{"id":"0b489e16-ef8e-4a79-abe5-fb581422eda3","arxiv_id":"2411.11903","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DiHuR uses learnable SMPL-vertex tokens and a 2D diffusion prior to reconstruct detailed 3D humans from sparse multi-view images, beating prior models on standard datasets.","lead":"The paper presents DiHuR, a model that reconstructs 3D human bodies from three photos with little overlap by combining learned SMPL-vertex tokens with a diffusion-based geometry prior. It reports state-of-the-art reconstruction and view-synthesis accuracy on THuman, ZJU-MoCap and HuMMan without 3D supervision during training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-scene SDS finetuning is applied only to DiHuR and no frozen feed-forward baseline is reported, so the claimed generalizable-architecture advantage is confounded with test-time optimization.","rationale":"The reader's CONDITIONAL verdict is appropriate. The most load-bearing issue is not whether the diffusion prior is perfect, but that the reported SOTA numbers are produced by a test-time optimization stage applied only to DiHuR, while the paper's title and central claim emphasize generalizable reconstruction. Section 3.4/3.5 describe 150 iterations of per-scene SDS finetuning with additional RGB, Eikonal, and smoothness losses. Tables 1-5 report this final finetuned model; no table reports the frozen feed-forward model, so the reader cannot separate architectural generalization from optimization. The ablation in Table 7 ('w/o diffusion guidance') is suggestive but ambiguous because the text implies per-scene finetuning still occurs without the SDS term. The gain from diffusion is substantial (NC +0.066, CD -0.227), so the comparison against baselines that lack such a stage is confounded. I agree with the reader's weakest-assumption identification and would keep the verdict CONDITIONAL, with the added condition that the authors report frozen feed-forward numbers and a baseline granted the same per-scene optimization. If those numbers show the architecture alone is competitive, the paper's central claim stands; if not, the claim should be restated as 'feed-forward reconstruction plus per-scene diffusion refinement.' The paper does provide useful evidence: ablations, view-count scaling, and cross-dataset HuMMan results, but no code release and the missing zero-shot evaluation prevent full verification.","tokens_in":10746,"tokens_out":6826,"duration_ms":78033,"concrete_test":"Run DiHuR in pure feed-forward mode on the exact THuman and ZJU-MoCap test splits: load the pretrained weights, render SDF/normals with no inference-time optimization (no SDS, no RGB finetuning, no Eikonal/smoothness finetuning), and report CD/NC under the same 3-view protocol. Then run one strong baseline (e.g., GP-NeRF or SIFU) through the same 150-iteration per-scene finetuning with the same normal-map SDS loss and text condition and report CD/NC. If frozen DiHuR does not beat the feed-forward baselines on both metrics, or if the SDS-finetuned baseline matches DiHuR's final numbers, the central claim of a generalizable architecture is not supported by the current experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Every headline 3D-reconstruction comparison (Tabs 1 and 3) reports DiHuR after the Sec. 3.4/3.5 inference-time finetuning: 150 iterations of SDS plus image-reconstruction and regularizer losses. Most baselines (SparseNeuS, GP-NeRF, SiTH, SIFU) are evaluated as frozen feed-forward networks, so the comparison bundles the proposed architecture with an extra per-scene optimization budget. The paper never reports the zero-shot output of the pretrained DiHuR network before any finetuning. The 'w/o diffusion guidance' row in Tab. 7 cannot resolve this: Sec. 3.5 states that only the SDS term is added during the per-scene finetuning stage, implying the stage itself (with Lrgb, Leik, Lsm) may still run when Lsds is removed. If so, that ablation still includes test-time optimization, and the true contribution of the learned SMPL tokens and multi-target training to generalization remains unknown. The magnitude of the effect is large: removing diffusion guidance drops NC from 0.779 to 0.713 and CD from 1.117 to 1.344 on THuman, so the reported margin over SIFU (NC 0.779 vs 0.741) is largely produced by the per-scene SDS refinement. Because the diffusion model is a super-resolution model trained on natural RGB images but is here fed rendered normal maps, the geometric validity of its SDS gradients is also not independently checked against ground-truth normals or depth.