{"id":"7b0542bd-96cf-4678-9724-47c608963784","arxiv_id":"2412.03407","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Skel3D conditions a Free3D-style diffusion model on target-view skeleton images, showing small but significant metric gains on Objaverse, yet the evaluation relies on oracle skeletons rather than skeletons estimated from the input.","lead":"This paper introduces Skel3D, a diffusion-based method that uses an object's skeleton image as conditioning to synthesize novel views from a single input view. It reports moderate metric gains over the Free3D baseline, but only when the exact skeleton of the target view is provided as ground truth, not when it must be inferred from the single image.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported NVS evaluation appears to use animation frames rather than camera viewpoints, and the target skeleton is an oracle; either way the monocular NVS claim is unsupported.","rationale":"The reader's verdict of REJECT is correct, but the more fundamental problem may be that the task itself is not novel view synthesis. The paper's data pipeline in Section 3.2 renders animation frames of objects with bone structures, without any description of camera motion or viewpoint sampling, and the main model variant drops the ray-conditioning layer that encodes camera information. This makes it plausible that source-target pairs are different animation frames under a fixed camera, so the model is doing skeleton-conditioned pose synthesis rather than view synthesis. The reader's weakest assumption, the oracle target skeleton, is also valid: even if camera views are varied, the test-time conditioning uses ground-truth skeletons rendered from the Objaverse 3D models, which a monocular system cannot access. The paper is transparent about this limitation in Section 6, mentioning future integration of 2D-3D lifting models, but the abstract claims a monocular open-set NVS result that the experiments do not support. The statistical testing with Mann-Whitney U is a strength, and the architecture itself is internally coherent; the problem is the evaluation-to-claim gap. Both identified issues independently justify rejection: if the frames are not views, the whole evaluation is invalid; if they are views, the oracle skeleton still invalidates the monocular claim. Therefore the reader's verdict stands unchanged.","tokens_in":9556,"tokens_out":7934,"duration_ms":80104,"concrete_test":"Audit the test-set generation script or Blender scene files for the 260-object split: for a sample of objects, list the camera matrix associated with each rendered frame and check whether source-target pairs differ by camera viewpoint or only by animation frame. If the camera is fixed across pairs, the paper has not evaluated novel view synthesis. If cameras do vary, rerun Table 1 with target skeletons estimated from the source image via a category-agnostic 3D skeleton-lifting model instead of rendered from the 3D model; if the estimated-skeleton metrics do not beat Free3D on L1, SSIM, PSNR, and LPIPS, the monocular claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing issue is that the experiments may not test novel view synthesis at all. Section 3.2 states: \"we render every fourth frame of the first 24 animation frames, providing a diverse set of poses.\" No camera pose sampling or camera motion is described anywhere in the data pipeline, and the main Skel3D variant replaces all Ray Conditioning Normalization layers with Skeleton Conditioning Normalization (Section 3.3), removing the camera-conditioning mechanism used by Free3D. The model is conditioned only on the source image and the target skeleton image (Eq. 4). Thus the target skeleton determines the output; if target and source are different animation frames, the task is pose-conditioned frame synthesis, not viewpoint change. The paper calls the frames \"views\" but provides no evidence of actual camera movement. If this reading is correct, the quantitative gains in Table 1 measure the model's ability to copy a pose from the skeleton oracle, not to synthesize novel views. Even granting that cameras were varied, the skeleton supplied at test time is the ground-truth skeleton of the target view rendered from the 3D model (Section 3.2), which is unavailable in the monocular setting. The paper's own Figure 4 shows gains depend on skeleton quality, and Figure 5 shows worse skeletons can hurt, so replacing the oracle with an estimated skeleton is likely to reduce or reverse the reported improvements.