{"id":"4dcc55cd-e984-4e51-87f5-31e4425603e6","arxiv_id":"2412.08412","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Pragmatist turns sparse unposed photos of an object into a high-fidelity 3D mesh by generating consistent canonical views with a diffusion model, reconstructing a triplane mesh, then refining camera poses and texture from the original views.","lead":"Pragmatist reconstructs 3D objects from a few photos taken at unknown camera positions by first generating extra consistent views with a diffusion model, then building a mesh and refining poses and texture. The paper reports strong benchmark gains, but ships no code or data and leaves several evaluation details ambiguous, so the practical result is conditional on confirmation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Canonicalization is the load-bearing step: the pose-free generator must infer a globally consistent object frame from appearance alone, but Table 3 validates this only in a self-referential way; if that frame drifts, Eq. 4 and Eq. 5 inherit the error.","rationale":"For the central claim to hold, the generator must do more than produce pleasing images: its output must be geometrically consistent and anchored to a single metric frame of the input object. The paper fixes the output pose grid but gives no loss, metric, or ablation that verifies the implied input-to-canonical alignment; the ablation in Figure 5 only shows that adding the generator helps, not that the canonical frame is correct. The reported pose accuracy in Table 3 cannot settle this because it evaluates the model against itself. The reader's conditional verdict is therefore appropriate; my read does not move it. I do not claim the method is wrong: the principle is plausible and the qualitative results are suggestive, but the canonicalization assumption is the load-bearing point and currently lacks non-circular evidence. Release of code and per-instance metrics would help, but the two-input-set consistency test is the decisive experiment.","tokens_in":15593,"tokens_out":7003,"duration_ms":82615,"concrete_test":"Canonicalization-consistency test on the GSO split: for each of ~100 objects, sample two independent sets of 4 unposed inputs with deliberately different camera distributions (e.g., elevations ±30° vs ±60°, distances 1.2 vs 2.0). Generate canonical views and reconstruct meshes without refinement; align both meshes to the ground-truth mesh by ICP. Record the relative rigid/scale transform between the two canonical frames and each mesh's CD to GT. If the relative transform exceeds about 10° in rotation or 15% in scale, or if CD degrades substantially relative to Table 4, the pose-free canonicalization is not stable. As a secondary check, recompute Table 3 after ICP-aligning the reconstructed mesh to GT, not to the canonical mesh.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the Canonical Camera System (Method, Multiview Conditional Multiview Generator): target views are fixed at elevation 0°, distance 1.5, with azimuth offsets β±k·45°, but condition views Icond arrive with no poses. Eq. 1 writes Pcond as conditioning, yet at inference it is dropped; the model is expected to infer the input object's orientation, scale, and center in a canonical frame from RGB alone. This is the ill-posed part of the problem, relocated into the diffusion generator. If the inferred canonical frame is wrong, the triplane reconstruction, the pose inversion in Eq. 4, and the texture refinement in Eq. 5 are all anchored to the wrong coordinate system; refinement then minimizes photometric error against input views under a false pose, so it can blur or warp instead of sharpen. The only quantitative pose evidence, Table 3, is self-referential: errors are measured with respect to the model's own canonical mesh rather than ground-truth object coordinates (the GT-alignment procedure described for CD/IoU is not applied there). Thus the central claim that the pipeline 'circumvent[s] the direct resolution of highly ill-posed problems' is currently supported by qualitative figures and by a metric that presupposes the canonicalization it is meant to establish.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Pragmatist, a three-stage pipeline for reconstructing a 3D object from sparse unposed views. First, a pose-free multi-view conditional diffusion model generates several novel views in a canonical object coordinate system. Second, a feed-forward large reconstruction model predicts a triplane NeRF and extracts a mesh from these generated views. Third, the input views are registered to the canonical frame by optimizing camera poses against the reconstructed mesh, and the texture is refined using the high-resolution input views. The authors report improvements over pose-free and pose-based baselines on GSO and OmniObject3D, plus ablation studies supporting the role of the generative prior and refinement.","tokens_in":15875,"tokens_out":3911,"duration_ms":41634,"significance":"If the results hold, the paper offers a practical reformulation of unposed sparse-view reconstruction: instead of solving an ill-posed inverse problem directly, the method synthesizes canonical observations with a generative model and then uses a feed-forward reconstructor plus test-time refinement. This is a plausible and potentially useful system contribution, and the empirical comparison covers a broad set of recent baselines. The paper does not provide code or pre-trained models, and