{"id":"efb8a287-d421-45af-94ac-9f9921b22478","arxiv_id":"2505.03394","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"EOPose transfers the pose of an exemplar object image onto a source object image using DINO-ViT keypoint correspondences, a learned warp, and a GAN refinement stage, outperforming TPS, UFO-PT, and ControlCom on a synthetic Objaverse benchmark.","lead":"A new picture-editing pipeline, EOPose, moves an object from one image into a new pose shown by another image of a similar object, while keeping the original colors, textures, and logos intact. It combines pre-trained correspondence detection, image warping, and a refinement network, and reports better image-quality scores than three existing methods on a new synthetic dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's 'All together' row cannot be a simple aggregate of the four category rows: pooled SSIM 0.44 lies below all category SSIMs and pooled LPIPS 0.33 lies above all category LPIPSs, so the headline numbers rest on an unexplained protocol.","rationale":"The reader's weakest_assumption (keypoint reliability) is plausible: no ablation removes the Pa/Pp channels, and DINO BBP correspondences can be noisy under large pose changes. However, that is a robustness or attribution concern, not a contradiction in the reported evidence. The single load-bearing issue is the quantitative table itself. The central claim is explicitly about Table 1. The 'All together' row cannot be a weighted average of the category rows for SSIM and LPIPS, so either the table has a typo, the categories were evaluated with different models, or the aggregate was computed differently; any of these undermines the headline numbers as stated. This does not mean the method is fraudulent or useless: the per-category results and the user study (88.4% preference) support the general conclusion. Therefore the appropriate disposition is to keep the paper conditional and require a corrected, reproducible Table 1 rather than reject outright. The keypoint ablation should still be added (e.g., zero out Pa/Pp channels), but it is secondary to resolving the metric inconsistency.","tokens_in":11378,"tokens_out":7475,"duration_ms":76121,"concrete_test":"Ask the authors to release the per-image SSIM/LPIPS scores and the exact metric script for Table 1, and recompute the 'All together' row as the weighted average of the four category means, weighted by test sample count. If the reproduced pooled SSIM/LPIPS still fall outside the category ranges, the table is erroneous and must be corrected; if the category rows come from per-category models and 'All together' comes from a single multi-class model, the paper must state this explicitly and report all numbers under a single protocol. Additionally, run a paired per-sample significance test (e.g., bootstrap) on the corrected aggregate to confirm EOPose beats TPS, UFO-PT, and ControlCom.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim is Table 1. The 'All together' row is internally inconsistent with the category rows above it. SSIM and LPIPS are per-image averages; when pooling the same test samples from four disjoint categories, the pooled mean must lie between the minimum and maximum category means. For EOPose, category SSIMs are 0.89/0.86/0.82/0.77 but 'All together' is 0.44; category LPIPSs are 0.04/0.06/0.07/0.08 but 'All together' is 0.33. TPS shows the same problem (category SSIM 0.68/0.61/0.60/0.55 vs 'All together' 0.34; category LPIPS 0.28/0.35/0.37/0.27 vs 0.58). Thus the quoted comparison 'EOPose achieves SSIM 0.44, LPIPS 0.33, FID 18.22 vs next best 0.34/0.58/75.22' cannot be a pooled aggregate of the reported category results unless the categories use separate models or different evaluation protocols, which the paper does not state. This is a direct threat to the strongest claim: the headline numbers may be misreported, or the comparison may mix incompatible settings. A reviewer cannot verify superiority from this table as presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces EOPose, an end-to-end exemplar-based object reposing method. Given an appearance image and a pose-guidance image of another object from the same class, EOPose first extracts sparse correspondences using DINO-ViT descriptors with Best-Buddies Pairs and k-means clustering, then warps the appearance image with a U-Net that predicts multi-scale flow fields, and finally re-renders the warped image with a GAN-based generator that injects multi-scale texture features into pose encodings. The authors also contribute a new paired dataset rendered from Objaverse models across four object categories and report quantitative results (SSIM, LPIPS, FID), qualitative