{"id":"f61e56d4-4571-4156-8a0b-39295e3e6e8a","arxiv_id":"2509.07920","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"ScoreHOI refines coarse monocular estimates of human-object pose by score-guided diffusion with contact and penetration guidance, improving contact F-Score on BEHAVE and InterCap.","lead":"From a single photo, a diffusion model is steered toward a physically consistent 3D human and object reconstruction by enforcing contact and non-penetration constraints. It reports better contact accuracy and much faster inference than optimization-based alternatives on BEHAVE and InterCap.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hyperparameters are selected on the BEHAVE test split (Sec. 4.3, Table 4) and final results are reported on that same split (Table 1), so the 9% contact F-Score gain may reflect test-set overfitting rather than a genuine method advantage.","rationale":"The reader's weakest_assumption points to the DPS-style approximation in Eq. (4). I do not find that to be the most load-bearing concern: the approximation is standard in score-guided inverse problems, and because tau=50 is a small noise level, the one-step denoised estimate x_hat_0 is likely accurate enough for gradient guidance. Even if the approximation were imperfect, the empirical gains could still hold. The decisive weakness is the evaluation protocol: Section 4.3 explicitly performs all ablations on the BEHAVE test split and Table 4 selects the final hyperparameters by best test chamfer distance, while Table 1 reports results on that same split. This makes the 9% contact F-Score gain a selected optimum rather than a prediction. Since no validation split or error bars are provided, the claimed superiority over CONTHO is not established by the reported numbers. A re-split with validation-based hyperparameter selection, plus paired confidence intervals, is a concrete check that would settle this. The paper does provide code and detailed ablations, which is creditworthy, but the central comparative claim requires re-evaluation. The reader's CONDITIONAL verdict is therefore appropriate, with the condition made explicit. My agreement is partial because the reader mentioned test-set selection in the rationale but did not identify it as the weakest assumption.","tokens_in":15510,"tokens_out":5391,"duration_ms":46102,"concrete_test":"Split the BEHAVE training sequences into a training set and a held-out validation set (e.g., 80/20), select N, tau, and Delta t on the validation set using the same chamfer-distance criterion as Table 4, then evaluate the chosen configuration on the test split and recompute Table 1 (CD_human, CD_object, contact F-Score). Additionally, compute paired bootstrap 95% confidence intervals for the Ours-vs-CONTHO contact F-Score difference on the test split. If the advantage drops below the reported 9% or the confidence interval includes zero, the claimed superiority is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 states, 'We conduct all ablation studies on the BEHAVE test split as a standard benchmark for fair comparison,' and Table 4 says 'The optimal results measured by chamfer-distance are selected to establish our baseline,' choosing N=10, tau=50, Delta t=2. The final 'Ours' row in Table 1 is then computed on the same BEHAVE test split. This means the hyperparameters were tuned on the test set and the best configuration was reported on that same set. The headline contact F-Score improvement over CONTHO (0.554 to 0.609) is therefore a selected optimum, not an unbiased estimate of generalization. No validation split, standard errors, or significance tests are provided, so the magnitude of any true advantage is unknown. This is more load-bearing than the DPS approximation in Eq. (4): even if that approximation were imperfect, the method could still work empirically, but if the evaluation protocol is contaminated by test-set selection, the central claim of superiority is unsupported. The DPS concern is secondary; the experimental design is the decisive issue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ScoreHOI, a diffusion-based optimizer for single-image joint 3D human-object interaction reconstruction. An affordance-aware regressor produces initial SMPL-H body parameters and object pose; at inference, DDIM inversion maps this estimate into a noisy latent which is then denoised under score guidance augmented with physical constraints (human-object contact, object-floor contact, and penetration avoidance), with contact masks iteratively refined across N rounds. Experiments on BEHAVE and InterCap report improved chamfer distances and contact F-Score over CHORE, CONTHO, and PHOSA, along with much faster inference than Adam-based optimizers.","tokens_in":15924,"tokens_out":6918,"duration_ms":60992,"significance":"If the claims hold, the paper