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DiHuR proposes a generalizable human reconstruction method from sparse, minimally overlapping multi-view images. The method combines learnable tokens attached to SMPL vertices that aggregate sparse-view image features, an SDF-based volume rendering pipeline, and a per-scene inference-time finetuning stage driven by score distillation sampling (SDS) from a pretrained super-resolution diffusion model. The paper claims state-of-the-art results on THuman, ZJU-MoCap, and HuMMan datasets for both 3D reconstruction and novel view synthesis, with all training done using multi-view images only and no 3D supervision. The main architectural components are a learnable SMPL-token cross-view attention module, KNN-based feature interpolation, a multi-target patch sampling strategy, and the diffusion-guided finetuning step.","tokens_in":11064,"tokens_out":4586,"duration_ms":44311,"significance":"If the claims hold, DiHuR would be a practically useful contribution to sparse-view 3D human reconstruction, combining a feed-forward generalizable radiance field with a strong 2D diffusion prior without requiring 3D supervision. The paper is clearly written, the ablations in Table 7 confirm that each proposed component contributes to the reported quality, and the evaluation covers both within-dataset and cross-dataset settings. The main limitation is that the reported headline numbers are obtained after a per-scene inference-time finetuning step that is not applied to most baselines, so the claimed advantage of the generalizable architecture is not cleanly established. No code or repeated-run statistics are provided, which limits the verifiability of the moderate performance margins over prior work.","major_comments":[{"comment":"The comparisons in Tables 1–3 are not apples-to-apples. DiHuR is evaluated after 150 iterations of per-scene finetuning (Sec. 3.4) that includes the SDS loss together with color, Eikonal, and smoothness losses, whereas the baselines SparseNeuS, GP-NeRF, SiTH, and SIFU are evaluated as frozen feed-forward networks. The manuscript never reports the zero-shot output of the pretrained DiHuR network before any finetuning. The 'w/o diffusion guidance' row in Table 7 does not resolve this confound: per Sec. 3.5, the per-scene finetuning stage still includes Lrgb, Leik, and Lsm even when Lsds is removed. Therefore the true contribution of the learnable SMPL tokens and multi-target training to generalization remains unknown, and the reported margins over SIFU (e.g., THuman NC 0.779 vs 0.741) may be substantially produced by the extra per-scene optimization budget.","section":"§3.4–§3.5, Tables 1–3"},{"comment":"The P.S. (per-scene optimization) column in Table 4 is mislabeled for the proposed method. DiHuR performs per-scene finetuning with SDS at inference, so it should be marked as ✓ under the same definition used for NB. Marking Ours as ✗ obscures a key difference between DiHuR and the feed-forward baselines (NHP, MPS-NeRF, GP-NeRF) in the novel-view-synthesis comparison and makes the table misleading as reported.","section":"Table 4"},{"comment":"The diffusion prior is a latent super-resolution model trained on natural RGB images, but it is applied here to rendered normal maps conditioned on a text prompt. The paper provides no independent validation of the geometric correctness of these SDS gradients, e.g., by comparing SDS-refined normals or depth against ground truth. The final CD/NC metrics conflate the diffusion prior's geometric contribution with the per-scene finetuning, so it is unclear whether the diffusion guidance is a reliable geometric prior or a source of hallucinated detail. An experiment that compares SDS-refined normals to ground-truth normals, or an ablation with an equal-budget per-scene optimization without SDS, is needed to substantiate the claim that the diffusion prior improves geometry.","section":"§3.4, Eq. (11)"},{"comment":"No error bars, standard deviations, or repeated-run statistics are reported. The performance margins over prior methods, such as THuman NC 0.779 vs 0.741 and CD 1.117 vs 1.521, are reported without any measure of variability, making it impossible to assess whether the improvements are statistically significant, especially because each test subject receives an independent per-scene finetuning step.","section":"Tables 1–3, Table 2"}],"minor_comments":[{"comment":"The sentence 'subsequently, generalizable NeRFs are [2,8,25,28] are proposed' contains a duplicated 'are' and should be rephrased for clarity.","section":"Sec. 2"},{"comment":"The citation in Sec. 2 appears as '[?, 9, 13]' with an unresolved placeholder, which should be fixed before publication.","section":"Sec. 2, reference list"},{"comment":"Equation (2) is missing a closing parenthesis in the SoftMax expression, and Eq. (9) writes 'i' where the per-point color 'c_i' should appear in the volume rendering accumulation.","section":"Eq. (2) and Eq. (9)"},{"comment":"The row label 'w/o leanable code' contains a typo; it should read 'w/o learnable code'.","section":"Table 7"},{"comment":"The caption contains an erroneous space in 'V olume rendering' and should be corrected to 'Volume rendering'.","section":"Fig. 