comparisons, ablations, and a user study. The central claim is that EOPose substantially outperforms TPS warping, UFO-PT, and ControlCom on this dataset.","tokens_in":11760,"tokens_out":5950,"duration_ms":58524,"significance":"If the reported results are correct, the paper would make a useful practical contribution: it addresses a clear e-commerce need, proposes a category-agnostic pose representation via pretrained visual descriptors, and includes a new dataset that could enable further work on generic object reposing. The three-stage design is reasonable, and the ablations support the importance of end-to-end fine-tuning and the number of keypoints. The paper also ships several strengths: a concrete dataset construction protocol, a user study, and comparisons with multiple adapted baselines. However, the central quantitative evidence is currently undermined by an internally inconsistent aggregated row in Table 1 and by missing supplementary material that is explicitly referenced for key loss definitions and training details. These issues must be resolved before the comparative claims can be accepted.","major_comments":[{"comment":"The 'All together' row cannot be a simple pooled mean of the four category rows for SSIM and LPIPS. For EOPose, the category SSIM values are 0.89, 0.86, 0.82, and 0.77, so any weighted average of per-image SSIM values over these disjoint groups must lie between 0.77 and 0.89; the reported 0.44 is outside this range. Likewise, the category LPIPS values 0.04, 0.06, 0.07, and 0.08 cannot average to 0.33. The same impossibility holds for the TPS row. Please state exactly how the 'All together' row was computed, whether it uses the same test samples as the category rows, and recompute all reported pooled values. As presented, the headline comparison 'EOPose achieves SSIM 0.44, LPIPS 0.33, FID 18.22 versus next-best 0.34, 0.58, 75.22' is not verifiable from the table.","section":"Section 5, Table 1"},{"comment":"Several load-bearing definitions are deferred to a supplementary document that is not included in the arXiv submission: the TPS-based flow loss L_f low and the full form of Equation (3), the training schedule and epoch counts, all loss-weight hyperparameters (beta1...beta5, alpha_l1, alpha_per, alpha_sty, alpha_adv, alpha1, alpha2), and the dataset-split details. Without these, Equations (2) and (3) are not fully specified and the method cannot be reproduced. Please either include the supplementary material with the submission or move the essential definitions into the main text.","section":"Sections 3.2, 3.3, and 3.4"},{"comment":"The paper's central novelty is the use of unsupervised DINO-ViT keypoint correspondences for class-agnostic pose transfer, but no ablation removes or corrupts the keypoint channels. The only related experiment varies the number of keypoints k (15, 25, 35, 45); this does not test whether the correspondences are necessary, because the warping network also receives both original images and could in principle learn pose alignment from the images alone. Please add an ablation in which the Pa and Pp channels are removed or replaced with random keypoints, and report the resulting metrics. This is needed to support the claim that the correspondence module is an essential component.","section":"Section 3.1 and Table 2"},{"comment":"The baseline comparison is not fully specified. The paper says UFO-PT and ControlCom are 'fine-tuned using our dataset' and that the TPS baseline is 'adapted from [39]', but it does not report the fine-tuning protocol (number of steps, learning rate, loss weighting, or model selection) or the TPS adaptation details. Without this information, the reader cannot assess whether the baselines were given a fair comparison; the reported margins may partly reflect suboptimal baseline tuning. Please provide the complete baseline training and evaluation protocol.","section":"Section 4, Baselines"}],"minor_comments":[{"comment":"The abstract and Section 4 mention PSNR as an evaluation metric, but Table 1 and the experimental section report only SSIM, LPIPS, and FID. Please either add PSNR results or remove the mention.","section":"Abstract and Section 4"},{"comment":"Equation (5) appears to have the assignment direction reversed: the text states that the output e_i_out undergoes a noise modulation operation B to produce e_i_infused, but the equation reads e_i_out = B(e_i_infused). Please correct the notation so the text and equation agree.","section":"Section 3.3, Equation (5)"},{"comment":"The summation in Equation (2) is typeset incorrectly; the limits 'l=K-1 ... l=0' are ambiguous. Please write the summation with clear lower and