{"id":"efc2e42f-1fed-4a80-82bb-6b255a3861f9","arxiv_id":"2507.10347","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Picard iteration is ported to SO(3) diffusion sampling, yielding up to 4.9x faster pose generation on SYMSOL without retraining the model.","lead":"This paper adapts a parallel sampling technique for diffusion models to the SO(3) manifold used in 6D object pose estimation. The authors report up to 4.9x lower latency with no measured loss in minimum angular distance on SYMSOL.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The no-degradation claim is not yet established: Eq. (9)'s log-linearization is first-order wrong on SO(3), and Table 1's quality comparison has no error bars or significance test.","rationale":"The speedup measurement is concrete and the 4.9x number is plausible, given the batched Picard structure. I did not find a reason to doubt the wall-clock timings. The fragile part is quality preservation. The reader's BCH concern is correct: for a general clean rotation X0, the commutator term in Eq. (9) is first order in the noise, not negligible. But a wrong derivation does not by itself refute the empirical claim; the algorithm could still be a useful heuristic. The decisive gap is that the reported evaluation cannot support 'no measurable degradation': single scalar values, no variance, no test, and an improvement over baseline that is surprising. A repeated-seed distributional benchmark would settle whether the parallel sampler actually preserves the target distribution. Since the speedup evidence is solid and the quality issue is testable, the conditional verdict is appropriate, and I would not move it.","tokens_in":7118,"tokens_out":18056,"duration_ms":222858,"concrete_test":"Use a fixed pretrained score network on the same SYMSOL images; run the sequential geodesic random walk and Algorithm 1 at T=100 for 10 independent sampling seeds, each generating 2,000 rotations. Report mean +/- std of the minimum angular distance, a paired 95% CI for the parallel-minus-sequential difference, and a rotation-space distributional distance (e.g., MMD or histogram divergence) between the two sample sets. Pre-register a non-inferiority margin, say 0.1 deg; if the CI exceeds the margin or the distributional distance is significantly larger than the sequential self-distance, the 'no measurable degradation' claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires the parallel sampler to produce pose distributions equivalent to the sequential sampler. The derivation intended to guarantee this is invalid: Eq. (9) replaces the relative log Log(X0^{-1}Xi) with the difference of absolute logs Log(Xi)-Log(X0). Writing y0=Log(X0) and delta=Log(Xi)-Log(X0), BCH gives Log(Exp(-y0)Exp(y0+delta)) = delta + (1/2)[-y0,delta] + ...; for general X0, y0=O(1) and delta=O(sigma), so the omitted commutator is O(sigma), the same order as the noise. Hence Eqs. (10)-(12) do not follow from the SO(3) perturbation kernel Eq. (8). The algorithm may still be a viable right-invariant Euler/Picard heuristic, but the paper does not prove it solves the score model's probability-flow ODE. The empirical support for 'no measurable degradation' is also thin: Table 1 reports one minimum angular distance per configuration, with no error bars, repetitions, or significance test, and the parallel value (1.40) is lower than the sequential baseline (1.48), which is more consistent with noise than with a faithful parallelization. So the no-degradation part of the central claim is currently unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a parallel-in-time sampling algorithm for score-based diffusion models on SO(3), adapting the Picard-iteration approach of ParaDiGMS to the rotation manifold. The authors derive a Lie-algebra SDE approximation from the wrapped Gaussian perturbation kernel, convert it to a probability-flow ODE, and obtain the exponential-map update in Eq. (12); Algorithm 1 executes this update over batch windows. Experiments on SYMSOL with LiePoseDiffusion report up to 4.9x wall-clock speedup at 100 denoising steps and comparable minimum angular distances, plus qualitative samples of symmetric pose distributions.","tokens_in":7440,"tokens_out":6192,"duration_ms":70514,"significance":"If the theoretical derivation were correct, the paper would be a useful extension of parallel ODE sampling to SO(3), with a clear practical benefit: no retraining and reduced latency for pose-diffusion models. The reported speedups are concrete and the qualitative samples are encouraging. However, the derivation rests on an approximation that is not generally valid, and the no-degradation claim is not statistically supported, so the central guarantee is not yet established. With a corrected or reframed derivation and stronger empirical validation, the method could be a solid contribution.","major_comments":[{"comment":"Eq. (9) is not valid for general non-commuting rotations. With y0 = Log(X0), delta = Log(Xi) - Log(X0), the left side is Log(Exp(-y0) Exp(y0 + delta)); the Baker-Campbell-Hausdorff expansion gives delta + (1/2)[-y0, delta] + ..., so the omitted commutator is O(sigma) while delta is also O(sigma). Since y0 is O(1) in general, the relative error