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Skel3D, a diffusion-based approach to single-view novel view synthesis conditioned on object skeletons. Building on the Free3D architecture, it replaces ray conditioning normalization (RCN) with a skeleton conditioning normalization (SCN) layer and trains on a curated subset of Objaverse animated objects with bone structures. The authors report quantitative improvements over Free3D on 260 held-out objects across L1, SSIM, PSNR, LPIPS, and FID, and claim that skeleton guidance improves pose accuracy and multi-view consistency for monocular open-set NVS without explicit 3D representations.","tokens_in":9844,"tokens_out":5662,"duration_ms":53213,"significance":"If the claimed results held in a true monocular setting, injecting a category-agnostic 2D skeleton image as conditioning into a diffusion NVS model would be a practically useful idea, particularly because it avoids fixed keypoint definitions and could in principle work with any 2D pose estimator. The architectural modification is clearly described, and the use of a non-parametric significance test is a positive step. However, the significance is conditional on an evaluation protocol that the paper does not provide: the target-view skeletons are oracle renderings from the 3D model, and the test frames appear to be animation frames rather than camera viewpoints. As it stands, the contribution is closer to oracle-skeleton-conditioned frame synthesis than to monocular novel view synthesis, so the main claimed advance is not demonstrated.","major_comments":[{"comment":"The test-time conditioning signal S in Eq. (4) is the ground-truth skeleton of the target view, rendered directly from the Objaverse 3D model as described in Section 3.2. In the claimed monocular setting this skeleton is not available; it would have to be estimated from the source image. The paper does not test this, and Section 6 explicitly defers \"integrating 2D-3D skeleton lifting models\" to future work. Consequently, Table 1 compares Skel3D with oracle pose information against Free3D without it, so the reported gains are an upper bound and do not support the abstract's monocular NVS claim.","section":"Section 3.2, Eq. (4), Section 6"},{"comment":"The data pipeline renders \"every fourth frame of the first 24 animation frames\" and no camera motion or viewpoint sampling is described anywhere. Moreover, Skel3D replaces all Ray Conditioning Normalization layers with Skeleton Conditioning Normalization (Section 3.3), so the model no longer receives camera rays and is conditioned only on the source image and the target skeleton (Eq. 4). Under the described pipeline, the target skeleton encodes an animation frame, not a camera viewpoint. The paper never operationalizes what is meant by a \"view\" in terms of camera pose, so Table 1 appears to measure pose-conditioned frame synthesis rather than novel view synthesis.","section":"Section 3.2, Section 3.3, Table 1"},{"comment":"The comparison set consists of a single baseline, Free3D, while the abstract claims to outperform \"existing state-of-the-art NVS techniques\"; no results for Zero-1-to-3 or other recent NVS methods are reported. In addition, the paper states that fine-tuning the original Free3D architecture on the curated training set decreased performance, but the numbers for that fine-tuned baseline are not reported. This missing baseline matters because Skel3D is fine-tuned on the curated Objaverse subset while the reported Free3D result is the original pretrained model, so the reader cannot separate the effect of skeleton conditioning from the effect of additional fine-tuning on the training distribution.","section":"Section 5.1, Table 1"},{"comment":"The paper repeatedly claims improvements in \"pose accuracy\" and \"multi-view consistency,\" but Section 5 lists only image-similarity metrics (L1, SSIM, PSNR, LPIPS, FID); no pose error metric and no consistency metric between generated views is reported. Furthermore, the authors' own analysis shows that when the skeleton fits the object poorly, Skel3D performs worse than Free3D (Figure 5) and that average improvement scales with skeleton-object IoU (Figure 4). Since all test skeletons are ground-truth renderings, the reported average improvements are conditional on perfect skeleton quality, and no experiment with an estimated skeleton is presented.","section":"Section 5, Section 6, Figures 4 and 5"}],"minor_comments":[{"comment":"The abstract says the method works \"without relying on explicit 3D representations,\" but Section 3.2 uses the 3D bone structures from Objaverse, projects them into 2D skeleton images, and uses those as conditioning; this wording should be clarified.","section":"Abstract, Section 3.2"},{"comment":"Figure 1's caption says \"predicted skeleton,\" but