upper limits, such as sum_{l=0}^{K-1}.","section":"Section 3.2, Equation (2)"},{"comment":"The phrase 'try-on output' is used to describe the reposing result; for consistency with the rest of the paper, use 'reposing output'.","section":"Section 6.2"},{"comment":"The '35 keypoints' row and the 'EOPose (OURS)' row report identical numbers; please label the 35-keypoint row clearly as the full model rather than as an ablation, or explain the difference.","section":"Table 2"},{"comment":"The user study reports an 88.4% preference rate but provides no confidence interval or significance test; please report the number of comparisons per participant and the variance across participants.","section":"Section 5, User Study"}],"recommendation":"major_revision","confidential_remarks":"The inconsistency in Table 1 is serious: if the authors cannot provide a corrected, internally consistent table, the central comparative claim would not be supported. I would also ask the editor to verify that the referenced supplementary material is included in any revised submission, since several definitions essential to the method currently reside only in that missing document."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is real: exemplar-based reposing for generic objects, not just humans or garments. EOPose takes an appearance image and a pose exemplar, finds correspondences with DINO-ViT features, warps with a U-Net, and re-renders with a GAN. That three-stage combination is not in the prior work I know of, and the new Objaverse-derived dataset is a useful asset even if it is small. The ablations support the design choices: keypoint count, pose-image input, and end-to-end fine-tuning all move the metrics in the expected direction.\n\nThe problem is Table 1. The 'All together' row is not a weighted average of the category rows above it. For EOPose, category SSIMs are 0.89, 0.86, 0.82, 0.77; any pooled mean has to lie between 0.77 and 0.89, yet the row says 0.44. Category LPIPS values are 0.04, 0.06, 0.07, 0.08; the aggregate says 0.33. TPS shows the same impossibility. So the sentence \"EOPose achieves significantly better SSIM of 0.44, LPIPS of 0.33 and FID of 18.22\" is either computed on a different test set or there is a bug. The paper does not say. That is a load-bearing flaw: the headline comparison is the main evidence of superiority over the baselines.\n\nOther soft spots are more standard. The referenced supplementary material is missing from the arXiv version. There are no error bars or significance tests. The evaluation is entirely on synthetic renders; the application is e-commerce product photos, but no real images are tested. The correspondence module itself is not ablated by corrupting keypoints, so the weakest link in the pipeline is untested. On the positive side, the writing is clear, the related work is relevant, and the user study, though small, shows a strong preference.\n\nWho is this for? People working on object reposing, product image generation, and mask-free pose transfer. The paper deserves a serious referee, but it should not be accepted as-is. The authors need to explain or recompute the aggregate row, upload the supplementary, and ideally show at least a few real-image qualitative results. If the table cannot be fixed, the claims should be scaled back to per-category results only.","headline":"A clean problem framing and a plausible three-stage pipeline, but the headline numbers in Table 1 cannot be pooled averages of the per-category rows, so the main quantitative claim needs checking before this is citable.","tokens_in":12255,"tokens_out":2242,"would_cite":false,"duration_ms":22083,"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":"EOPose is a three-stage, class-agnostic pipeline that transfers the pose of a target product photo onto a source product photo using unsupervised keypoint correspondences, and it reports substantially better image-quality scores than the…","keywords":["object reposing","exemplar-based image synthesis","self-supervised keypoint correspondences","DINO-ViT dense descriptors","GAN image generation","Objaverse dataset","pose transfer","e-commerce product imaging"],"falsifier":"Randomly permute or zero the 70 Gaussian keypoint heatmap channels in the warping module's input while keeping the appearance and pose images unchanged; if the reported SSIM, LPIPS, and FID do not substantially worsen toward the TPS baseline, the keypoint correspondences are not the load-bearing component the paper claims.","tokens_in":11219,"feed_emoji":"🖼️","tokens_out":7568,"duration_ms":62752,"temperature":0.7,"pith_summary":"EOPose claims that generic object reposing — changing a product