the quantitative evidence currently lacks statistical robustness and a non-self-referential pose evaluation.","major_comments":[{"comment":"The pose accuracy evaluation in Table 3 is self-referential and needs clarification. The text describes pose recovery as optimizing the camera pose p of I_cond against the mesh obtained from the feed-forward reconstruction (Eq. 4), but it does not specify what the ground-truth pose is for unposed input views. The CD/IoU evaluation aligns meshes with ICP, but no analogous alignment step is described for pose evaluation. If the reported rotation and translation errors are computed with respect to the model's own canonical mesh, then Table 3 measures only the convergence of the pose optimization, not the accuracy of the canonicalization. This is load-bearing because the refinement in Eq. (5) and the 'Ours Pred.' row in Table 2 depend on these poses. Please specify the reference frame, how ground-truth poses are obtained, and whether any alignment to the GT object is applied before computing R. error and T. error.","section":"Table 3, Eq. (4)"},{"comment":"There is an inconsistency between the formal model in Eq. (1), which conditions generation on Icond and Pcond, and the later statement that the model works 'without explicitly using the camera pose information of the conditional views.' The conditioning feature extraction paragraph describes only an image encoder; it does not explain how Pcond is encoded or whether it is dropped at inference. If Pcond is used only during training to define the canonical target views relative to the condition views, that should be stated explicitly. As written, the reader cannot determine whether the model at inference is being asked to do something different from its training objective, which is central to the canonicalization claim.","section":"Eq. (1) and 'Conditioning Feature Extraction'"},{"comment":"The evaluation protocol is under-specified. No error bars, number of test objects, number of sampled views per object, or random seed information is reported, so it is unclear whether the improvements in Tables 2, 4, and 5 are statistically significant. Furthermore, the rows 'Ours' and 'Ours Pred.' in Table 2 are not defined in the text; the reader cannot tell what each variant does (e.g., whether 'Ours' uses the canonical generated views without refinement, or uses predicted poses for the input views). Please provide a precise evaluation protocol, standard deviations, and a clear definition of every reported variant.","section":"Tables 2-5 and 'Datasets' paragraph"},{"comment":"The load-bearing assumption that the diffusion model can map arbitrary unposed views to a canonical object coordinate system is not quantitatively validated. The 'Canonical Camera System' fixes target views at elevation 0 and distance 1.5 with azimuth offsets, but the condition views have arbitrary pose, and the model must infer the object's orientation, scale, and center from RGB alone. If the inferred canonical frame drifts, the pose optimization in Eq. (4) and the texture refinement in Eq. (5) are anchored to a wrong coordinate system, so the photometric loss can blur or warp instead of sharpen. The only quantitative evidence of canonicalization is Table 3, which is self-referential as noted above. Please evaluate canonicalization directly, for example by comparing the generated canonical views to ground-truth canonical renderings of the ICP-aligned GT mesh, or by reporting failure cases.","section":"Canonical Camera System, Eq. (5)"}],"minor_comments":[{"comment":"In the 'Ours' row of Table 4, the pose column contains 'Pred.' at the end of the row, which appears to be a typo.","section":"Table 4"},{"comment":"The loss weight λ in Eq. (2) is defined only in the experimental section, and the text does not clearly connect λ in Eq. (2) to λp in the implementation. Please give the value in the equation or refer to the implementation paragraph at the first use.","section":"Eqs. (2)-(3) and Implementation details"},{"comment":"The abbreviation 'Gen Occ.' in Table 1 is not expanded in the table or the text; please spell out 'Generative Occlusion' or whatever is intended.","section":"Table 1"},{"comment":"References Wei et al. 2023a and 2023b appear to be the same paper (NeuralManifold) and should be merged; the supplementary material cites Sohl-Dickstein without a corresponding reference entry.","section":"References"},{"comment":"The claim that 'optimizing the triplane feature and color MLP at the same time achieves better results than optimizing the color MLP alone' is not supported by any table or figure; please provide quantitative evidence.","section":"Ablation of Virtual Camera Guided Refinement"},{"comment":"The comparison with HLoc in Table 3 is unclear: HLoc is a full SfM pipeline, and it is not stated whether it was run under the same sparse-view setting as the other methods. Please clarify the exact input to each baseline.","section":"Pose Estimation Results"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid systems contribution, and the central idea is interesting, but the current evidence is not yet convincing for acceptance. The self-referential pose evaluation is the most serious issue; without a non-circular evaluation the canonicalization claim is unsupported. I also think the missing definition of the Ours vs Ours Pred. variants and the lack of error bars should be addressed. This is likely fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: a plausible, well-engineered pipeline that may be a real step forward for unposed sparse-view reconstruction, but as reported the evidence does not yet support the headline numbers.