offers a timely and potentially useful contribution: a diffusion-based refinement loop that injects physical plausibility into human-object reconstruction without slow per-sample Adam optimization. The idea of combining a learned diffusion prior with differentiable physical guidance is well motivated, and the proposed contact-driven iterative refinement is a sensible mechanism for stabilizing contact prediction. Strengths include the clear system design, the efficiency comparison (Table 2), and the ablation study (Table 3) that decomposes contributions of the diffusion module, conditions, and guidance terms. However, the current evaluation protocol contains a load-bearing test-set selection issue and omits key hyperparameters, and the headline contact F-Score gain is partly a direct consequence of the optimized loss; these problems prevent the state-of-the-art claim from being accepted without revision.","major_comments":[{"comment":"The optimization hyperparameters N, tau, and Delta t in Table 4 are selected on the BEHAVE test split, and the final 'Ours' row in Table 1 is reported on that same split. Section 4.3 states that 'the optimal results measured by chamfer-distance are selected to establish our baseline,' making the reported 9% contact F-Score improvement a selected optimum rather than an unbiased estimate of generalization. No validation split, error bars, or significance tests are provided. Please re-evaluate using a held-out validation split (or nested cross-validation) and report standard errors or confidence intervals for the main comparisons.","section":"Section 4.3, Table 4 and Table 1"},{"comment":"The physical guidance weights lambda_ho, lambda_of, lambda_pt in Eq. (5) and the guidance scale rho in Eq. (6) are never reported. These values directly control the strength of the contact, floor, and penetration losses and therefore the results in Tables 1 and 3, but the paper only lists N, tau, and Delta t in Section 4.3. Without these numbers the inference procedure is not reproducible, and the ablation study cannot be quantitatively interpreted. Please provide the exact values and a sensitivity analysis.","section":"Eqs. (5)-(6), Sections 3.6 and 4.3"},{"comment":"The evaluation's contact F-Score is computed by classifying human vertices within 5 cm of the object mesh, while the loss L_ho directly minimizes the Euclidean distance between human and object vertices at the predicted contact regions. Consequently, part of the reported 9% contact F-Score gain is a direct and expected consequence of the optimization objective, not an independent measure of improved physical plausibility. For a fairer assessment, report additional metrics that are not so directly optimized (e.g., chamfer distance alone, or contact F-Score evaluated after excluding vertices used in the contact loss), or otherwise quantify how much of the contact gain is attributable to the diffusion prior versus the explicit loss.","section":"Eq. (5) and Section 4.1 (contact F-Score)"},{"comment":"The Table 1 comparison between ScoreHOI and CONTHO is not fully apples-to-apples because ScoreHOI uses a new affordance-aware regressor with Hand4Whole initialization and PointNeXt object features, whereas CONTHO uses its own regressor. The ablation 'w/o diffusion' in Table 3 shows the diffusion component helps within the authors' pipeline, but it does not establish that the reported improvement over CONTHO comes from the proposed optimizer rather than from the improved initial regressor. A head-to-head comparison that refines the same initial estimates with Adam and with ScoreHOI would directly attribute the gain to the proposed method.","section":"Section 3.2 and Table 1"}],"minor_comments":[{"comment":"The InterCap CONTHO row is missing spacing in the numerical values ('5.96 9.500.6610.432 0.522'); please fix the formatting.","section":"Table 1"},{"comment":"In the contact mask update line, 'object-to-human contact mask Mh' should be written as M_o (or M_oh) to match the notation {M_i}_{i in {h,o,f}} used elsewhere.","section":"Algorithm 1, line 5"},{"comment":"The loop header 'for t = tau to Delta t with step size Delta t' is ambiguous because the sampling proceeds from a larger tau down to Delta t; please write the loop as descending, e.g., 'for t = tau down to Delta t step -Delta t'.","section":"Supplementary Algorithm A"},{"comment":"The caption says 'Bottom row: the ablation study of L_ho,' but the accompanying text and the described effect (penetration penalty) indicate the bottom row should refer to L_pt.","section":"Figure 5 caption"},{"comment":"The 'Hand4Whole framework' is not cited; please add the appropriate reference.","section":"Section 3.6"},{"comment":"The criterion 'optimal results measured by chamfer-distance' is not fully precise: it should state whether this is the sum of CD_human and CD_object, and it should acknowledge that the selected configuration (N=10, tau=50, Delta t=2) does not maximize contact F-Score in Table 4 (the tau=100 row gives 0.615), so the selection rule should be stated unambiguously.