{"id":"8f11b748-ad6a-44f6-a992-052366c10768","arxiv_id":"2506.09416","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"NCVSD trains a conditional generator to sample from denoising posteriors, enabling one-step and multi-step image generation and plug-and-play inverse problem solving with fewer function evaluations.","lead":"This paper proposes NCVSD, a method that distills a pretrained diffusion model into a generative denoiser, a network that samples from the denoising posterior at any noise level. The result is one model that supports fast one-step generation, higher-quality multi-step generation, and zero-shot probabilistic inference for inverse problems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The NCVSD gradient's correctness depends on the online conditional score model D_phi tracking the evolving generator distribution; no tracking or convergence guarantee is provided, so the multi-step and PnP-GD exactness claims inherit this unverified assumption.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the online conditional score model D_ϕ is updated only one step per generator step, and the theoretical claims (Proposition 2, PnP-GD exactness) presuppose an accurate sampler of the denoising posterior. The paper provides no convergence or tracking analysis for D_ϕ, so the correctness of the gradient in Eq. (9) and all downstream guarantees rest on an empirical hope. This is genuinely load-bearing because if D_ϕ systematically lags, the generator may not converge to q(x0|yσ), invalidating the multi-step marginal preservation and the asymptotic exactness of PnP-GD. The concern is not a mathematical inconsistency in the paper's derivations—Proposition 1 and the VSD gradient are standard and correct—but rather a missing guarantee in the training dynamics. Since the empirical results are plausible and the theoretical statements are qualified (the paper notes the ideal scenario for PnP-GD), the appropriate verdict remains CONDITIONAL rather than REJECT or UNVERDICTED. The concrete test proposed would directly probe whether the one-step tracking is sufficient, and if not, would confirm the concern. Other possible issues (e.g., overstatement in the abstract about outperforming teachers, modified PnP-DM baseline, EMA in PnP-GD breaking exactness) are less central because they affect the strength of empirical claims rather than the core mechanism of the method.","tokens_in":25941,"tokens_out":15123,"duration_ms":159050,"concrete_test":"Train NCVSD on a tractable posterior where the true q(x0|yσ) is known (e.g., a Gaussian mixture prior in low dimensions, or 64x64 FFHQ with a synthetic known posterior). Compare two training regimes: (a) the paper's one-step D_ϕ update per generator step, and (b) D_ϕ updated to near-convergence (e.g., 50 inner steps) at each generator step. If regime (a) yields a significantly worse fit to the true posterior (e.g., higher KL divergence or FID) than regime (b), the tracking assumption is load-bearing. Alternatively, on the actual ImageNet setup, log the denoising MSE of D_ϕ on a fixed validation set of newly generated samples throughout training; if this MSE is markedly higher than the MSE of a converged score model on the same generated distribution, the gradient in Eq. (9) is biased.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 derives the NCVSD gradient (Eq. 9) from the reverse-KL objective (Eq. 8). Its correctness requires ∇_xt log p_θ(xt|yσ) to be estimated accurately by the online score model D_ϕ, trained via Eq. (13) with only one gradient step per generator step (Algorithm 3, Appendix B.2). No tracking or convergence guarantee is given for D_ϕ relative to the non-stationary generator distribution. If D_ϕ lags, the bias in Eq. (9) propagates to the generator update, so G_θ need not converge to the denoising posterior q(x0|yσ). Proposition 2 (multi-step marginals) and the PnP-GD asymptotic exactness (Section 4) both assume the generator is an exact sampler of q(x0|yσ), so the central claims of multi-step refinement and zero-shot probabilistic inference rest on this unverified tracking assumption. The empirical FID/LPIPS results do not disentangle a properly converged denoiser from one that merely performs well on the tested tasks.