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim of state-of-the-art generalizable reconstruction is likely overstated because the reported numbers include a per-scene SDS finetuning stage that is not applied to baselines. If the authors can supply zero-shot results before finetuning and an apples-to-apples comparison with same-budget per-scene optimization, the contribution may stand. The mislabeling in Table 4 and the missing error bars are also concerning and should be addressed in revision. The unresolved citation placeholder in Sec. 2 should be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on DiHuR. The paper has a genuine new combination—SMPL-anchored learnable tokens for cross-view feature aggregation, SDF volume rendering, multi-target patch sampling, and an inference-time SDS finetuning on rendered normal maps using a super-resolution diffusion model. The ablations in Table 7 show each piece contributes, and the cross-dataset results on HuMMan are credible. The feed-forward training without 3D supervision is a nice property.\n\nThe soft spot is exactly what the stress-test note flags: every headline comparison in Tables 1 and 3 runs DiHuR after 150 steps of per-scene SDS finetuning while the baselines are frozen feed-forward networks. The w/o diffusion guidance row in Table 7 doesn't fix this, because the finetuning stage still runs with the other losses. So the reported margin over SIFU is largely produced by the extra optimization budget, not by the learned tokens or architecture. We never see the zero-shot output of the pretrained network. Also, no error bars, no repeated runs, no code, and the diffusion prior is an RGB-trained super-resolution model applied to normal maps without any independent check on geometric validity.\n\nNone of this makes the paper nonsense. The method is sensible and the components are well-chosen. But the central claim—that this architecture generalizes better—is confounded with test-time optimization. The fix is straightforward: report the zero-shot numbers, run a fair baseline comparison with the same finetuning protocol (or at least a stronger frozen baseline), and give variance over seeds. A careful referee should push for exactly that.\n\nWho is this for? Someone working on sparse-view human reconstruction or generalizable neural fields. It deserves a serious referee, because the confound is addressable and the idea is worth testing properly.\n\nMy recommendation: send to peer review, but make the authors close the apples-to-oranges gap before acceptance.","headline":"Real method, unfair comparison: DiHuR's gains are confounded with test-time SDS finetuning that baselines don't get.","tokens_in":11624,"tokens_out":2099,"would_cite":false,"duration_ms":20797,"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 generalizable radiance field can reconstruct a detailed 3D clothed human from as few as three minimally overlapping images, trained only on multi-view photos, by anchoring learnable tokens to SMPL body vertices and refining the…","keywords":["3D human reconstruction","sparse multi-view","generalizable radiance field","SMPL body model","signed distance function","score distillation sampling","diffusion prior","novel view synthesis"],"falsifier":"Take a held-out subject in complex clothing with a ground-truth 3D scan, reconstruct it from three views with and without the SDS finetune, and compare each mesh to the scan by Chamfer distance and normal consistency: if enabling SDS systematically lowers the 2D normal-map error while leaving Chamfer distance unchanged or worse, the diffusion prior is adding plausible but ungrounded detail and the paper's central explanation for its gains is wrong.","tokens_in":1867,"feed_emoji":"🧍","tokens_out":2640,"duration_ms":109931,"temperature":0.7,"pith_summary":"This paper establishes that a generalizable radiance field can reconstruct detailed 3D clothed humans from three minimally overlapping views using only multi-view images for training, with no 3D scans. The key move is to anchor learnable feature tokens on the fixed vertices of the SMPL body model, so that sparse view features are aggregated through cross-view attention in a way that transfers across identities. To recover folds and clothing detail that the body model