{"id":"788523f6-d35a-4f9a-aa91-06df95036a82","arxiv_id":"2411.12773","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"ADMMDiff decouples the diffusion prior from condition guidance with an ADMM algorithm and reports state-of-the-art results on guided image generation and controllable motion synthesis.","lead":"This paper introduces a training-free way to guide an existing diffusion model toward a desired condition, like a text prompt or an image layout, without retraining the model. It splits generation into two linked subproblems and uses an optimization method called ADMM to balance them, reporting improved quality and condition matching over several baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's proof does not analyze Algorithm 1: the z-update uses a Tweedie-evaluated gradient, while the proof bounds an exact z*-subproblem, and the theorem's rho <= 1/(6L) contradicts the proof's rho > 6L.","rationale":"The reader's weakest assumption already flags the unproven diffusion-proximal bridge and the missing theorem for the z-step. The present attack sharpens this into a direct proof mismatch: even granting Proposition 1 and the summability of delta_t, the Appendix's convergence argument bounds an exact-ADMM z-iteration, not the Tweedie-gradient update in Algorithm 1, and the rho condition in the theorem is the opposite of the condition used in the proof. This is an internal inconsistency, not a disagreement with prevailing consensus. The empirical results are promising and the decoupling idea is interesting, but they do not repair the formal guarantee. The reader's REJECT verdict remains appropriate; a future version with a corrected proof for the actual update, released code, and error bars would be worth re-evaluating.","tokens_in":18847,"tokens_out":14010,"duration_ms":154504,"concrete_test":"Simulate a 1D toy version of Algorithm 1 with guidance c_theta(tilde z_0(z),y)=exp(-(0.5z-1)^2/2), y=1, and take the proof's g(z)=(z-1)^2/2 (so L=1). Use eta=1/(rho+L), many inner iterations K_t, and any rho satisfying the theorem's rho <= 1/(6L). Track m_k=||z_{k+1}-z_k||^2+||x_{k+1}-x_k||^2 and the Lyapunov function in (A.6). If the residuals do not satisfy min_{j<=T} m_j=o(1/T), or if the Lyapunov decrease used in the proof fails, the theorem's z-subproblem is not the update Algorithm 1 executes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central formal claim is Theorem 2, stated as convergence of Algorithm 1 to a stationary point of (P) with min_j(||z_{j+1}-z_j||^2+||x_{j+1}-x_j||^2)=o(1/T). The proof in Appendix A.2 does not analyze the algorithm as written. Step 1 defines z*_{k+1} := argmin_z L(x_{k+1}, z, mu_k) with g(z) := -log c_theta(z,y), then bounds ||z_{k+1}-z*_{k+1}|| by gradient-descent contraction on that subproblem. But Algorithm 1's z-update is z_t^{(k+1)} = z_t^{(k)} - eta rho (z_t^{(k)} - x_t - nu_t) + eta grad_z log c_theta(tilde z_0(z_t^{(k)}), y), evaluated at Tweedie's estimate. Unless c_theta(tilde z_0(z),y)=c_theta(z,y), which is not assumed and is generically false for off-the-shelf parsers, sketch generators, CLIP, and motion encoders, the sequence bounded by (A.6) is not the sequence Algorithm 1 produces. Separately, Theorem 2 assumes rho <= 1/(6L), while Step 4 of the same proof says 'Because rho > 6L' before defining c1..c4. The two conditions cannot both be imposed as stated. This is not cosmetic, because the x-step is tied to Proposition 1 with rho=beta/(1-beta); standard DDPM schedules (beta ~ 0.01-0.02) give rho ~ O(0.01), so the proof's rho > 6L regime is not shown to be satisfied. Thus the convergence guarantee is unsupported even if Proposition 1 is granted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ADMMDiff, a training-free conditional generation framework that decouples the unconditional diffusion prior from a differentiable guidance loss by introducing an auxiliary variable z and the constraint x=z. The algorithm alternates a single reverse diffusion step, interpreted as an approximate proximal update for -log q_phi(x), with gradient descent on a Tweedie-estimated guidance loss for z, followed by a dual-variable update. The authors claim convergence to a stationary point of problem (P) at a sublinear rate under smoothness and summability assumptions, and they report experiments on non-linear guidance, linear inverse problems, and controllable motion synthesis. The central theoretical assertion is that one diffusion reverse step is a first-order approximation of the proximal operator of -log q_phi(x), so a pretrained diffusion model can be plugged into ADMM without retraining.","tokens_in":19273,"tokens_out":7919,"duration_ms":82613,"significance":"Conditional generation with off-the-shelf guidance functions is an active problem, and a principled decoupling that