photo into a new pose — can be done without per-class pose models by using dense visual correspondences from a pretrained self-supervised vision transformer. The paper proposes to find 35 matching keypoints between the appearance image and a pose-guidance image, warp the appearance image with a learned dense flow field, then re-render the warped image with a GAN that keeps fine-grained detail such as colors, textures, and brand marks. On a new Objaverse-based dataset of rendered shoes, vases, briefcases, and file cabinets, EOPose reports SSIM 0.44, LPIPS 0.33, and FID 18.22, ahead of the compared baselines, and a user study prefers it 88.4% of the time. If correct, the method offers e-commerce a way to produce multiple product views from two ordinary photos, without 3D models or manual editing.","feed_headline":"EOPose reposes objects with two photos and beats diffusion baselines","feed_subtitle":"A DINO-ViT correspondence stage, a warping U-Net, and a re-rendering GAN transfer pose while preserving brand marks.","key_machinery":"The load-bearing object is the generalized pose correspondence: 35 ordered keypoints per image found by extracting spatial descriptors from a pre-trained DINO-ViT, keeping only mutual nearest-neighbour (Best-Buddies) pairs between the appearance image and the pose image, and clustering the descriptor set with k-means. Each keypoint is encoded as a Gaussian heatmap channel, so the warping U-Net receives a 76-channel input (two RGB images plus 35 heatmaps per image) and predicts a coarse-to-fine flow pyramid; the flow is refined with convex upsampling and used to warp the source image. The final generator fuses pose and texture embeddings through texture injection blocks (scaling and shifting pose features by texture-derived parameters, followed by noise modulation) before tRGB blocks synthesize the output. This three-step design is what, in the paper's argument, preserves exact colors and brand marks while still transferring the new pose.","core_discovery":"EOPose's central claim is that exemplar-based reposing of arbitrary object classes is achievable by replacing class-specific pose representations with unsupervised keypoint correspondences derived from DINO-ViT dense descriptors, matched by mutually-nearest-neighbour Best-Buddies Pairs and reduced to k=35 points by k-means clustering. These correspondences, rendered as Gaussian heatmap channels, guide a U-Net that predicts a pyramid of dense flow fields; the coarsest flow warps the appearance image, and intermediate flows are supervised with L1, perceptual, style, total-variation, and TPS-based losses. A StyleGAN-style generator then re-renders the warped image, injecting multi-scale texture encodings into pose embeddings through 2D style modulation and refining occluded or missing regions. The trained pipeline is evaluated on a new paired dataset built from 60 manually filtered Objaverse models rotated by random Euler angles, and the paper reports that EOPose achieves SSIM 0.44, LPIPS 0.33, and FID 18.22, substantially better than the next best values of SSIM 0.34, LPIPS 0.58, and FID 75.22 from the compared baselines.","pith_inferences":["A direct test the paper leaves open: ablate the keypoint channels; if EOPose's quality holds with random or zeroed keypoints, the claimed role of 'generalized pose correspondences' would be unsupported.","DINO-ViT correspondences were trained on natural images, and the four rendered Objaverse classes are relatively rigid, so applying EOPose to articulated or deformable objects would likely require denser or semantically finer correspondences than k=35.","Since the dataset is rendered from 3D models with controlled backgrounds, transferring the pipeline to real product photos will need extra robustness steps such as background removal, scale normalization, and camera calibration, which the paper's metrics do not measure.","The pipeline is a lightweight alternative to diffusion for controlled product-shot generation: it needs only a U-Net and a GAN, both far cheaper to run than iterative denoising, at the cost of relying on the source image to contain the appearance to preserve."],"forward_implications":["For e-commerce settings where a brand has no 3D models, two ordinary photos of a same-class object suffice to synthesize any desired pose.","Because the first two stages warp rather than regenerate, the output preserves stickers, logos, patterns, and sharp color transitions that generative and diffusion models tend to hallucinate away.","The approach generalizes across object classes without retraining pose representations, provided the class has enough DINO-ViT-correspondence structure, as demonstrated