is first-order and does not vanish in the small-noise limit. Consequently Eqs. (10)–(12) do not follow from the perturbation kernel in Eq. (8), and the claim that the algorithm solves the probability-flow ODE of the SO(3) diffusion model is not established.","section":"§3, Eq. (9)–(12)"},{"comment":"Eq. (10) writes g(t) = sqrt(sigma^2(t)/dt), which is dimensionally inconsistent: a differential dt cannot appear inside the coefficient. The variance-exploding SDE should use g(t) = sqrt(d sigma^2/dt) (or an explicit discretization); as written this equation is not a well-defined SDE and cannot justify the subsequent ODE.","section":"§3, Eq. (10)"},{"comment":"Table 1 reports a single minimum angular distance per configuration, with no error bars, no number of runs, and no significance test. The parallel value 1.40 at 100 steps is below the sequential 1.48, which is more consistent with run-to-run noise than with 'no measurable degradation.' The authors should report means and standard deviations over repeated trials and, ideally, distributional metrics (e.g., coverage of the symmetric pose modes) before claiming no degradation.","section":"§4.2, Table 1 (right)"},{"comment":"Algorithm 1, line 5 does not match Eq. (12): the equation uses the initial iterate X^k_0 and a product from i=0 to t-1, whereas the pseudocode uses X^k_t and a product from i=t to t+j. This is either a typo or an additional algorithmic variant (a windowed Picard update); either way it must be stated explicitly, since the derivation of Eq. (12) is what justifies the algorithm.","section":"§3, Algorithm 1 vs. Eq. (12)"}],"minor_comments":[{"comment":"The paper uses 'we proposed' in the abstract and introduction; use 'we propose' to describe the current contribution.","section":"Abstract and Introduction"},{"comment":"Line 12 returns 'xk T' using lowercase x, while the update variables are capital X; this typo should be fixed.","section":"Algorithm 1, line 12"},{"comment":"Eq. (3) writes the initialization condition xt ~ N(0,I) next to the probability-flow ODE; on SO(3) the initial condition should be specified on the manifold (e.g., X_T from the prior), not as a Euclidean Gaussian.","section":"§2.2, Eq. (3)"},{"comment":"The paper uses 'minimum angular distance' without defining whether it is the minimum over sampled modes, the geodesic distance to the nearest symmetric pose, or something else; this should be clarified.","section":"§4.1, Evaluation Method"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the derivation issue in §3 is load-bearing and cannot be ignored, but the practical speedup is plausible and the method may be salvageable. A revision that either proves the approximation under a clearly stated regime or explicitly reframes the method as a heuristic, together with proper statistical support for the no-degradation claim, would be worth considering. The novelty relative to ParaDiGMS is incremental but could be sufficient for the journal if the central claims are made rigorous."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read of 2507.10347. The headline: the speedup is probably real, but the paper's own derivation for why it should work does not hold together. The authors adapt ParaDiGMS's Picard iteration to diffusion sampling on SO(3) and report up to 4.9x latency reduction on the SYMSOL pose ambiguity task. The wall-clock numbers are concrete and the parallelization idea transfers cleanly: the MLP head dominates forward time and can be evaluated across timesteps in parallel. That is a genuinely useful contribution and, as far as I know, the first Picard-iteration sampler on SO(3).\n\nWhat is new: the right-invariant update in Eq. (12) and the adaptive batch-window scheme in Algorithm 1. The experiments include multi-sample inference and an ablation on window size. That is solid engineering.\n\nThe soft spot is Section 3. Eq. (9) replaces Log(X0^{-1} Xi) by Log(Xi) - Log(X0). For rotations that do not commute this is not a small error. Baker-Campbell-Hausdorff gives Log(Exp(-y0) Exp(y0+δ)) = δ + 1/2[-y0,δ] + ...; if y0 is O(1) and δ is O(σ), the commutator term is O(σ), the same order as the noise. So the VE-SDE in Eq. (10) and the Picard update in Eqs. (11)-(12) are not consequences of the SO(3) kernel in Eq. (8). The derivation is load-bearing because it is the stated justification for why the parallel sampler should track the sequential distribution. The algorithm might still work as a heuristic--the empirical results suggest it does--but the paper needs to either fix this derivation or reposition the method as an approximate parallel sampler whose accuracy is established empirically.\n\nThere are also smaller issues. Eq. (10) writes sqrt(sigma^2(t)/dt), which is dimensionally odd; probably should be sqrt(d sigma^2/dt). And Table 1 reports single minimum angular distance numbers with no error bars or repetitions. The parallel 1.40 vs sequential 1.48 is more consistent with noise than with 'no measurable degradation.' The 4.9x speedup claim itself is supported by the timing data, so I am not worried about that.