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Noise Conditional Variational Score Distillation (NCVSD), a method to distill a pretrained diffusion model into a conditional generator that approximates samples from the denoising posterior q(x0|yσ) across a range of noise levels. The key theoretical contribution is Proposition 1, a closed-form expression for the score of the denoising posterior q(xt|yσ) in terms of the unconditional score evaluated at an effective noise level. This is integrated into a VSD-style objective, with an additional adversarial loss and a carefully engineered parameterization. The trained generator supports one-step sampling, multi-step refinement (Proposition 2), and plug-and-play posterior sampling via a Split Gibbs Sampler (PnP-GD) claimed to be asymptotically exact. Experiments on ImageNet 64/512 and FFHQ-256 inverse problems show competitive FID and LPIPS with substantially fewer function evaluations than diffusion-based baselines.","tokens_in":26205,"tokens_out":10585,"duration_ms":107143,"significance":"If the method works as claimed, it offers a practical bridge between one-step distillation and iterative refinement, and it provides a fast, flexible posterior sampler for inverse problems. The derivation of the conditional score estimator is self-contained and correct, and the multi-step marginal construction is a clean extension of DDIM. The code is provided, and the experimental results are strong on both class-conditional generation and inverse problems. However, the paper's principal theoretical guarantees are conditional on untested assumptions about training convergence and sampler exactness, so the significance of the empirical findings is currently supported more by the experiments than by the theory.","major_comments":[{"comment":"The NCVSD gradient is unbiased only if D_phi(x_t,t,y_sigma,sigma) accurately estimates the score of the current generator distribution p_theta(x_t|y_sigma). In Algorithm 3, D_phi is updated once per generator step and is never guaranteed to track the non-stationary distribution of G_theta. If D_phi lags, Eq. (9) is biased and G_theta need not converge to q(x0|y_sigma). Since Proposition 2 and the Section 4 exactness claim both assume mu_theta = q(x0|y_sigma), the multi-step and PnP-GD guarantees inherit this unverified assumption. The paper should either provide a two-timescale convergence analysis, or empirical tracking diagnostics (e.g., monitor the training loss of Eq. (13) on held-out generator samples, or compare D_phi with a reference score), or explicitly state that the trained generator is only a biased approximation.","section":"Section 3.1, Eq. (9), Algorithm 3"},{"comment":"The statement that PnP-GD achieves \"asymptotically exact\" sampling is only true for the idealized SGS in which the prior step samples exactly from q(x0|y_sigma) and the likelihood step runs full ULA. The implemented algorithm uses a finite, pre-specified number of ULA steps with an adaptive step size gamma_sigma = C1*(beta^{-1}*C2 + sigma^{-2})^{-1}, and applies an EMA operation to the x0 samples (Algorithm 1, lines 6-8). Both the ULA discretization and the EMA introduce a bias that is not bounded by the TV-convergence argument of Vono et al. The paper should either derive a bias bound under these modifications or revise the abstract and Section 4 to state that PnP-GD is an approximate sampler.","section":"Section 4, Algorithm 1"},{"comment":"The marginal-preservation guarantee requires that each intermediate draw x0 ~ mu_theta(x0|y_sigma_eff) is an exact sample from q(x0|y_sigma_eff). The training distribution of sigma in Eq. (42) does not necessarily cover the effective noise levels sigma_eff encountered during the multi-step schedule, especially at early steps where sigma_eff can be very small. The paper does not evaluate the generator's accuracy as a function of sigma_eff. Please provide a calibration analysis (e.g., FID or conditional denoising error at different sigma_eff) or restrict the multi-step claim to the range where the generator is validated.","section":"Section 3.3, Proposition 2"}],"minor_comments":[{"comment":"The abstract claims the method \"outperforms teacher diffusion models\" when scaling test-time compute, but Table 1 on ImageNet-64 shows 4-step NCVSD-L FID 1.53 versus EDM2-L FID 1.33 and EDM2-XL FID 1.33, so the claim is too strong as stated.","section":"Abstract"},{"comment":"The training iteration count \"32k×3\" is ambiguous; state whether this means 32k generator updates plus 32k score-model and 32k discriminator updates, and explain how this is compared fairly against the single-number iteration counts of sCD and ECM.","section":"Section 5.1, Table 1"},{"comment":"There are several typos: \"direclty\" should be \"directly\", \"unconditonal\" should be \"unconditional\", and \"repectively\" in Section 5.1 should be \"respectively\".","section":"Section 3.1"},{"comment":"The claim \"record-breaking LPIPS\" is supported only on the FFHQ-256 dataset with a relatively small XS teacher model; the paper should state this scope explicitly in the abstract or conclusion.","section":"Section 5.2, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper's central ideas are sound and the experiments are compelling, but the main theoretical claims (multi-step marginals and asymptotic exactness of PnP-GD) rest on the untested assumption that the generator and the online score model converge. This is a common issue in VSD/GAN-style training, but the paper should either acknowledge the approximation with more precision or provide empirical tracking evidence. The manuscript fits the scope of the journal and the code release is a plus."