{"id":"75cd50bb-2033-431f-b652-347f49b9a500","arxiv_id":"2506.19741","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Noise Consistency Training adds new controls to pre-trained one-step generators by training a lightweight adapter with a noise-space consistency loss, matching conditional generation quality at a fraction of the compute.","lead":"This paper introduces a new way to teach a fast one-step image generator to follow extra instructions, like edge maps, depth, or reference images, without retraining the base model. The approach adds a small adapter and trains it with a noise-space consistency loss, achieving controllable generation in a single step.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 2's claimed equivalence between the one-particle consistency loss and MMD is algebraically false: for k(x,y)=-||x-y||^2, E||X-Y||^2 = ||mu_X-mu_Y||^2 + tr Sigma_X + tr Sigma_Y, whereas MMD_k^2 = 2||mu_X-mu_Y||^2.","rationale":"The reader's conditional verdict identifies the gap between the practical surrogate and the theorem's MMD condition, but my check shows the gap is stronger: the claimed 'practical estimation' is not an estimator of MMD at all. The one-particle objective contains within-distribution variance terms that the population MMD cancels. For k=-||x-y||^2, the MMD collapses to a mean-matching criterion, so even if Lemma 2 were corrected, Theorem 1's characteristic-kernel assumption is not met by the paper's own chosen kernel. Because the theoretical bridge between Algorithm 1 and Theorem 1 is broken, and the remaining empirical evidence (internal dataset, FID computed against unconditional generator outputs, no error bars, no disclosed dataset) cannot substantiate the 'state-of-the-art' claim on its own, the central claim is not established. I would reject the paper in its current form. A resubmission could address this by replacing the surrogate with a proper characteristic-kernel MMD estimator, proving a corrected theorem, and providing a publicly reproducible evaluation.","tokens_in":15257,"tokens_out":11290,"duration_ms":130816,"concrete_test":"Analytical check: for independent X,Y, expand E||X-Y||^2 = ||mu_X-mu_Y||^2 + tr Sigma_X + tr Sigma_Y, and compute MMD^2 for k(x,y)=-||x-y||^2 as 2||mu_X-mu_Y||^2. For the concrete case P=Q=N(0,1), the surrogate equals 2 while MMD_k^2 equals 0, directly falsifying Remark 1's claimed 'practical implementation' of the MMD loss. If the pseudo-Huber kernel is intended instead, repeat the same check: a one-particle expected distance cannot equal MMD^2, which requires the within-sample terms E[k(X,X')] and E[k(Y,Y')]; the paper should either provide a corrected identity or replace the surrogate with a proper multi-particle characteristic-kernel MMD estimator.","verdict_should_be":"REJECT","load_bearing_attack":"Section 3.2 and Appendix A.3: Theorem 1 requires zero MMD between adjacent pushed-forward distributions under a characteristic kernel. The practical loss, Eq. (6) and Remark 1, is a one-particle paired distance with k(x,y)=-||x-y||^2 or pseudo-Huber. The paper's own Remark 1 admits -||x-y||^2 is not a proper positive-definite kernel; for this k, MMD^2(P,Q)=2||mu_P-mu_Q||^2, which characterizes first moments only, not distributions. More importantly, the one-particle surrogate is not an estimator of this MMD. For independent X~P and Y~Q, E||X-Y||^2 = ||mu_P-mu_Q||^2 + tr Sigma_P + tr Sigma_Q. The variance terms remain positive even when P=Q, so the surrogate can be minimized by collapsing variance rather than by matching the interpolated distributions. Additionally, Algorithm 1 uses the same z and eps for the paired noisy latents, so the two draws are correlated and cannot serve as the independent two-sample estimator required by MMD. Consequently, Lemma 2 and the chain of equalities (T_phi)#nu_{t_k} = ... = (T_phi)#nu_{t_0} in Theorem A.1 do not follow from Algorithm 1's update. The theorem is vacuous for the trained model; the central theoretical justification for NCT is unsupported. The experimental SOTA claim then rests on an internal dataset, an unconventional FID protocol, and no error bars, so it cannot independently carry the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Noise Consistency Training (NCT), a method for adding conditioning signals to a pre-trained one-step generator without retraining the base diffusion model. An adapter network is trained using a noise-consistency loss that compares the adapted generator's outputs on two noisy versions of the same latent noise, together with a boundary loss that anchors the adapter to the frozen one-step generator on fully coupled (z,c) pairs. The authors prove a theorem (Theorem 1 / Theorem A.1) stating that if the boundary loss is zero and the summed MMD between