no skeleton prediction module is described in the paper; the skeleton is obtained from the 3D object or would need an external estimator. The caption should say \"provided skeleton\" or an estimation pipeline should be described.","section":"Figure 1"},{"comment":"The FID-Score is reported without a standard deviation and without details on how the FID is computed over the test set (number of generated samples, reference distribution, and whether the same random seeds are used); this should be specified.","section":"Section 5.1, Table 1"},{"comment":"The Mann-Whitney U test appears to treat each image as an independent sample, but images from the same object are likely correlated; a paired or per-object analysis would be more appropriate.","section":"Section 5.1, Table 2"},{"comment":"The paper mentions alternative architectures and their results in the supplementary material, but the arXiv version does not include a supplementary document; this material should be provided or the reference removed.","section":"Section 6"},{"comment":"Training details such as learning rate, optimizer, noise schedule, and exact number of optimization steps are not reported; these details are needed for reproducibility.","section":"Section 3.5 and Section 4"}],"recommendation":"reject","confidential_remarks":"The core problem is that the experiments do not test the claimed setting: target-view skeletons are oracle renderings from the 3D model, and the data appear to be animation frames rather than camera viewpoints. This is not a local fix; it would require collecting new data with actual camera motion, adding estimated-skeleton experiments, expanding the baseline set, and adding pose/consistency metrics. As is, the paper may be more accurately framed as oracle-skeleton-conditioned animation frame synthesis, which is a different contribution from the monocular NVS claimed in the abstract. I would not encourage a minor revision; the authors would need to substantially redesign the evaluation to address the central claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The skeleton-conditioning idea is simple and worth a few minutes: replace the ray conditioning in a Free3D-style diffusion model with a skeleton conditioning layer, feeding in a rendered skeleton image of the target view. The experiments, however, do not support the paper's central claim about monocular novel view synthesis. The strongest reading of the text is that the 'views' are actually animation frames rendered with a fixed camera — no camera trajectory is ever described — and the target skeleton is ground truth obtained from the 3D model. Under that reading, the task is pose-conditioned image synthesis, not NVS, and the improvements in Table 1 measure how well the model copies a pose from an oracle rather than any real viewpoint generalization.\n\nGive credit where it's due. The SCN layer is cleanly specified (modulated group norm with MLP from skeleton embeddings), and the choice to represent skeletons as images is a smart way around fixed keypoint counts (it's why ControlNet-style guides can work for arbitrary objects). The curated 260-object test set is deliberately disjoint from Free3D's training data, and they run Mann-Whitney U tests instead of just reporting means. Figure 4's skeleton-quality correlation and Figure 5's failure cases are genuinely informative. These are careful experimentalists.\n\nThe soft spots, in proportion. (1) The NVS framing is load-bearing and not demonstrated. Section 3.2 says they render every fourth frame of the first 24 animation frames, with no camera motion described. Worse, the model replaces all RCN layers with SCN layers, so the network has no ray or camera input at all. If the camera did not move, the 'learned views' are just different animation frames at the same camera pose. (2) The target skeleton at test time is ground truth rendered from the 3D model, exactly the explicit 3D structure the abstract claims not to use. The limitations paragraph in Section 6 does flag this, but there is no estimation experiment, despite citing 3D-LFM as a path. (3) Minor: the only baseline is original Free3D; the fine-tuned Free3D is mentioned as worse but not shown; Zero-1-to-3 is discussed but not evaluated. Effect sizes are small (about +1.1 dB PSNR), although the nonparametric tests argue they are real.\n\nThe stress-test note holds up on reading. The data-pipeline language is consistent with animation-frame rendering, and the RCN-to-SCN replacement is consistent with the model having no camera information.