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the NCVSD paper. My take: it's a legitimate, well-executed contribution, and the core theoretical observation, though elementary, is put to genuinely new use. Proposition 1 is just Gaussian conditioning plus Tweedie, but it lets you distill a teacher into a generative denoiser that can sample the denoising posterior at arbitrary noise levels. That framing is what's new, and it's useful: you get one-step generation, multi-step refinement, and plug-and-play posterior sampling from a single model. The experiments back this up. The FID numbers are competitive with consistency models at similar or smaller training budgets, and the inverse problem results are impressive, especially the 50 NFE phase retrieval. I appreciate that the code is promised and the training details are unusually complete.\n\nThe soft spots are real but not disqualifying. The asymptotic exactness language is doing too much work. Proposition 2 and the PnP-GD guarantee are conditional on the generator being an exact sampler of the denoising posterior, and the online score model D_phi tracking the non-stationary generator distribution is assumed. There is no convergence or tracking guarantee. That is a standard assumption in VSD-style methods, and I wouldn't call it a fatal flaw, but it means the theory is idealized. The abstract's claim that the method 'outperforms teacher diffusion models' is an overstatement; that only happens at 4 steps with the largest model, and at 1 step it is still behind the teacher. The inverse problem tables are honest about the distortion-perception tradeoff, and the PnP-DM baseline uses a single sample rather than the original 20, which is disclosed. These are minor caveats.\n\nWho is this for: people working on diffusion distillation, consistency models, and zero-shot inverse problem solvers. It deserves a serious referee. I would accept it for review and push the authors to soften the exactness claims and add an ablation or discussion of the tracking assumption. The core idea is sound and the empirical work is solid.","headline":"A solid diffusion distillation paper that repurposes a classical Gaussian conditioning identity to train generative denoisers, with strong experiments and idealized exactness claims that need softening.","tokens_in":26760,"tokens_out":2511,"would_cite":true,"duration_ms":27204,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","68T45","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"NCVSD shows the unconditional score of a diffusion model encodes every denoising posterior, so one distilled generative denoiser generates in one step, refines with added compute, and acts as a plug-and-play prior for inverse problems.","keywords":["diffusion distillation","generative denoiser","denoising posterior sampling","variational score distillation","noise conditional score estimator","plug-and-play inverse problems","few-step image generation","consistency models"],"falsifier":"On a tractable target such as a low-dimensional Gaussian mixture, compute both sides of the Proposition 1 identity exactly or by dense Monte Carlo; any mismatch falsifies the conditional-score estimator. Then distill a generative denoiser from an exact teacher score on the same target and compare its conditional samples at several noise levels with the analytic $q(x_0|y_\\sigma)$; if the sample histograms deviate measurably at converged loss, the premise that the NCVSD objective drives the generator to the denoising posterior is refuted.","tokens_in":25739,"feed_emoji":"⚡","tokens_out":13945,"duration_ms":120649,"temperature":0.7,"pith_summary":"NCVSD claims that a single distilled network can replace iterative diffusion sampling without giving up any of its benefits. The enabling move is a conditional-score identity: the score of the denoising posterior $q(x_0|y_\\sigma)$ can be computed exactly from a pretrained unconditional score evaluated at an effective noise level, so the otherwise intractable 'data score' inside Variational Score Distillation becomes a single teacher-network call. Training a one-step generator against this score, together with an adversarial