on vases, briefcases, file cabinets, and shoes.","End-to-end fine-tuning after stage-wise training is essential: it improves FID from 34.86 to 18.22, so the two-stage training scheme alone is not enough for the reported quality.","The number of correspondence points has a sweet spot: 35 points outperform 15, 25, and 45, with FID degrading from 18.22 to 21.62, 20.65, and 22.48 respectively, showing that both too few and too many points hurt the flow quality."],"supporting_citations":[{"why":"Supplies the DINO-ViT feature extraction and salience-based patch selection that the keypoint module is built on.","marker":"[2]"},{"why":"Supplies the mutual-nearest-neighbour Best-Buddies matching rule used to find correspondences between the two images.","marker":"[27]"},{"why":"Supplies the 3D object models from which the paired reposing dataset is rendered.","marker":"[8]"},{"why":"Source of the adapted TPS baseline and the TPS-based initialization loss for the flow.","marker":"[39]"},{"why":"The mask-guided pose transformation baseline (UFO-PT) that EOPose compares against.","marker":"[36]"},{"why":"The diffusion-based image-composition baseline that EOPose compares against and argues hallucinates details.","marker":"[42]"},{"why":"Supplies the 2D style modulation / texture-injection design used in the generator.","marker":"[1]"},{"why":"Supplies the convex upsampling of flow maps that refines the warping module's predicted flows.","marker":"[37]"}],"fun_headline_variants":["Two photos, one pose: EOPose preserves brand marks in reposing","Generalized object reposing via unsupervised keypoint correspondences","EOPose transfers pose with DINO-ViT keypoints, beats GAN baselines","Warp and re-render: EOPose keeps colors and textures in pose transfer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline's correctness rests on the assumption that the 35 correspondence points found by DINO-ViT and Best-Buddies Pairs stay semantically consistent across different instances and large pose changes, an assumption the paper never isolates by ablating the keypoint channels.","fun_headline_variants_meta":{"raw":{"variants":["Two photos, one pose: EOPose preserves brand marks in reposing","Generalized object reposing via unsupervised keypoint correspondences","EOPose transfers pose with DINO-ViT keypoints, beats GAN baselines","Warp and re-render: EOPose keeps colors and textures in pose transfer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1643,"prompt_tokens":996,"completion_tokens":647,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":562}},"tokens_in":612,"tokens_out":647,"duration_ms":5912,"temperature":1.0,"reasoning_tokens":562,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:52:45.739171+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Randomly permute or zero the 70 Gaussian keypoint heatmap channels in the warping module's input while keeping the appearance and pose images unchanged; if the reported SSIM, LPIPS, and FID do not substantially worsen toward the TPS baseline, the keypoint correspondences are not the load-bearing component the paper claims.","supporting_citations":[{"cited_title":"Deep vit features as dense visual descriptors","cited_arxiv_id":null,"evidence_quote":"Supplies the DINO-ViT feature extraction and salience-based patch selection that the keypoint module is built on."},{"cited_title":"Freeman, and Shai Avidan","cited_arxiv_id":null,"evidence_quote":"Supplies the mutual-nearest-neighbour Best-Buddies matching rule used to find correspondences between the two images."},{"cited_title":"Toward characteristic- preserving image-based virtual try-on network","cited_arxiv_id":null,"evidence_quote":"Source of the adapted TPS baseline and the TPS-based initialization loss for the flow."},{"cited_title":"General object pose transformation network from unpaired data","cited_arxiv_id":null,"evidence_quote":"The mask-guided pose transformation baseline (UFO-PT) that EOPose compares against."},{"cited_title":"Controlcom: Controllable im- age composition using diffusion model, 2023","cited_arxiv_id":null,"evidence_quote":"The diffusion-based image-composition baseline that EOPose compares against and argues hallucinates details."},{"cited_title":"Pose with Style: Detail- preserving pose-guided image synthesis with conditional stylegan","cited_arxiv_id":null,"evidence_quote":"Supplies the 2D style modulation / texture-injection design used in the generator."},{"cited_title":"Raft: Recurrent all-pairs field transforms for optical flow, 2020","cited_arxiv_id":null,"evidence_quote":"Supplies the convex upsampling of flow maps that refines the warping module's predicted flows."}],"review_version":1}