adjacent pushed-forward distributions is zero, then the adapted generator maps independent p(z)p(c) to the target joint distribution p_theta(x)p(c|x). They further claim in Lemma 2 that the practical noise-consistency loss is a single-particle estimator of this summed MMD. Experiments on Canny, HED, depth, super-resolution, and image-prompted generation report lower FID and consistency scores than ControlNet, DI+ControlNet, and JDM, using one network evaluation per image.","tokens_in":15666,"tokens_out":11665,"duration_ms":117377,"significance":"If the theoretical result applied to the implemented method, NCT would be a valuable lightweight approach to controllable one-step generation, avoiding additional distillation and preserving the base generator. The paper is also commendable for tackling a relevant problem and for providing a modular adapter formulation with experiments across several control modalities. However, the theoretical guarantee is the central justification of the method, and the connection between the theorem and the practical loss is not established: Lemma 2 is algebraically incorrect, and the kernels used in the implementation are not characteristic. The experimental evaluation is weakened by an unconventional FID protocol, an internal dataset, and the absence of error bars. The idea is interesting, but the paper does not currently provide a sound theoretical foundation or a fully convincing empirical validation for its central claims.","major_comments":[{"comment":"The claim that the noise-consistency loss is a practical estimation of the summed MMD is algebraically incorrect. For the implemented kernel k(x,y)=-||x-y||^2, one has MMD^2(P,Q)=2||μ_P-μ_Q||^2, whereas the single-particle paired expectation E||X-Y||^2 with independent X~P and Y~Q equals ||μ_P-μ_Q||^2 + tr Σ_P + tr Σ_Q. The variance terms remain positive even when P=Q, so the surrogate can be minimized by collapsing variance rather than by matching the interpolated distributions. Moreover, Eq. (6) and Algorithm 1 use the same z and ϵ in both z_{t_n} and z_{t_{n-1}}, so the two evaluations are not independent draws from p_{θ,φ,t_n} and p_{θ,φ,t_{n-1}}; the one-particle MMD estimator interpretation therefore does not apply. The appendix provides no proof of Lemma 2; Remark 1 only restates the implementation. Consequently, the chain of equalities (T_φ)#ν_{t_N}=...=(T_φ)#ν_{t_0} in Theorem A.1 is not connected to the loss that Algorithm 1 actually minimizes.","section":"§3.2, Lemma 2 and Appendix A.3, Remark 1"},{"comment":"Even if the surrogate loss were an MMD estimator, the theorem's consistency condition requires a characteristic kernel, while the practical loss uses kernels that are not characteristic. The paper itself notes that k(x,y)=-||x-y||^2 is not a proper positive-definite kernel; for this kernel, MMD^2(P,Q)=2||μ_P-μ_Q||^2, which characterizes only first moments rather than full distributions. The pseudo-Huber kernel mentioned in Remark 1 is conditionally positive definite and induces an energy distance, but the paper does not establish that it is characteristic on the relevant image space, nor that the single-particle paired loss estimates the corresponding MMD. Thus the conclusion that f_{θ,φ} maps independent p(z)p(c) to p_θ(x)p(c|x) is not established for the objective actually minimized.","section":"§3.2, Theorem 1 and Remark 1"},{"comment":"The FID protocol is nonstandard: the FID score is computed by comparing images generated by the unconditional base model against images generated with the conditional inputs, rather than by comparing generated images to a real-image reference distribution. This measures the marginal distance from the base generator, not image quality in the usual FID sense. In addition, the paper reports no error bars or statistical significance tests, and the dataset is internal. The abstract's claim of 'state-of-the-art controllable generation' is therefore not independently supported by the reported numbers.","section":"§4.1, Evaluation Metric"}],"minor_comments":[{"comment":"The notation E_{z,c|z,ϵ} in the simplified expression is ill-formed; the expectation should be over z, c, and ϵ separately.","section":"§3.2, Eq. (6)"},{"comment":"The symbol N is used both for the number of time steps in Eq. (7) and for the particle number in Remark 1; please disambiguate these two uses.","section":"§3.2 and Appendix A.3"},{"comment":"Please specify how the consistency measure h(x) is defined for the super-resolution task, where the condition is a low-resolution image rather than an edge or depth map.","section":"§4.1"},{"comment":"There are several typos: 'find find a shortcut solution', 'Contorlnet', and 'Geneartion' should be corrected.","section":"§4.3 and Appendix