\n\nWho this is for: someone working on skeleton-guided pose-conditioned generation might borrow the SCN design and the skeleton-as-image principle. As an NVS contribution, it needs reframing and re-evaluation with estimated skeletons plus a camera-motion baseline before it can claim any SOTA status. I would not cite it in the next 12 months.\n\nRecommendation for peer review: send it to reviewers, but with an explicit request that they pin down the data pipeline. If the camera did not move, the paper needs a new title and abstract, and a practical-skeleton experiment must be added. The idea is real; the framing is not.","headline":"Reasonable skeleton-conditioning idea, but the experiments appear to test pose-conditioned frame synthesis with an oracle target skeleton, not monocular novel view synthesis.","tokens_in":10349,"tokens_out":4290,"would_cite":false,"duration_ms":40680,"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":"Skel3D claims that conditioning a diffusion-based novel view synthesizer on the target object's 2D skeleton improves pose accuracy and multi-view consistency, outperforming the Free3D baseline on Objaverse objects.","keywords":["novel view synthesis","diffusion model","skeleton conditioning","Objaverse","pose accuracy","multi-view consistency","adaptive normalization","single-view 3D"],"falsifier":"Run the same 260-object evaluation with target-view skeletons produced by an automatic 2D-to-3D skeleton estimator instead of rendered ground-truth bones; if the metric improvements over Free3D shrink below statistical significance or reverse, the method's advantage depends on oracle skeleton access rather than on the conditioning mechanism itself.","tokens_in":9364,"feed_emoji":"🦴","tokens_out":3742,"duration_ms":30433,"temperature":0.7,"pith_summary":"This paper claims that feeding a diffusion-based novel view synthesizer an explicit 2D skeleton of the target object, rather than relying only on camera pose, makes generated views more pose-accurate and consistent across viewpoints. It builds on Free3D and replaces its ray-conditioning normalization with a skeleton-conditioning normalization, training on a curated subset of Objaverse objects that have bone animations. On a 260-object held-out test set, the skeleton-guided model improves L1, SSIM, PSNR, LPIPS, and FID relative to the Free3D baseline, and the gains are statistically significant. The caveat is that the evaluation supplies the skeleton of the target view directly from the ground-truth 3D model, so the paper does not test the monocular setting in which the skeleton must first be estimated.","feed_headline":"Skeleton conditioning beats Free3D on novel view synthesis","feed_subtitle":"Adding the target object's 2D skeleton to a Stable Diffusion UNet improves pose accuracy and view consistency.","key_machinery":"The central mechanism is the Skeleton Conditioning Normalization (SCN) layer, an adaptive normalization inserted after every sub-module of the diffusion UNet. It applies group normalization to each activation, then modulates it with scale and shift parameters produced by an MLP from the encoded skeleton-image embedding, replacing Free3D's ray-conditioning normalization. This lets pre-trained image-generation features be steered by explicit structural cues without retraining the whole network.","core_discovery":"Skel3D establishes that explicit skeletal structure can serve as an effective conditioning signal for single-image novel view synthesis. The method intakes a source image and a target-view skeleton image, encodes both with the Stable Diffusion image encoder, and modulates every sub-module of the UNet through a Skeleton Conditioning Normalization (SCN) layer, replacing Free3D's Ray Conditioning Normalization. Training on roughly 12,000 animated Objaverse objects with rendered bone overlays, the model outperforms the Free3D baseline across all five evaluated metrics on a 260-object test set, with Mann-Whitney U tests showing p<.01 for the main comparisons. The paper further shows a monotone relationship between skeleton-to-object bounding-box IoU and the size of the improvement, indicating that the benefit is driven by how well the skeleton fits the object.","pith_inferences":["In a genuine single-image deployment, the target-view skeleton would come from an estimator rather than the ground-truth model; the paper's Figure 4 suggests that imperfect skeletons, measured by low IoU, erase or reverse the gains, so the practical value hinges on unseen estimation error.","The skeleton-image representation is a natural interface for category-agnostic 3D skeleton lifters such as 3D-LFM, which the paper points to; a testable next step is pairing that estimator with Skel3D and measuring end-to-end performance versus Free3D.","The same conditioning