avoids a hand-tuned guidance weight would be a valuable contribution. The empirical results are suggestive: ADMMDiff leads most nonlinear image-guidance metrics in Table 1, most linear-inverse metrics in Table 3, and the trajectory objective in Table 4. The paper also gives a clear conceptual framing of diffusion reverse steps as proximal operators. However, the convergence analysis is an advertised core contribution, and as written the proof of Theorem 2 does not analyze the algorithm that is actually run. The theoretical significance is therefore not realized in the current manuscript. The empirical evidence alone would need fuller reproducibility details to carry the paper.","major_comments":[{"comment":"Theorem 2's proof does not analyze the z-update of Algorithm 1. Step 1 defines z*_{k+1} := argmin_z L(x_{k+1}, z, mu_k) with g(z) := -log c_theta(z,y), and then bounds ||z_{k+1} - z*_{k+1}|| by a contraction of K_k gradient-descent steps on that exact subproblem. Algorithm 1 instead updates z with the gradient of log c_theta(tilde_z0(z), y), where tilde_z0 is the Tweedie estimate. These two sequences coincide only if log c_theta(tilde_z0(z),y) equals log c_theta(z,y) up to a z-independent constant, a condition that is neither stated nor satisfied by BiSeNet, sketch generators, CLIP, or motion encoders. Consequently, the Lyapunov argument in (A.5)-(A.6) bounds a different sequence from the one produced by Algorithm 1, and Theorem 2 is not established for ADMMDiff.","section":"§4, Algorithm 1; Appendix A.2, Step 1"},{"comment":"There is a direct contradiction between the theorem's stated hypothesis rho <= 1/(6L) and the proof's Step 4, which invokes \"Because rho > 6L\" before defining the constants in (A.6). Both conditions cannot hold for L > 0. Since Proposition 1 gives rho = beta/(1-beta), and standard DDPM schedules have beta in the range 0.01-0.02, the proof's regime rho > 6L is not shown to be compatible with the diffusion schedules used in the experiments. The convergence rate claim therefore has no valid proof under the stated assumptions.","section":"Theorem 2 statement and Appendix A.2, Step 4"},{"comment":"The proof of Theorem 1 begins with the assertion \"Note that x_{t-1} = x_t - eta * grad h_t(x_t)\", but Algorithm 1 does not define h_t or a gradient step for x; the x-update is the single reverse-diffusion formula. No derivation is supplied showing that this reverse step is equivalent to the gradient descent form, and Proposition 1 is delegated to reference [60] rather than proved. This gap is load-bearing because Theorem 1 provides the delta_t error terms that enter the summability assumption of Theorem 2.","section":"Appendix A.1, Proof of Theorem 1"},{"comment":"The assumption that sum_t delta_t < +infinity is effectively an assumption that the errors of the inexact proximal steps vanish sufficiently fast. In Theorem 1, delta_t contains a term proportional to ||x_t - tilde_x0||^2 (Appendix A.1), and no argument is given that this term is summable for the pretrained diffusion models and off-the-shelf guidance functions tested. Since the paper presents this convergence result as the rigorous justification of ADMMDiff, the summability condition needs to be either proved for the tested settings or explicitly recognized as a substantive restriction rather than a mild assumption.","section":"Theorem 2 assumptions"}],"minor_comments":[{"comment":"In the x-update block, the second line evaluates s_theta at hat{x}_t, but only hat{x}_{t-1} was defined earlier in the block; the notation should be corrected, for example by defining hat{x}_t = z_t - nu_t/rho and then computing x_{t-1} from hat{x}_t.","section":"Algorithm 1, x-update block"},{"comment":"The z-update writes -eta*rho*(z_t^{(k)} - x_t - nu_t), but Section 3's update is -eta*rho*(z_t - x_t - (1/rho)*nu_t). The missing 1/rho factor gives the dual variable incorrect units relative to the augmented Lagrangian formulation.","section":"Algorithm 1, z-update block"},{"comment":"The text states that ADMMDiff achieves the best FID in 9 of 10 tasks, but in the FFHQ Inpaint(random) row DPS has FID 21.19 versus ADMMDiff's 22.04; the statement should be made task-specific rather than leaving the reader to identify the exception.","section":"Section 5.2 and Table 3"},{"comment":"The paper does not report the values of rho, eta, and K_t used in the experiments for any of the three tasks, which makes the strong empirical tables difficult to reproduce without additional information.","section":"Section 5, general reproducibility"}],"recommendation":"reject","confidential_remarks":"The core problem is that the main convergence theorem is proved for a different algorithm than the one actually used: the proof bounds an exact subproblem for z while Algorithm 1 uses a Tweedie-evaluated