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The test-set hyperparameter selection is the deciding issue for me: the headline improvement over CONTHO is a selected optimum, and the missing loss weights and guidance scale prevent reproduction. The paper has a solid core idea and useful ablations, and I would be willing to reconsider once a clean validation protocol, error bars, and the missing hyperparameters are provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a real, competent piece of engineering for monocular human-object reconstruction, but the headline result—a 9% contact F-Score gain over CONTHO on BEHAVE—is weaker than it looks because the hyperparameters were chosen on the BEHAVE test split and the final numbers come from that same split. The method itself is worth a serious referee, but the evaluation needs to be redone with a clean split.\n\nWhat's actually new: ScoreHOI replaces the usual Adam-based joint optimization with a DDIM inversion-and-sampling loop, conditions the diffusion model with both image features and object geometry via an IG-Adapter, and iteratively updates contact masks during refinement. The speed improvement is real and clearly shown: two orders of magnitude over CHORE/VisTracker, plus a faster N=2 variant. Code is promised. That part is solid.\n\nThe soft spots are mostly in the experimental protocol. Section 4.3 says all ablations were run on the BEHAVE test split, and Table 4 selects N=10, tau=50, Δt=2 by best chamfer distance on that same split. Table 1 then reports 'Ours' on the same test set. That is textbook test-set overfitting, and it makes the 9% contact gain a selected optimum rather than a predictive claim. The stress-test is right to call this load-bearing.\n\nSecond, L_ho in Eq. (5) directly minimizes the vertex-to-vertex distance that the contact F-Score measures (vertices within 5cm). So the contact gain is partly a consequence of optimizing the metric. That doesn't make the method useless—it makes it less surprising.\n\nThird, reproducibility gaps: the loss weights λ_ho, λ_of, λ_pt and the guidance scale ρ are never reported, and there are no error bars on any Table 1 number. The DPS-style approximation in Eq. (4) is standard but unvalidated; I'd call that minor.\n\nBottom line: this is a useful contribution to the subfield, with a clean idea and honest code release, but the central quantitative claim is not yet supported. I'd send it to peer review with the requirement that the authors rerun the ablation selection on a validation split (or at least report cross-validated selection) and report all hyperparameters and error bars.","headline":"Solid diffusion-based HOI refinement with a contaminated evaluation: test-set-tuned hyperparameters make the headline 9% contact gain unsupported as stated.","tokens_in":16293,"tokens_out":3023,"would_cite":false,"duration_ms":26105,"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":"ScoreHOI claims that a score-guided diffusion optimizer with physical constraints reconstructs human-object interactions more accurately and far faster than Adam-based optimization, improving contact F-Score by 9% on BEHAVE.","keywords":["human-object interaction reconstruction","score-guided diffusion","physical plausibility","contact refinement","DDIM inversion","3D human mesh recovery","single-image reconstruction","diffusion prior"],"falsifier":"Measure the error between the one-step denoised estimate $\\hat{x}_0(x_t)$ and the posterior mean obtained from a full DDIM sample at the same noise level ($\\tau = 0.05$) on BEHAVE frames; if the error is large in regions where hands contact objects, the DPS approximation in Eq. (4) is miscalibrated and the 9% contact F-Score gain cannot be attributed to the diffusion prior alone.","tokens_in":15341,"feed_emoji":"🧍","tokens_out":7585,"duration_ms":61766,"temperature":0.7,"pith_summary":"This paper sets out to show that a diffusion model can act as a high-quality optimizer for joint reconstruction of a human body and an interacting object from a single image. The proposed pipeline, ScoreHOI, inverts an initial regression estimate into a noisy latent, then runs DDIM sampling while steering each denoising step with physical constraints: human-object contact, object-floor contact, and penetration avoidance. A contact-driven iterative refinement loop updates the predicted contact masks between sampling rounds. The authors report that this approach outperforms previous optimization and regression methods, including a 9% relative improvement in contact F-Score on the BEHAVE benchmark, while running one to two orders of