\n\nThe genuinely new element is the framing: instead of predicting geometry directly from unposed inputs, generate a complete set of canonical views with a pose-free multiview diffusion model, then reconstruct from those views and recover input poses by inverting the triplane. The individual pieces—multi-view diffusion, LRM, DiffMC, pose/refinement—are borrowed, but the integrated pose-free canonicalization is not present verbatim in the cited work. The ablations are in the right direction and do support the claim that generative priors and the refinement stage matter. The method also handles a variable number of input views, which is practically useful.\n\nThe soft spots are about evidence, not the core idea. There is no code, no data, no error bars, an unspecified Objaverse filter, an unclear evaluation subset, and no explanation of what 'Ours Pred.' means in Table 2 (as opposed to 'Ours' with unposed inputs). The pose benchmark in Table 3 is self-referential: errors are measured relative to the model's own canonical mesh, not ground-truth object coordinates, so it cannot establish that the canonicalization is actually correct. There is also an expository inconsistency in Eq. 1, which lists Pcond as conditioning even though the text elsewhere says poses of condition views are not used. Those are fixable, but they matter because canonicalization is the load-bearing step. If the canonical frame drifts, Eq. 4 and Eq. 5 inherit the error. The paper does not currently show that this doesn't happen.\n\nOverall: this is a serious engineering effort on a practical problem. The reader should not accept the central benchmark claim without artifacts and a clean pose-evaluation protocol, but the idea deserves referee time. I'd send it to review with a request for code/data, error bars, clarification of the evaluation protocol, and a pose benchmark against ground truth rather than the model's own frame.","headline":"A well-engineered pipeline that may be a real step forward for unposed sparse-view reconstruction, but the reported evidence does not yet prove the headline gains.","tokens_in":16392,"tokens_out":2736,"would_cite":false,"duration_ms":28592,"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":"Generate first, reconstruct later: 3D from unposed views","keywords":["3D reconstruction","unposed sparse views","multi-view diffusion","novel view synthesis","canonical camera system","triplane representation","pose refinement"],"falsifier":"Render test objects from extreme elevations and distances far from the canonical 0° elevation / 1.5 distance setting, run the pipeline, and measure the multi-view consistency of the generated canonical views (e.g., feature-metric reprojection error under recovered poses) and the final mesh Chamfer distance. If consistency or reconstruction degrades sharply at out-of-distribution viewpoints, the canonical camera system is the load-bearing constraint.","tokens_in":15353,"feed_emoji":"🧊","tokens_out":6231,"duration_ms":55843,"temperature":0.7,"pith_summary":"The paper argues that sparse unposed 3D reconstruction should be split into two stages: first, a pose-free multi-view diffusion model generates a complete set of canonical views of the object; second, geometry is reconstructed from those generated views. The claim is that this reformulation avoids the direct, ill-posed regression from pixels to 3D and lets the model plausibly fill in unseen regions using learned geometric priors. If true, casual sparse captures—images with no camera poses—can yield high-fidelity meshes and textures, and even single images become competitive with dedicated single-view methods. Recovering input-view poses by inverting the reconstructed 3D model, rather than estimating them from 2D images, is part of the mechanism that makes this work.","feed_headline":"Generate first, reconstruct later: 3D from unposed views","feed_subtitle":"Pragmatist turns sparse unposed images into canonical views, then reconstructs mesh and pose from the generated set.","key_machinery":"The central object is the canonical camera system: a fixed set of target views at 0° elevation and distance 1.5 with azimuth offsets β, β±45°, β±90°, β±135°, and β±180°. The diffusion model conditions on unposed inputs through a lightweight encoder and injects pixel-level camera conditions via Plücker ray embeddings, while extending self-attention to cross-frame attention across target views and cross-attention between input and target views. This machinery embodies the pose-free consistency learning that maps arbitrary casual viewpoints into a shared canonical frame. The triplane representation produced by the reconstruction model is the object that gets inverted for pose recovery and refined with input-view textures.","core_discovery":"Pragmatist's central claim is that unposed sparse-view 3D reconstruction can be recast as conditional novel view synthesis in a canonical object coordinate system. A multi-view conditional diffusion model, conditioned on arbitrary unposed input views via a lightweight encoder and cross-frame attention, generates several novel views at fixed elevation and distance with azimuth offsets. These canonical views are fed to a feed-forward triplane reconstruction model, which predicts a NeRF