gradient, and the stated rho condition contradicts the proof's rho condition. This is a load-bearing error in the paper's advertised theoretical contribution. The empirical results may still be of interest, but a publishable version would need a new convergence analysis that honestly handles the Tweedie-evaluated z-update and verifies or removes the summability assumption. That is a substantial rewrite rather than a minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. The algorithmic idea is good: decoupling the unconditional diffusion prior and the guidance function into separate variables tied by an ADMM constraint is a fresh twist on training-free guidance, and the experimental evidence for it is broad and numerically strong. But the convergence proof as written does not analyze the algorithm it claims to analyze, and the theorem statement conflicts with its own proof. Treat the theory as unsupported.\n\nWhat's actually new: the two-variable formulation (x for the diffusion prior, z for the guidance), the use of a single diffusion reverse step as an approximate proximal operator for x, and Tweedie-evaluated gradient steps for z. That is a legitimate algorithmic contribution relative to DPS, LGD-MC, FreeDoM, and MPGD, which all balance prior and guidance with a tuned weight. The experiments cover a wide range—segmentation, sketch, text, linear inverse problems, and motion—and ADMMDiff beats the baselines on most metrics, often by a large margin. That is a substantive empirical claim.\n\nThe problems are in the theory. Algorithm 1's z-update uses ∇_z log c_θ(\\tilde{z}_0(z), y) — the guidance function evaluated at Tweedie's estimate. The proof in Appendix A.2 bounds gradient descent on the exact subproblem g(z) = -log c_θ(z,y). Those are different updates unless c_θ(\\tilde{z}_0(z)) = c_θ(z), which is generically false. So the contraction argument in Step 1 can't be applied to the sequence the algorithm actually produces. Second, Theorem 2 assumes ρ ≤ 1/(6L), but Step 4 of the proof says 'Because ρ > 6L' to build the Lyapunov inequality. The two conditions are compatible only if L < 1/6, which is not assumed and is unlikely for the neural-network energies at hand. This is not cosmetic: the x-step's ρ is tied to β/(1-β) by Proposition 1, which for standard DDPM schedules is O(0.01), far from the ρ > 6L regime. Theorem 1's proof also uses an update x_{t-1} = x_t - η∇h_t(x_t) that is not in the algorithm. And Proposition 1 is delegated to [60] without a local derivation, plus its ρ scaling is ambiguous.\n\nTwo more things: the experiments have no error bars and no hyperparameter values (η, K_t, ρ), and code is not released, so the numbers can't be independently checked. Theorem 2's assumption Σ δ_t < ∞ is effectively assuming the approximation errors vanish; that's not established for off-the-shelf guidance functions.\n\nWho is this for? Someone working on training-free guided diffusion will get useful empirical signals from this. The decoupling idea might inspire better algorithms. But nobody should cite the convergence result. The paper deserves a serious referee — the idea is novel and the empirical scope is impressive — and a revision that fixes (or drops) the convergence analysis, releases code, and documents hyperparameters could be a solid contribution. I'd send it to review, but the current theory should not survive.","headline":"Good algorithm, bad proof: the ADMM decoupling idea and the empirical results are promising, but Theorem 2's convergence analysis doesn't analyze the algorithm and its rho bounds contradict.","tokens_in":19834,"tokens_out":10325,"would_cite":false,"duration_ms":96354,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ADMMDiff rewrites conditional generation as two coupled variables, one for the diffusion prior and one for the condition, and proves the coupled iteration converges to a stationary point.","keywords":["guided diffusion","ADMM","training-free guidance","proximal operator","plug-and-play generation","motion synthesis","inverse problems"],"falsifier":"Run Algorithm 1 on CelebA-HQ segmentation guidance and measure $\\min_j (\\|z_{j+1}-z_j\\|^2 + \\|x_{j+1}-x_j\\|^2)$ against $T$; Theorem 2 predicts this quantity decays as $o(1/T)$, so any plateau above zero contradicts the claimed stationary-point convergence. Also record $\\sum_t \\delta_t$ for that guidance loss; if it diverges, the theorem's assumptions are not satisfied.","tokens_in":18626,"feed_emoji":"🎯","tokens_out":7113,"duration_ms":82890,"temperature":0.7,"pith_summary":"This paper proposes ADMMDiff, a training-free framework that guides an off-the-shelf unconditional diffusion model with any differentiable condition function without tuning a task-specific weight. The key move is to give the diffusion prior and the