could be extended to video: Objaverse animations provide temporal skeleton sequences, so the method may serve as a base for temporally consistent synthesis from a single frame.","Since skeleton quality directly modulates improvement, a promising research direction is learning a skeleton-refinement module inside the diffusion loop, which the paper does not explore."],"forward_implications":["Replacing ray conditioning with skeleton conditioning improves four pixel- and perception-level metrics on the Objaverse test set, with statistically significant differences at p<.01.","Keeping both skeleton and ray conditioning (Skel3D+RCN) gives the best average metrics, though the only significant gain over skeleton-only is in LPIPS at p≈0.044.","The method generalizes across diverse non-anthropomorphic object categories, not just human-like shapes.","Because skeletons are represented as images, the approach can accept skeletons from any current or future pose estimator, avoiding a fixed keypoint format.","The dependency of gains on skeleton quality shows that the method's advantage is bounded by how well the provided skeleton matches the object's true structure."],"supporting_citations":[{"why":"The Free3D baseline architecture and the Ray Conditioning Normalization layer that Skel3D replaces and extends.","marker":"[5]"},{"why":"Zero-1-to-3 provides the diffusion-based single-view NVS framework and training objective that Skel3D inherits.","marker":"[6]"},{"why":"The latent diffusion model and UNet backbone that Skel3D fine-tunes and modulates with skeleton conditioning.","marker":"[1]"},{"why":"Objaverse is the source dataset whose animated objects carry the bone structures used for skeleton rendering.","marker":"[13]"},{"why":"Diffusion4D supplies the curated 12K high-quality animated subset that Skel3D uses for training and filtering.","marker":"[10]"},{"why":"ControlNet demonstrates skeleton-conditioned image generation and motivates using skeletons as structural guides.","marker":"[15]"},{"why":"3D-LFM is cited as a potential method for lifting 2D skeletons to 3D, which would enable operation without ground-truth bones.","marker":"[18]"}],"fun_headline_variants":["Skeleton guide beats Free3D in view synthesis","Adding skeleton cues boosts novel view synthesis","Skel3D: Skeleton-aware NVS without 3D models","Skeleton conditioning sharpens multi-view consistency","Pose accuracy improved by skeleton-guided diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes the ground-truth skeleton of the target view is available at test time; in the claimed monocular setting that skeleton must be estimated from the source image, a step the paper does not test.","fun_headline_variants_meta":{"raw":{"variants":["Skeleton guide beats Free3D in view synthesis","Adding skeleton cues boosts novel view synthesis","Skel3D: Skeleton-aware NVS without 3D models","Skeleton conditioning sharpens multi-view consistency","Pose accuracy improved by skeleton-guided diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000262,"raw_usage":{"total_tokens":1546,"prompt_tokens":845,"completion_tokens":701,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":627}},"tokens_in":461,"tokens_out":701,"duration_ms":6694,"temperature":1.0,"reasoning_tokens":627,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:25:47.428356+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 260-object evaluation with target-view skeletons produced by an automatic 2D-to-3D skeleton estimator instead of rendered ground-truth bones; if the metric improvements over Free3D shrink below statistical significance or reverse, the method's advantage depends on oracle skeleton access rather than on the conditioning mechanism itself.","supporting_citations":[{"cited_title":"Free3d: Con- sistent novel view synthesis without 3d representa- tion","cited_arxiv_id":null,"evidence_quote":"The Free3D baseline architecture and the Ray Conditioning Normalization layer that Skel3D replaces and extends."},{"cited_title":"Zero-1-to-3: Zero-shot one image to 3d object","cited_arxiv_id":null,"evidence_quote":"Zero-1-to-3 provides the diffusion-based single-view NVS framework and training objective that Skel3D inherits."},{"cited_title":"High- resolution image synthesis with latent diffusion mod- els","cited_arxiv_id":null,"evidence_quote":"The latent diffusion model and UNet backbone that Skel3D fine-tunes and modulates with skeleton conditioning."},{"cited_title":"3d- lfm: Lifting foundation model","cited_arxiv_id":null,"evidence_quote":"3D-LFM is cited as a potential method for lifting 2D skeletons to 3D, which would enable operation without ground-truth bones."}],"review_version":1}