term, yields a generative denoiser that samples the denoising posterior at any noise level. From that one model the paper obtains one-step generation from pure Gaussian noise, multi-step refinement whose intermediate distributions provably stay on the target, and an asymptotically exact plug-and-play sampler for inverse problems. If the results hold, diffusion models compress into networks that generate instantly, spend extra compute only when quality matters, and solve inverse problems with roughly twenty times fewer function evaluations than current diffusion solvers.","feed_headline":"Four distilled steps beat the diffusion teacher's 63","feed_subtitle":"One generative denoiser also solves inverse problems in 50 function evaluations, matching 1,000-step solvers.","key_machinery":"The load-bearing object is a conditional-score identity the paper calls the noise conditional score estimator (Proposition 1): restricting to the denoising posterior and pooling the two noise observations by precision gives $\\nabla_{x_t}\\log q(x_t|y_\\sigma) = t^{-2}(D_0(y_{\\sigma_{\\mathrm{eff}}}, \\sigma_{\\mathrm{eff}}) - x_t)$ with $y_{\\sigma_{\\mathrm{eff}}} = (\\sigma^{-2}y_\\sigma + t^{-2}x_t)/(\\sigma^{-2}+t^{-2})$ and $\\sigma_{\\mathrm{eff}} = (\\sigma^{-2}+t^{-2})^{-1/2}$. This one identity carries the argument: it converts the intractable data score of the conditional VSD gradient into a call to the pretrained teacher, leaving only the model score, which an online conditional score model $D_\\phi$ learns from the generator's own output. Around this core the paper assembles the auxiliary Jensen–Shannon adversarial loss, a magnitude-preserving ControlNet-style parameterization initialized from the teacher, uncertainty-weighted loss balancing, the DDIM-type latent construction of Proposition 2 that makes multi-step sampling marginal-preserving, and PnP-GD, a Split Gibbs sampler whose ULA likelihood step uses a step size that adapts to the annealing noise level.","core_discovery":"The paper's central claim is that a pretrained unconditional diffusion model already contains the score functions of every denoising posterior. Proposition 1 states that when $x_0$ is observed through two independent Gaussian channels with noise levels $\\sigma$ and $t$, the posterior $q(x_0|x_t, y_\\sigma)$ equals the denoising posterior at the precision-weighted average $y_{\\sigma_{\\mathrm{eff}}} = (\\sigma^{-2}y_\\sigma + t^{-2}x_t)/(\\sigma^{-2}+t^{-2})$ with effective noise level $\\sigma_{\\mathrm{eff}} = (\\sigma^{-2}+t^{-2})^{-1/2}$; consequently $\\nabla_{x_t}\\log q(x_t|y_\\sigma) = t^{-2}(\\mathbb{E}[x_0|y_{\\sigma_{\\mathrm{eff}}}] - x_t)$, computable through Tweedie's formula with a single forward pass of the unconditional denoiser. NCVSD inserts this estimator into the Variational Score Distillation gradient, conditioning the generator on noisy data $y_\\sigma$ and estimating the model score with an online conditional score model, then adds a Jensen–Shannon adversarial loss so the student can exceed the teacher. The resulting generative denoiser $G_\\theta(y_\\sigma, \\sigma, z)$ approximates samples of $q(x_0|y_\\sigma)$ across noise levels and delivers three capabilities: one-step generation at high noise, marginal-preserving multi-step sampling through a DDIM-style latent construction (Proposition 2), and asymptotically exact plug-and-play posterior sampling via a Split Gibbs sampler whose prior step is a single denoiser evaluation. Empirically, 4-step NCVSD-L reaches FID 1.76 on ImageNet-512, below 2-step sCM-XXL (1.88) and the teacher EDM2-XXL (1.81), and PnP-GD achieves best or second-best LPIPS on linear and nonlinear inverse problems at 50 function evaluations where diffusion solvers typically need about 1,000.","pith_inferences":["The effective-noise substitution is a general precision-pooling fact: two independent noisy observations of $x_0$ are posterior-equivalent to their precision-weighted average at the combined noise level, so the same identity suggests a recipe beyond distillation for turning any unconditional denoiser into one that conditions on extra noisy evidence.","The same weights act as an unconditional sampler, a conditional sampler, and a Bayesian prior; this points toward a single model handling both text-to-image generation and inverse problems, a unification the paper demonstrates piecewise but does not push.","Because the adversarial loss is what allows the student to exceed the teacher, the method's ceiling may be set less by the teacher's score accuracy than by GAN-style training stability and discriminator quality, so the acknowledged tuning burden becomes the likely scaling bottleneck.","The $\\zeta$ parameter of the multi-step