\n\nOverall: the paper deserves serious referee attention because the adaptation is new and the speedup is practically relevant. But it needs major revision before acceptance: correct or reframe the theory, and add repeated trials with confidence intervals. I would send it to review.","headline":"Useful engineering adaptation of ParaDiGMS to SO(3) diffusion with a believable speedup, but the theoretical derivation in Section 3 has a load-bearing approximation that is not generally valid.","tokens_in":7866,"tokens_out":3854,"would_cite":false,"duration_ms":41813,"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":"Picard iteration on the rotation group SO(3) turns the sequential denoising loop of a pose-diffusion model into parallel time steps, yielding up to a 4.9x speedup with no measured accuracy loss.","keywords":["diffusion models","SO(3) manifold","Picard iteration","parallel sampling","score-based generative modeling","6D pose estimation","probability flow ODE","Lie algebra"],"falsifier":"Compare the exact quantity $\\operatorname{Log}(X_0^{-1}X_i)$ with $\\operatorname{Log}(X_i)-\\operatorname{Log}(X_0)$ on actual denoising trajectories of a trained SO(3) score model; if the typical angular distance between $X_0$ and $X_i$ is not small (say, above a few tenths of a radian) at the noise levels where the model operates, then the discrepancy between the two sides is first-order in the noise and the derived VE-SDE is not the limit of the true perturbation kernel, so the parallel samples solve a different diffusion process.","tokens_in":6952,"feed_emoji":"⚡","tokens_out":10771,"duration_ms":102718,"temperature":0.7,"pith_summary":"The paper sets out to accelerate diffusion-model sampling on the rotation group SO(3), where the denoising loop is inherently sequential. It adapts Picard fixed-point iteration to the Lie algebra of SO(3), rewriting the probability-flow ODE so that, within a fixed-point iteration, updates at different time steps are independent and can run in parallel. On the SYMSOL pose-ambiguity dataset, the paper reports up to 4.9x lower single-sample latency at 100 denoising steps, and 2.9x at 50 steps, with a minimum angular distance of 1.40 degrees versus 1.48 for the sequential baseline at 100 steps. The significance is that latency is cut by restructuring the computation graph, not by retraining the model or by reducing the number of denoising steps.","feed_headline":"4.9x faster pose-diffusion sampling on SO(3)","feed_subtitle":"A parallel fixed-point loop cuts single-sample latency without retraining or measurable pose-accuracy loss.","key_machinery":"The key object is the logarithmic map $\\operatorname{Log}: SO(3) \\to \\mathfrak{so}(3)$, which maps rotations to their Lie-algebra vectors. The derivation relies on the approximation $\\operatorname{Log}(X_0^{-1}X_i) \\approx \\operatorname{Log}(X_i) - \\operatorname{Log}(X_0)$, turning the manifold perturbation kernel into a Gaussian in the Lie algebra and yielding the variance-exploding SDE. The parallelization itself is carried by the Picard update in Eq. (12), a product of exponentials of score-function evaluations at each time step; because the product at iteration $k+1$ is built only from scores evaluated at iteration $k$, all time steps inside one iteration can be computed in parallel. The algorithm's adaptive stride and batch window then decide how many steps are advanced per iteration.","core_discovery":"The paper's central claim is that the SO(3) perturbation kernel used in score-based pose diffusion can be represented in the Lie algebra, where it matches the Euclidean variance-exploding SDE, and that the corresponding probability-flow ODE can be solved by Picard iteration. The resulting update, $X_{t}^{k+1} = X_0^k \\prod_{i=0}^{t-1} \\operatorname{Exp}\\left(-\\frac{g^2(i/T)}{2T} \\nabla_y \\log p_i(\\operatorname{Log} X_i^k)\\right)$, is a fixed-point step in which every time index $t$ depends only on quantities from the previous iteration, so the forward passes of the score network across time can be evaluated in parallel. The paper reports that this scheme matches the sequential sampler's pose distribution, with no measurable degradation in minimum angular distance, while reducing single-sample generation latency by up to 4.9x without any model retraining.","pith_inferences":["The approximation in Eq. (9) is first-order in noncommutativity, so the method is best understood as a parallel sampler for the Lie-algebra SDE; on rotations with large angular separation, the samples it produces may be those of a slightly different diffusion than the sequential manifold walk.","A testable extension is to apply the same Picard structure to SE(3) or to score models on other Lie groups, where the same Exp/Log approximation would carry over but the Baker-Campbell-Hausdorff corrections differ.","The adaptive stride in Algorithm 1 creates a latency-versus-accuracy dial that the paper does not tune per object: objects with wide symmetric pose distributions may require smaller strides to stay close to the true distribution, which suggests an object-dependent scheduling rule.","Reporting per-object convergence of the Picard iteration (number of iterations K versus tolerance) would make the speedup more interpretable, since the AI index already shows that larger batch windows cost more model