guidance function their own variables, $x$ and $z$, and couple them with the constraint $x = z$, so the balance between image quality and condition adherence is set adaptively by the ADMM dual variable. The paper also claims that one reverse diffusion step is a first-order approximation to the proximal operator of the prior's negative log-likelihood, which is what lets a pretrained diffusion model be plugged into the optimization as-is. Under smoothness assumptions, it proves convergence to a stationary point of the coupled objective with a sublinear rate. Experiments on CelebA-HQ image generation, FFHQ/ImageNet inverse problems, and human motion synthesis report better condition adherence and quality than several existing guided-diffusion methods.","feed_headline":"ADMM guides pretrained diffusion models with zero weight tuning","feed_subtitle":"A single reverse step is the prior's proximal update, so the same loop follows maps, text, and motion paths.","key_machinery":"The load-bearing identity is Proposition 1: a single reverse diffusion step, $x_{t-1} = \\frac{1}{\\sqrt{\\alpha_t}}(x_t + \\beta_t s_\\theta(x_t,t)) + \\sqrt{\\frac{1-\\bar\\alpha_{t-1}}{1-\\bar\\alpha_t}}\\beta_t \\epsilon$, is a first-order approximation to the proximal operator of $-\\frac{1}{\\rho}\\log q_\\phi(x)$ at $x_t$, with $\\rho = \\frac{\\beta_t}{1-\\beta_t}$. The auxiliary variable $z$ is updated by Tweedie's formula, $\\tilde z_0(z_t) = \\frac{1}{\\sqrt{\\bar\\alpha_t}}(z_t + (1-\\bar\\alpha_t)s_t)$, followed by $K_t$ gradient descent steps on the guidance loss $\\log c_\\theta(\\tilde z_0(z), y)$, which the paper treats as a first-order approximation to the proximal operator of $-\\log c_\\theta(z,y)$. These two inexact proximal steps are coupled by ADMM's dual variable $\\nu_t$ and penalty coefficient $\\rho$, which adaptively balance prior fidelity and condition satisfaction.","core_discovery":"The authors' discovery is that conditional generation can be formulated as a constrained optimization problem in which the diffusion prior and the guidance function live in separate variables, so the usual tuned weight is replaced by an ADMM coupling mechanism. They write the objective as $\\max_{x,z} \\log q_\\phi(x) + \\log c_\\theta(z,y)$ subject to $x = z$, solve it with alternating updates, and show that the diffusion reverse step serves as an inexact proximal operator for the prior. The paper's central theoretical claim is that this ADMM-based loop converges to a stationary point of the coupled objective, with the residual measure $\\min_j (\\|z_{j+1}-z_j\\|^2 + \\|x_{j+1}-x_j\\|^2)$ decaying as $o(1/T)$ under stated assumptions. Empirically, the paper reports state-of-the-art or competitive performance on nonlinear guided image synthesis, linear inverse problems, and controllable motion synthesis.","pith_inferences":["The proximal-step view suggests a compatibility criterion for other samplers: any sampler whose reverse step can be certified as a proximal approximation could be inserted into the ADMM loop, so the scheme is not inherently tied to the DDPM update used in the paper.","A testable extension is to adapt $K_t$ online from the primal and dual residuals instead of fixing it, which would make each subproblem's accuracy track the observed coupling error.","Because the two variables evolve separately before being fused, the framework could support swapping the guidance function mid-generation without restarting the diffusion process; the paper does not explore this.","The theorem's summability assumption is what separates convergence from failure in practice; measuring $\\sum_t \\delta_t$ for a given guidance loss would make convergence checking a one-time calibration rather than a faith-based assumption."],"forward_implications":["No tuned guidance weight is needed: ADMM's dual variable and penalty coefficient set the balance between prior fidelity and condition satisfaction adaptively.","The same algorithm works across nonlinear losses, such as segmentation maps, sketches, and CLIP text embeddings, and linear measurements such as super-resolution, inpainting, and deblurring.","If the assumptions of Theorem 2 hold, the coupled iteration converges to a stationary point of the posterior-approximation objective at rate $o(1/T)$, not merely in the limit.","The method extends to off-the-shelf motion diffusion models, giving trajectory-following behavior with lower objective and embedding distance than the tested guided-motion baselines."],"supporting_citations":[{"why":"Supplies Proposition 1, the equivalence between a reverse diffusion step and the proximal operator that lets ADMMDiff use a pretrained diffusion model without retraining.","marker":"[60]"},{"why":"One of the foundational ADMM references from which the splitting algorithm and