magnitude faster than Adam-based optimizers. This matters because physically plausible human-object reconstruction is seriously ill-posed from a single view, and existing optimizers either ignore image evidence or are too slow for practical use.","feed_headline":"Diffusion prior lifts human-object contact F-Score by 9 percent","feed_subtitle":"Score-guided sampling with contact and floor constraints beats Adam-based fitting and runs two orders of magnitude faster.","key_machinery":"The load-bearing mechanism is the modified noise-prediction update $\\epsilon'_{\\phi} = \\epsilon_{\\phi}(x_t, t, c) + \\rho \\sqrt{1-\\alpha_t} \\nabla_{x_t} L_P$, where $L_P$ combines human-object contact, object-floor contact, and penetration-penalty losses. To compute those gradients from a noisy latent, the method uses the DPS approximation $\\nabla_{x_t} \\log p(P|c, x_t) \\approx \\nabla_{x_t} \\log p(P|c, \\hat{x}_0(x_t))$, replacing the latent with its one-step denoised estimate $\\hat{x}_0$ and differentiating the physical losses through it. An IG-Adapter cross-attention block injects image features and object-geometry features as conditions into the diffusion model, and the contact-driven iterative refinement updates the contact masks each outer loop using a trained contact predictor.","core_discovery":"On its own terms, the paper's central discovery is that diffusion priors over human-object interaction parameters are strong enough to serve as the optimizer itself, not just a generative prior over point clouds. Starting from a coarse SMPL-H and object pose estimate, DDIM inversion maps the parameters into a noisy latent that carries the image evidence, and guided DDIM sampling with physical losses reshapes the estimate into a more accurate, physically plausible interaction. The contact masks, which determine where the human-object and object-floor penalties apply, are themselves iteratively refined during inference, and the ablations show this iterative contact update is responsible for a large share of the contact F-Score gain. The paper positions this as a controllable, efficient alternative to joint Adam-based optimization that keeps image-level features in the loop.","pith_inferences":["The same score-guided refinement loop could extend to hand-object or human-scene reconstruction, since it needs only a differentiable parametric model, a denoiser over the parameters, and contact/penetration losses.","Because the method requires a canonical template for each object and trains affordance features on known categories, its gains may not transfer to unseen object shapes; the paper itself flags this, so an object-agnostic template or per-instance shape module is the natural next step.","The DPS-style approximation in Eq. (4) is shared with other diffusion-guidance methods; annealing the noise level or using a consistency-model one-step estimator could make the physical guidance robust at higher $\\tau$ and further improve contact recall.","On InterCap the method trades precision for recall (precision drops from 0.661 to 0.627 while recall rises from 0.432 to 0.590), so applications with asymmetric error costs would need to reweight the contact objectives."],"forward_implications":["If the central claim holds, single-image human-object reconstruction can be refined without per-sample Adam iteration, cutting inference time from roughly 0.0035–0.036 frames per second to 0.29–2.0 frames per second with equal or better accuracy.","Contact quality improves: the method raises contact F-Score by 9% relative on BEHAVE and boosts contact recall on InterCap from 0.432 to 0.590, meaning predicted contact regions align far better with ground truth.","The contact-driven iterative refinement is a necessary component: removing it drops contact F-Score from 0.609 to 0.577, showing that updating contact masks during sampling is what lets the physical penalties act on the right vertices.","The diffusion prior also functions without physical guidance, since the generation tests from intermediate noise levels produce plausible interactions; the guidance terms then correct rather than create the contact pattern."],"supporting_citations":[{"why":"Supplies the contact predictor, mesh regressor, and training losses that ScoreHOI wraps its diffusion loop around, and serves as the strongest baseline it must beat.","marker":"[37]"},{"why":"CHORE is the primary optimization-based baseline; its BEHAVE numbers and Adam-based loop are the efficiency comparison point.","marker":"[59]"},{"why":"Introduces score-guided diffusion for 3D human recovery, the technique this paper extends to joint human-object parameter optimization.","marker":"[50]"},{"why":"DPS diffusion posterior sampling is the source of the approximation in Eq. (4) used to compute physical guidance gradients.","marker":"[7]"},{"why":"DDIM provides the