evaluations."],"forward_implications":["A trained SO(3) pose-diffusion model becomes 2.9-4.9x faster at inference with no retraining, which makes the acceleration usable in latency-critical settings.","The same Lie-algebra/Picard rewriting should transfer to any score-based diffusion on a matrix Lie group whose perturbation kernel has the form in Eq. (7), since the derivation uses only Exp/Log and the VE-SDE form.","The 4.9x speedup at 100 steps versus 2.9x at 50 steps means the method gives the largest gain precisely when the sequential baseline is slowest, i.e., when many denoising steps are required.","At small sample counts the parallelized loop runs above 30 fps, bringing real-time single-object pose sampling within reach on the evaluated hardware.","At 2,000 samples the method keeps a 1.3x speed advantage, so the parallel structure does not collapse at larger batch sizes, though the relative gain shrinks."],"supporting_citations":[{"why":"supplies the sequential SE(3)/SO(3) score-diffusion pose estimator, its perturbation kernel in Eq. (7), and the baseline timings and accuracy numbers.","marker":"[3]"},{"why":"provides the SDE/probability-flow-ODE framework and the variance-exploding SDE that the Lie-algebra derivation invokes.","marker":"[4]"},{"why":"defines the SMLD perturbation kernel in Euclidean space that the approximation in Eq. (9) is meant to reproduce.","marker":"[5]"},{"why":"provides the SYMSOL dataset and rotation-manifold distributions used to evaluate pose-ambiguity handling.","marker":"[7]"},{"why":"introduces Picard-iteration parallel sampling for diffusion models in Euclidean space, the structure adapted here to SO(3).","marker":"[17]"},{"why":"is the source of Picard-Lindelöf and Banach fixed-point theorems used to justify iteration convergence.","marker":"[18]"}],"fun_headline_variants":["Parallel Picard speeds SO(3) diffusion 4.9x","SO(3) diffusion parallelized: 4.9x faster, no loss","Picard loop accelerates SO(3) pose sampling 4.9x","4.9x faster SO(3) diffusion via parallel fixed-point","Fixed-point parallelization for SO(3) diffusion: 4.9x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The derivation assumes that the rotations encountered during denoising are close enough to the identity that the logarithm of a product can be replaced by the difference of the individual logarithms, an approximation that is exact only for commuting rotations and fails as rotation angles grow.","fun_headline_variants_meta":{"raw":{"variants":["Parallel Picard speeds SO(3) diffusion 4.9x","SO(3) diffusion parallelized: 4.9x faster, no loss","Picard loop accelerates SO(3) pose sampling 4.9x","4.9x faster SO(3) diffusion via parallel fixed-point","Fixed-point parallelization for SO(3) diffusion: 4.9x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000351,"raw_usage":{"total_tokens":1861,"prompt_tokens":837,"completion_tokens":1024,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":922}},"tokens_in":453,"tokens_out":1024,"duration_ms":9956,"temperature":1.0,"reasoning_tokens":922,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:32:43.952300+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the exact quantity $\\operatorname{Log}(X_0^{-1}X_i)$ with $\\operatorname{Log}(X_i)-\\operatorname{Log}(X_0)$ on actual denoising trajectories of a trained SO(3) score model; if the typical angular distance between $X_0$ and $X_i$ is not small (say, above a few tenths of a radian) at the noise levels where the model operates, then the discrepancy between the two sides is first-order in the noise and the derived VE-SDE is not the limit of the true perturbation kernel, so the parallel samples solve a different diffusion process.","supporting_citations":[{"cited_title":"Confronting ambiguity in 6D object pose estimation via score-based diffusion on SE(3),","cited_arxiv_id":null,"evidence_quote":"supplies the sequential SE(3)/SO(3) score-diffusion pose estimator, its perturbation kernel in Eq. (7), and the baseline timings and accuracy numbers."},{"cited_title":"Score-based gen- erative modeling through stochastic differential equations,","cited_arxiv_id":null,"evidence_quote":"provides the SDE/probability-flow-ODE framework and the variance-exploding SDE that the Lie-algebra derivation invokes."},{"cited_title":"Generative modeling by estimating gradients of the data distribution,","cited_arxiv_id":null,"evidence_quote":"defines the SMLD perturbation kernel in Euclidean space that the approximation in Eq. (9) is meant to reproduce."},{"cited_title":"Implicit-PDF: Non- parametric representation of probability distribu- tions on the rotation manifold,","cited_arxiv_id":null,"evidence_quote":"provides the SYMSOL dataset and rotation-manifold distributions used to evaluate pose-ambiguity handling."},{"cited_title":"Parallel sampling of diffusion models,","cited_arxiv_id":null,"evidence_quote":"introduces Picard-iteration parallel sampling for diffusion models in Euclidean space, the structure adapted here to SO(3)."},{"cited_title":"Berinde, Iterative Approximation of Fixed Points, vol","cited_arxiv_id":null,"evidence_quote":"is the source of Picard-Lindelöf and Banach fixed-point theorems used to justify iteration convergence."}],"review_version":1}