Gauss-Seidel updates are taken.","marker":"[16]"},{"why":"The other foundational ADMM reference, providing the augmented Lagrangian and alternating update framework.","marker":"[17]"},{"why":"Provides the DPS baseline and the experimental setting for linear inverse problems and nonlinear guidance that ADMMDiff is compared against.","marker":"[8]"},{"why":"Provides the LGD-MC baseline, the Monte Carlo loss estimation approach, and the motion evaluation metrics used in Section 5.3.","marker":"[48]"},{"why":"Provides the FreeDoM baseline and the exact CelebA-HQ experimental protocol with BiSeNet, sketch, and CLIP guidance.","marker":"[58]"},{"why":"Provides the MPGD baseline and the Algorithm 1 and 3 settings used in the image experiments.","marker":"[20]"},{"why":"Provides the PnP-ADMM baseline, a prior ADMM-based image restoration method whose denoiser-as-proximal idea ADMMDiff extends to diffusion models.","marker":"[7]"},{"why":"Provides the DDRM baseline for linear inverse problems, one of the main comparisons in the restoration experiments.","marker":"[26]"}],"fun_headline_variants":["ADMM decouples diffusion guidance, no tuning needed","Diffusion guidance via ADMM: decouple and converge","No more weight tuning: ADMM for guided diffusion","ADMM makes guided diffusion self-balancing","ADMMDiff: decoupled guidance, guaranteed convergence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Proposition 1 — that a single reverse diffusion step is a first-order approximation to the proximal operator of the prior's negative log-likelihood, a statement the paper cites rather than proves — together with the unproved gradient-descent proxy for the guidance subproblem and the convergence theorem's assumption that the accumulated approximation errors stay finite.","fun_headline_variants_meta":{"raw":{"variants":["ADMM decouples diffusion guidance, no tuning needed","Diffusion guidance via ADMM: decouple and converge","No more weight tuning: ADMM for guided diffusion","ADMM makes guided diffusion self-balancing","ADMMDiff: decoupled guidance, guaranteed convergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000813,"raw_usage":{"total_tokens":3565,"prompt_tokens":943,"completion_tokens":2622,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":2546}},"tokens_in":559,"tokens_out":2622,"duration_ms":18758,"temperature":1.0,"reasoning_tokens":2546,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:53:52.755429+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on CelebA-HQ segmentation guidance and measure $\\min_j (\\|z_{j+1}-z_j\\|^2 + \\|x_{j+1}-x_j\\|^2)$ against $T$; Theorem 2 predicts this quantity decays as $o(1/T)$, so any plateau above zero contradicts the claimed stationary-point convergence. Also record $\\sum_t \\delta_t$ for that guidance loss; if it diverges, the theorem's assumptions are not satisfied.","supporting_citations":[{"cited_title":"walking” “walking backwards","cited_arxiv_id":null,"evidence_quote":"Supplies Proposition 1, the equivalence between a reverse diffusion step and the proximal operator that lets ADMMDiff use a pretrained diffusion model without retraining."},{"cited_title":"A dual algorithm for the solution of nonlinear variational problems via finite element approximation","cited_arxiv_id":null,"evidence_quote":"One of the foundational ADMM references from which the splitting algorithm and Gauss-Seidel updates are taken."},{"cited_title":"Sur l’approximation, par éléments finis d’ordre un, et la résolution, par pénalisation-dualité d’une classe de problèmes de dirichlet non linéaires","cited_arxiv_id":null,"evidence_quote":"The other foundational ADMM reference, providing the augmented Lagrangian and alternating update framework."},{"cited_title":"Loss-guided diffusion models for plug- and-play controllable generation","cited_arxiv_id":null,"evidence_quote":"Provides the LGD-MC baseline, the Monte Carlo loss estimation approach, and the motion evaluation metrics used in Section 5.3."},{"cited_title":"Freedom: Training-free energy-guided conditional diffusion model","cited_arxiv_id":null,"evidence_quote":"Provides the FreeDoM baseline and the exact CelebA-HQ experimental protocol with BiSeNet, sketch, and CLIP guidance."},{"cited_title":"Plug-and-play admm for image restoration: Fixed- point convergence and applications.IEEE Transactions on Computational Imaging, 3(1):84–98, 2016","cited_arxiv_id":null,"evidence_quote":"Provides the PnP-ADMM baseline, a prior ADMM-based image restoration method whose denoiser-as-proximal idea ADMMDiff extends to diffusion models."},{"cited_title":"Denoising diffusion restoration models","cited_arxiv_id":null,"evidence_quote":"Provides the DDRM baseline for linear inverse problems, one of the main comparisons in the restoration experiments."}],"review_version":1}