cannot express, the paper then refines the signed-distance field during inference using score distillation from a pretrained 2D diffusion model, conditioned on rendered normal maps. The authors report state-of-the-art Chamfer distance and normal consistency on THuman, ZJU-MoCap, and a zero-shot HuMMan evaluation, and improved novel-view synthesis as a byproduct. If correct, this means practical sparse multi-camera capture can replace dense rigs and 3D supervision for human digitization.","feed_headline":"3 views, no 3D scans: DiHuR tops human reconstruction benchmarks","feed_subtitle":"The method tops prior work on THuman, ZJU-MoCap, and zero-shot HuMMan.","key_machinery":"The mechanism is a pipeline built around SMPL, a parametric human body model with a fixed 6,890-vertex mesh. Learnable tokens attached to those vertices act as query tokens in a multi-head cross-view attention module; a softmax over source views yields per-vertex aggregated features that a self-attention layer then correlates across the whole body. For each query point on a camera ray, the K nearest SMPL vertices are interpolated by inverse distance, and the resulting feature is concatenated with the mean and variance of directly projected image features to predict the signed distance value; NeuS-style volume rendering converts the signed distance into density for color accumulation. During the inference finetune, the same field renders normal maps from eight evenly spaced views; those are upsampled 4x and run through a pretrained latent super-resolution diffusion model, and the score distillation loss back-propagates the diffusion model's denoising error into the SDF network, while a fixed low-resolution normal map and a text prompt serve as conditioning. A multi-target training strategy samples patches from the same body segment across several target views at once, so rays from different views intersect and implicitly constrain the surface.","core_discovery":"On the paper's own terms, the central claim is that a sparse-view human reconstruction network fails not because there is too little image evidence but because naive feature averaging cannot tell occluded from visible points. DiHuR solves this by making the 6,890 SMPL vertex positions the carriers of correspondence: learned tokens query each source view, self-attention exchanges information among vertices, and KNN interpolation delivers a dense feature field that guides SDF prediction. The second half of the claim is that a super-resolution latent diffusion model, fed an upsampled rendered normal map and the text condition 'Best quality, human, normal map', provides enough geometric gradient via score distillation to add wrinkles and folds in a roughly two-minute per-scene finetune, without 3D supervision. Combined, the two priors achieve Chamfer distance 1.117 and normal consistency 0.779 on THuman, 0.790 and 0.767 on ZJU-MoCap, and 1.23 and 0.753 on zero-shot HuMMan, beating the strongest compared method, SIFU, in every row.","pith_inferences":["Beyond the paper, the SMPL-token design is likely portable to other parametric models with consistent vertex semantics, such as articulated hands or animal bodies, though the paper only demonstrates humans.","A direct audit of the diffusion prior is still open: comparing SDS-refined normals against high-resolution ground-truth scans would show how much of the added detail is recovered geometry versus plausible hallucination, which aggregate Chamfer and normal-consistency numbers cannot separate.","Because the tokens carry the cross-identity prior through SMPL correspondences, their quality is bounded by SMPL registration accuracy; a poor pose or shape fit would likely corrupt the semantic alignment and spoil the geometry even with diffusion guidance.","The multi-target patch sampling principle, forcing rays from different views to intersect the same body part, could be adopted by other sparse-view reconstruction methods as a cheap way to impose multi-view consistency."],"forward_implications":["With three cameras spaced roughly 120 degrees apart, the reported Chamfer distance on THuman falls to 1.117, indicating a minimal capture rig can produce avatar-quality surfaces.","A single trained network runs zero-shot on HuMMan without retraining, so new subjects can be absorbed from just a handful of views.","Novel view synthesis improves alongside geometry: the method reports higher PSNR and SSIM than prior generalizable human NeRFs on ZJU-MoCap and THuman, attributed to blending weights that benefit from better geometry.","Per-scene refinement takes about 150 iterations, roughly two minutes, so adding diffusion guidance remains practical for capture workflows.","Because