B"},{"comment":"The IP-Adapter result is taken from the official report; please clarify whether the base model, prompt set, and evaluation protocol are identical to those used for NCT so that the comparison is meaningful.","section":"§4.2, Table 2"}],"recommendation":"reject","confidential_remarks":"The paper contains a high concentration of self-citations (Diff-Instruct, JDM, Reward-Instruct, trajectory distribution matching), which is not disqualifying but makes the related-work positioning appear narrower than necessary. The decisive issue is the false Lemma 2: the theoretical guarantee is the paper's main justification, and it does not apply to the implemented loss. Fixing this would require either a substantially different theoretical analysis or a reformulation of the method, and the experimental section would also need to be redone with a standard evaluation protocol. I therefore recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the quick take: Noise Consistency Training is a genuinely new way to graft a control signal onto a frozen one-step generator — diffuse the input latent, enforce consistency between outputs at adjacent noise levels, and anchor one end with a boundary loss. The experiments show it works: on Canny/HED/depth/super-res it beats a 50-step ControlNet at one NFE, and the IP-Adapter numbers are competitive at 1 NFE vs 100. That is a real practical result, and the qualitative figures back it up.\n\nThe problem is the theory. Theorem 1 is a clean statement about population MMDs, but the loss in Eq. (6) and Algorithm 1 is not an MMD estimator. For the kernel the paper itself cites in Remark 1, k(x,y)=-||x-y||^2, the expected one-particle paired squared distance is ||mu_P-mu_Q||^2 + tr Sigma_P + tr Sigma_Q, whereas MMD^2 is 2||mu_P-mu_Q||^2. The variance terms are positive even when P=Q, so the loss can be minimized by shrinking variance rather than matching distributions. And the two noisy latents share the same z and epsilon, so they are not independent draws from nu_{t_{k+1}} and nu_{t_k}. Lemma 2's 'practical estimation' claim is therefore false in the strict sense, and the chain of equalities in Theorem A.1 does not follow from what the algorithm optimizes. The paper's own Remark 1 concedes the kernel isn't positive definite, but then still leans on MMD language. This is a load-bearing gap: the stated theory is vacuous for the trained model.\n\nThe empirical side has its own soft spots: the FID is computed against images from the base uncontrolled model, not against a real-data reference set, which is an unusual protocol; there are no error bars; and the training set is internal and undisclosed. The baselines are somewhat narrow, and the JDM comparison is a reimplementation. These are fixable, and the ablations are genuinely informative — without the boundary loss the distribution collapses, without the consistency loss the condition is ignored. So the method is plausible, just not as settled as the text suggests.\n\nWho's this for? People working on single-step conditional generation or adapters for distilled models. It deserves a serious referee, but I'd ask for major revision: either make the theory match the actual loss (prove something about the biased surrogate) or drop the formal claims and present the loss as a heuristic. Also standard FID, error bars, and dataset details. I'd send it out.","headline":"A one-step control adapter with real practical appeal, but the theory section doesn't cover the actual loss being optimized.","tokens_in":16107,"tokens_out":5852,"would_cite":true,"duration_ms":58228,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Noise Consistency Training lets a pre-trained one-step generator take on new controls — edges, depth, reference images — by enforcing agreement across diffused noise levels, and proves this agreement makes the generator sample the…","keywords":["Noise Consistency Training","one-step generator","controllable generation","adapter","distribution matching","maximum mean discrepancy","diffusion distillation","conditional generation"],"falsifier":"Train NCT on a benchmark where $p(x|c)$ is deliberately multimodal, then compute the MMD with a characteristic kernel between the adapted generator's joint output distribution and $p_\\theta(x)p(c|x)$; if that characteristic-kernel MMD stays bounded away from zero while the NCT loss goes to zero, the surrogate is not enforcing the consistency condition the theorem requires.","tokens_in":15081,"feed_emoji":"🎨","tokens_out":6816,"duration_ms":65639,"temperature":0.7,"pith_summary":"This paper introduces Noise Consistency Training (NCT), a way to give a pretrained one-step image generator a new conditioning signal — edges, depth, a reference image — without retraining the base