inversion and accelerated sampling steps that map initial estimates into noisy latents and back.","marker":"[47]"},{"why":"BEHAVE is the main benchmark and the source of the reported 9% contact F-Score improvement.","marker":"[2]"},{"why":"InterCap is the second evaluation benchmark, showing the method's contact recall gains.","marker":"[23]"},{"why":"PHOSA is an earlier optimization baseline whose predefined contact heuristics the paper argues are less accurate.","marker":"[68]"},{"why":"VisTracker is the efficiency comparison baseline, showing the Adam-based optimizer's FPS that ScoreHOI outstrips.","marker":"[61]"}],"fun_headline_variants":["ScoreHOI uses diffusion priors as optimizer for HOI reconstruction","Diffusion priors replace Adam for plausible human-object pose fitting","Contact-refined diffusion reconstructs human-object interactions","Score-guided diffusion with physical constraints improves HOI pose","Diffusion optimizer yields physically plausible human-object interaction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's guidance depends on the assumption that the one-step denoised estimate $\\hat{x}_0(x_t)$ is accurate enough at the chosen noise level that physical-loss gradients computed through it pull the parameters toward the image; if that estimate is off, the guidance can push the reconstruction away from the image rather than toward it.","fun_headline_variants_meta":{"raw":{"variants":["ScoreHOI uses diffusion priors as optimizer for HOI reconstruction","Diffusion priors replace Adam for plausible human-object pose fitting","Contact-refined diffusion reconstructs human-object interactions","Score-guided diffusion with physical constraints improves HOI pose","Diffusion optimizer yields physically plausible human-object interaction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001094,"raw_usage":{"total_tokens":4531,"prompt_tokens":868,"completion_tokens":3663,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":3584}},"tokens_in":484,"tokens_out":3663,"duration_ms":20730,"temperature":1.0,"reasoning_tokens":3584,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:09:05.066139+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the error between the one-step denoised estimate $\\hat{x}_0(x_t)$ and the posterior mean obtained from a full DDIM sample at the same noise level ($\\tau = 0.05$) on BEHAVE frames; if the error is large in regions where hands contact objects, the DPS approximation in Eq. (4) is miscalibrated and the 9% contact F-Score gain cannot be attributed to the diffusion prior alone.","supporting_citations":[{"cited_title":"Joint reconstruction of 3d human and object via contact-based refinement transformer","cited_arxiv_id":null,"evidence_quote":"Supplies the contact predictor, mesh regressor, and training losses that ScoreHOI wraps its diffusion loop around, and serves as the strongest baseline it must beat."},{"cited_title":"CHORE: Contact, human and object reconstruction from a single RGB image","cited_arxiv_id":null,"evidence_quote":"CHORE is the primary optimization-based baseline; its BEHAVE numbers and Adam-based loop are the efficiency comparison point."},{"cited_title":"Score-guided diffusion for 3d human recovery","cited_arxiv_id":null,"evidence_quote":"Introduces score-guided diffusion for 3D human recovery, the technique this paper extends to joint human-object parameter optimization."},{"cited_title":"Diffusion posterior sampling for general noisy inverse problems.ICLR, 2023","cited_arxiv_id":null,"evidence_quote":"DPS diffusion posterior sampling is the source of the approximation in Eq. (4) used to compute physical guidance gradients."},{"cited_title":"Denois- ing diffusion implicit models.ICLR, 2021","cited_arxiv_id":null,"evidence_quote":"DDIM provides the inversion and accelerated sampling steps that map initial estimates into noisy latents and back."},{"cited_title":"BEHA VE: Dataset and method for tracking human object in- teractions","cited_arxiv_id":null,"evidence_quote":"BEHAVE is the main benchmark and the source of the reported 9% contact F-Score improvement."},{"cited_title":"InterCap: Joint markerless 3D tracking of humans and objects in interaction","cited_arxiv_id":null,"evidence_quote":"InterCap is the second evaluation benchmark, showing the method's contact recall gains."},{"cited_title":"Perceiving 3D human-object spatial arrangements from a single image in the wild","cited_arxiv_id":null,"evidence_quote":"PHOSA is an earlier optimization baseline whose predefined contact heuristics the paper argues are less accurate."},{"cited_title":"Visibility aware human-object interaction tracking from sin- gle rgb camera","cited_arxiv_id":null,"evidence_quote":"VisTracker is the efficiency comparison baseline, showing the Adam-based optimizer's FPS that ScoreHOI outstrips."}],"review_version":2}