and then a mesh via differentiable marching cubes and rasterization. Input-view poses are recovered by inverting the triplane representation through differentiable rendering, and texture is refined using the high-resolution input views. The paper's experiments support that generating complete observations before reconstruction improves fidelity over direct pose-free prediction and over pipelines that estimate poses in image space.","pith_inferences":["The generate-then-reconstruct paradigm suggests that any sufficiently strong multi-view generative prior could replace the diffusion model, and the canonical coordinate design could be relaxed to estimate elevation and scale per object, extending the method to in-the-wild captures with varying camera intrinsics.","Because the generator is trained without pose conditioning on inputs, the same architecture could be trained on large-scale video data, potentially lifting the approach from objects to scenes—a direction the paper states as future work.","A testable extension is to provide the diffusion model with noisy or partial pose priors, such as rough azimuth ordering, and measure whether consistency and reconstruction improve; if they do, the pose-free encoding is not fully capturing viewpoint information."],"forward_implications":["Sparse unposed inputs can be turned into fully observed reconstruction problems, so the reconstruction stage sees complete observations rather than partial ones.","Recovering input-view poses by inverting the generated 3D representation is far more accurate than direct image-space pose estimation, with rotation error of 2.182 versus 67.580 on the GSO dataset.","The pipeline supports flexible numbers of input views, including a single image, and its reconstruction quality improves as more unposed views are added.","Separating density and color MLPs lets appearance refinement improve texture fidelity without degrading geometry."],"supporting_citations":[{"why":"LEAP is the main pose-free reconstruction baseline that the method compares against and improves upon.","marker":"(Jiang et al. 2023)"},{"why":"UpFusion is a pose-free novel-view diffusion baseline for unposed sparse inputs that defines the comparison setting.","marker":"(Kani et al. 2023)"},{"why":"Zero-1-to-3 supplies the single-view 3D-aware diffusion prior that the multi-view conditioning extends.","marker":"(Liu et al. 2023a)"},{"why":"LRM is the feed-forward large reconstruction model whose triplane and NeRF design the reconstruction module builds on.","marker":"(Hong et al. 2023)"},{"why":"LGM is a sparse-view reconstruction baseline that requires poses and is used to compare the generated-view pipeline against pose-conditioned methods.","marker":"(Tang et al. 2024a)"},{"why":"DiffMC provides the differentiable marching cubes used to extract meshes from the triplane density field.","marker":"(Wei et al. 2023a)"},{"why":"The nvdiffrast differentiable rasterizer is used in surface rendering and in camera and texture refinement.","marker":"(Laine et al. 2020)"},{"why":"Incremental SfM's relocalization concept motivates registering input views into the reconstruction through pose optimization.","marker":"(Schonberger and Frahm 2016)"},{"why":"The LPIPS perceptual loss is used in the volume and surface training objectives.","marker":"(Chen and Koltun 2017)"}],"fun_headline_variants":["From sparse unposed views: generate, then reconstruct","Diffusion generates views, then 3D reconstruction follows","Turn unposed sparse views into 3D by generating first","Pragmatist: generate complete views, then reconstruct 3D","Unposed views? Generate canonical views, then build mesh"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The multi-view diffusion model can learn, without any input camera poses, to map arbitrary unposed views into a shared canonical object coordinate system and produce geometrically consistent novel views from them.","fun_headline_variants_meta":{"raw":{"variants":["From sparse unposed views: generate, then reconstruct","Diffusion generates views, then 3D reconstruction follows","Turn unposed sparse views into 3D by generating first","Pragmatist: generate complete views, then reconstruct 3D","Unposed views? Generate canonical views, then build mesh"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000536,"raw_usage":{"total_tokens":2572,"prompt_tokens":941,"completion_tokens":1631,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":1547}},"tokens_in":557,"tokens_out":1631,"duration_ms":11232,"temperature":1.0,"reasoning_tokens":1547,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:51:58.268242+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render test objects from extreme elevations and distances far from the canonical 0° elevation / 1.5 distance setting, run the pipeline, and measure the multi-view consistency of the generated canonical views (e.g., feature-metric reprojection error under recovered poses) and the final mesh Chamfer distance. If consistency or reconstruction degrades sharply at out-of-distribution viewpoints, the canonical camera system is the load-bearing constraint.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LRM is the feed-forward large reconstruction model whose triplane and NeRF design the reconstruction module builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The nvdiffrast differentiable rasterizer is used in surface rendering and in camera and texture refinement."}],"review_version":1}