generator or running diffusion distillation. The key move is to diffuse the generator's input noise: a noise vector $z$ is interpolated toward pure noise by $z_t = \\sqrt{1-\\sigma_t^2}\\,z + \\sigma_t\\,\\epsilon$, so that pairs $(z_t, c)$ slowly decouple from the condition $c$. The training loss asks the adapted generator to produce the same image from a less-coupled noisy input as from a more-coupled one, with a stop-gradient, plus a boundary loss that pins the fully coupled input to the original generator's output. The paper proves that if both the boundary loss and the summed MMD between adjacent pushed-forward noise levels are exactly zero, the adapted generator maps an independent noise-condition pair to the joint distribution $p_\\theta(x)p(c|x)$. Experiments report that with one function evaluation NCT matches or beats a 50-step ControlNet baseline on edge, depth, and super-resolution controls, and matches a 100-step image-prompt adapter at 1/100th the compute.","feed_headline":"NCT: one-step generator adds new controls with a noise-consistency loss","feed_subtitle":"Matches 50-step ControlNet on edges, depth, and super-resolution using just one generation step.","key_machinery":"The mechanism is the noise-space consistency loss plus boundary loss on a variance-preserved diffusion of the latent $z$. For a shared noise sample $z$ and a condition $c$ drawn from $p(c|f_\\theta(z))$, the loss compares $f_{\\theta,\\phi}(z_{t_{k+1}}, c)$ with the stop-gradient of $f_{\\theta,\\phi}(z_{t_k}, c)$, where $z_t = \\sqrt{1-\\sigma_t^2}\\,z + \\sigma_t\\,\\epsilon$. As $t$ increases, the pair $(z_t, c)$ becomes less coupled; aligning outputs across adjacent $t$ levels teaches the adapter to rely on $c$ without destroying the one-step generator's mapping. The boundary loss $d(f_{\\theta,\\phi}(z,c), f_\\theta(z))$ anchors the boundary at $t=0$. Lemma 1 interprets the diffusion as an interpolation of joint latent-condition distributions; Lemma 2 identifies the summed single-particle consistency loss as a practical estimate of the summed MMD; Theorem 1 chains zero MMD over adjacent levels to conclude $(T_\\phi)_\\#\\rho = \\eta$ with $T_\\phi(z,c) = (f_{\\theta,\\phi}(z,c), c)$. The dual variable $\\lambda$ in the primal-dual update balances the consistency and boundary terms.","core_discovery":"The central claim is that a consistency condition imposed in the noise space, not the image space, is sufficient to inject a new control into a frozen one-step generator. For coupled pairs $(z,c)$ sampled from $p(z)p(c|f_\\theta(z))$, diffusing $z$ along a variance-preserving schedule creates an interpolation from fully coupled noise-condition pairs to independent pairs; the consistency loss forces the adapter's outputs at adjacent noise levels to agree. Together with the boundary loss, these two conditions imply, by Theorem 1, that $f_{\\theta,\\phi}(z,c)$ with $z$ and $c$ sampled independently yields samples distributed as $p_\\theta(x)p(c|x)$. The theorem itself requires zero population MMD with a characteristic kernel and zero boundary loss; in practice the method minimizes a single-particle surrogate with kernels such as $-\\|x-y\\|^2$ or pseudo-Huber, which the paper explicitly notes are not proper characteristic kernels. The reported result is controllable generation in one forward pass, with FID and condition-consistency better than ControlNet at 50 NFE, DI+ControlNet at 1 NFE, and JDM at 1 NFE on edge, depth, and 8x super-resolution tasks.","pith_inferences":["If the boundary-plus-consistency mechanism works as proven, NCT is a generic recipe: any condition expressible as a likelihood $p(c|x)$ can be injected into any one-step generator by training only a small adapter, and the same argument should extend beyond images to other latent-native generators.","A natural stress test is whether the single-particle, non-characteristic kernel surrogate is sufficient when the conditional distribution $p(x|c)$ is multimodal; a kernel of the form $-\\|x-y\\|^2$ cannot separate two distributions with the same mean, so the theoretical guarantee may need a characteristic kernel or a larger particle number to hold strictly.","The test-time composition result suggests that NCT adapters could be composed multiplicatively by merging condition likelihoods, provided each adapter's boundary condition is preserved; this looks testable without any further training.","The tight coupling between the boundary condition and the frozen generator's image prior suggests that NCT's quality ceiling is set by the original one-step generator; if the base generator drifts or is fine-tuned, adapters may need to be re-anchored."],"forward_implications":["With one function evaluation, a generator fitted with an NCT adapter can follow