no 3D ground truth is required, the same recipe could be applied to any calibrated multi-view capture of articulated subjects."],"supporting_citations":[{"why":"Supplies the SMPL parametric body model whose fixed vertex set anchors the learnable tokens and provides the human prior.","marker":"[15]"},{"why":"Supplies the NeuS signed-distance-to-density volume rendering formulation used for surface extraction and color accumulation.","marker":"[24]"},{"why":"Introduces the score distillation sampling loss that back-propagates diffusion gradients into the SDF network.","marker":"[19]"},{"why":"Provides the pretrained latent super-resolution diffusion model adapted as the 2D normal-map prior in the SDS loss.","marker":"[20]"},{"why":"Defines the generalizable human NeRF baseline whose dataset splits and evaluation protocol the paper follows.","marker":"[4]"},{"why":"SIFU is the strongest single-view baseline on THuman and ZJU-MoCap and the main comparison target for CD and NC.","marker":"[29]"},{"why":"Provides the ZJU-MoCap dataset and the per-scene Neural Body baseline for novel view synthesis and reconstruction.","marker":"[18]"},{"why":"Supplies the Eikonal regularization term used in the total loss to keep the signed distance field well-behaved.","marker":"[6]"},{"why":"Provides the HuMMan dataset used for the zero-shot cross-dataset generalization evaluation.","marker":"[1]"}],"fun_headline_variants":["Sparse views, no 3D labels: DiHuR redefines human reconstruction","Diffusion-guided DiHuR beats SIFU on all benchmarks","From 3 views to full body: DiHuR's dual-prior trick","DiHuR: SMPL tokens plus diffusion for sparse-view 3D humans","DiHuR wins without 3D supervision using diffusion prior"],"cache_read_input_tokens":13696,"weakest_assumption_plain":"The approach rests on the assumption that the 2D diffusion model, conditioned on a low-resolution normal map and the text prompt 'Best quality, human, normal map', provides honest geometric detail through its score-distillation gradients, rather than inventing wrinkles that look right in 2D but are wrong on the actual 3D body.","fun_headline_variants_meta":{"raw":{"variants":["Sparse views, no 3D labels: DiHuR redefines human reconstruction","Diffusion-guided DiHuR beats SIFU on all benchmarks","From 3 views to full body: DiHuR's dual-prior trick","DiHuR: SMPL tokens plus diffusion for sparse-view 3D humans","DiHuR wins without 3D supervision using diffusion prior"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000597,"raw_usage":{"total_tokens":2827,"prompt_tokens":1010,"completion_tokens":1817,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":1730}},"tokens_in":626,"tokens_out":1817,"duration_ms":11486,"temperature":1.0,"reasoning_tokens":1730,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:24:38.194101+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out subject in complex clothing with a ground-truth 3D scan, reconstruct it from three views with and without the SDS finetune, and compare each mesh to the scan by Chamfer distance and normal consistency: if enabling SDS systematically lowers the 2D normal-map error while leaving Chamfer distance unchanged or worse, the diffusion prior is adding plausible but ungrounded detail and the paper's central explanation for its gains is wrong.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the SMPL parametric body model whose fixed vertex set anchors the learnable tokens and provides the human prior."},{"cited_title":"Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction","cited_arxiv_id":null,"evidence_quote":"Supplies the NeuS signed-distance-to-density volume rendering formulation used for surface extraction and color accumulation."},{"cited_title":"Geometry-guided progressive nerf for generalizable and efficient neural human rendering","cited_arxiv_id":null,"evidence_quote":"Defines the generalizable human NeRF baseline whose dataset splits and evaluation protocol the paper follows."},{"cited_title":"Sifu: Side- view conditioned implicit function for real-world usable clothed human reconstruction","cited_arxiv_id":null,"evidence_quote":"SIFU is the strongest single-view baseline on THuman and ZJU-MoCap and the main comparison target for CD and NC."},{"cited_title":"Implicit geometric regularization for learning shapes","cited_arxiv_id":null,"evidence_quote":"Supplies the Eikonal regularization term used in the total loss to keep the signed distance field well-behaved."},{"cited_title":"HuMMan: Multi-modal 4d human dataset for versatile sensing and modeling","cited_arxiv_id":null,"evidence_quote":"Provides the HuMMan dataset used for the zero-shot cross-dataset generalization evaluation."}],"review_version":1}