free SOTA claim fails. Separately, to test the mutual-reinforcement thesis, ablate each prior one at a time on a held-out zero-shot benchmark and check that each prior improves every output task; if a depth prior improves only depth, the 'universal gains' claim is falsified.","supporting_citations":[],"review_version":1}