structural controls such as edge maps and depth maps at quality comparable to 50- or 100-step diffusion baselines.","New controls can be added to an already distilled one-step generator without retraining the base model, without diffusion distillation, and without access to the original training images.","Different adapters trained for different controls can be combined at inference time, so an image-prompt adapter and a structure adapter can be stacked to obtain joint text, image, and structure control.","The training objective is a distributional-matching objective in the generator's latent space, so the same machinery transfers from ControlNet-style adapters to IP-Adapter-style image-prompt adapters.","The primal-dual formulation gives an automatic early-stopping signal when the dual variable vanishes, avoiding expensive hyperparameter tuning for the boundary-loss weight."],"supporting_citations":[{"why":"Supplies the adapter architecture and the conditional denoising baseline that NCT builds on and compares against.","marker":"[15]"},{"why":"Provides the consistency-loss idea that NCT ports from image space to the generator's noise space.","marker":"[26]"},{"why":"Defines the MMD distributional distance used in Lemma 2 and Theorem 1.","marker":"[16]"},{"why":"Produces the pre-trained one-step generator used in the experiments through distillation.","marker":"[3]"},{"why":"Is the one-step controllable-generation baseline that requires an extra distillation phase, which NCT is designed to avoid.","marker":"[9]"},{"why":"Supplies the image-prompt adapter architecture used for the image-prompted generation experiments.","marker":"[39]"},{"why":"Provides the convergence guarantee for the primal-dual constrained optimization used in Algorithm 1.","marker":"[34]"}],"fun_headline_variants":["Noise-consistency loss injects new controls into frozen one-step generators","One-step control via noise-space consistency, no base retraining","Train a tiny adapter in noise space to control one-step generators","NCT: add controls to a one-step generator with a noise consistency loss","Control a one-step generator without retraining or original images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result depends on the assumption that the single-particle noise-consistency surrogate, using kernels such as $-\\|x-y\\|^2$ that are not strictly characteristic, drives the population MMD and the boundary loss all the way to zero for some reachable parameter set — a gap between Theorem 1 and Algorithm 1 that the paper leaves open.","fun_headline_variants_meta":{"raw":{"variants":["Noise-consistency loss injects new controls into frozen one-step generators","One-step control via noise-space consistency, no base retraining","Train a tiny adapter in noise space to control one-step generators","NCT: add controls to a one-step generator with a noise consistency loss","Control a one-step generator without retraining or original images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1470,"prompt_tokens":1057,"completion_tokens":413,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":673,"completion_tokens_details":{"reasoning_tokens":322}},"tokens_in":673,"tokens_out":413,"duration_ms":4142,"temperature":1.0,"reasoning_tokens":322,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:25:41.404109+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train NCT on a benchmark where $p(x|c)$ is deliberately multimodal, then compute the MMD with a characteristic kernel between the adapted generator's joint output distribution and $p_\\theta(x)p(c|x)$; if that characteristic-kernel MMD stays bounded away from zero while the NCT loss goes to zero, the surrogate is not enforcing the consistency condition the theorem requires.","supporting_citations":[{"cited_title":"Adding conditional control to text-to-image diffusion models","cited_arxiv_id":null,"evidence_quote":"Supplies the adapter architecture and the conditional denoising baseline that NCT builds on and compares against."},{"cited_title":"Diff- instruct: A universal approach for transferring knowledge from pre-trained diffusion models","cited_arxiv_id":null,"evidence_quote":"Produces the pre-trained one-step generator used in the experiments through distillation."},{"cited_title":"Adding Additional Control to One-Step Diffusion with Joint Distribution Matching","cited_arxiv_id":"2503.06652","evidence_quote":"Is the one-step controllable-generation baseline that requires an extra distillation phase, which NCT is designed to avoid."},{"cited_title":"Constrained Learning with Non-Convex Losses","cited_arxiv_id":"2103.05134","evidence_quote":"Provides the convergence guarantee for the primal-dual constrained optimization used in Algorithm 1."}],"review_version":2}