construction interpolates between stochastic and deterministic refinement, and the experiments fix $\\zeta = 1.0$ throughout; sweeping it at inference could expose a quality-versus-diversity trade-off the paper does not explore."],"forward_implications":["A single distilled generative denoiser spans the speed–quality trade-off without retraining: FID on ImageNet-512 improves monotonically from 2.56 at 1 step to 2.03 at 2 steps to 1.76 at 4 steps for NCVSD-L.","Multi-step sampling is marginal-preserving rather than heuristic: Proposition 2 guarantees $p(x_i) = q(x_i|y_\\sigma)$ for every intermediate latent, so added steps refine the sample without drifting from the target denoising posterior.","With the adversarial term the student overtakes the teacher: 4-step NCVSD-L on ImageNet-512 (FID 1.76) beats 2-step sCM-XXL (1.88) and the teacher EDM2-XXL (1.81) at a fraction of the function evaluations.","PnP-GD solves linear and nonlinear inverse problems, including phase retrieval, in about 50 function evaluations with best or second-best LPIPS, roughly 20 times fewer evaluations than diffusion solvers such as DAPS and DPS that use about 1,000.","The noise conditional score estimator is transportable: as the paper states, any score distillation method can adopt it to distill generative denoisers from pretrained score models."],"supporting_citations":[{"why":"Supplies the Variational Score Distillation gradient that NCVSD extends by conditioning on noisy data.","marker":"Wang et al., 2024"},{"why":"Provides the noise schedule $\\sigma_t = t$, the Tweedie-based score link, EDM-style preconditioning, and the design-space analysis used for the teacher and parameterization.","marker":"Karras et al., 2022"},{"why":"Tweedie's formula is the basis of Proposition 1, bridging the posterior mean and the conditional score.","marker":"Efron, 2011"},{"why":"Consistency models are the main comparison class the distilled denoisers are claimed to match or beat in FID.","marker":"Song et al., 2023c"},{"why":"PnP-DM is the SGS-based prior sampling approach that PnP-GD accelerates by replacing reverse diffusion with one generative denoiser call.","marker":"Wu et al., 2024"},{"why":"The Split Gibbs Sampler is the framework that gives PnP-GD its asymptotic exactness guarantee.","marker":"Vono et al., 2019"},{"why":"DPS is a representative approximate-conditional-score inverse solver used as a baseline and point of contrast.","marker":"Chung et al., 2023"},{"why":"Diff-Instruct shows VSD applied to image distillation and motivates estimating the model score with an online score model.","marker":"Luo et al., 2023"},{"why":"The GAN objective underlies the adversarial Jensen–Shannon loss used to train the discriminator.","marker":"Goodfellow et al., 2014"},{"why":"EDM2 provides the teacher models, training-dynamics analysis, and magnitude-preserving network designs reused in the parameterization.","marker":"Karras et al., 2024"}],"fun_headline_variants":["4-step NCVSD beats 63-step diffusion teacher on ImageNet-512","One denoiser distills all noise levels from a diffusion model","Generative denoiser solves inverse problems in 50 function evaluations","NCVSD: one-step generation and multi-step refinement from one model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the distilled generator actually sampling the true denoising posterior at every noise level: the combined reverse-KL and adversarial training has no convergence guarantee, and the auxiliary score model is updated only one step per generator step with no guarantee that it keeps tracking the generator's evolving output distribution.","fun_headline_variants_meta":{"raw":{"variants":["4-step NCVSD beats 63-step diffusion teacher on ImageNet-512","One denoiser distills all noise levels from a diffusion model","Generative denoiser solves inverse problems in 50 function evaluations","NCVSD: one-step generation and multi-step refinement from one model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000741,"raw_usage":{"total_tokens":3420,"prompt_tokens":1174,"completion_tokens":2246,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":790,"completion_tokens_details":{"reasoning_tokens":2167}},"tokens_in":790,"tokens_out":2246,"duration_ms":18243,"temperature":1.0,"reasoning_tokens":2167,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:49:37.556741+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a tractable target such as a low-dimensional Gaussian mixture, compute both sides of the Proposition 1 identity exactly or by dense Monte Carlo; any mismatch falsifies the conditional-score estimator. Then distill a generative denoiser from an exact teacher score on the same target and compare its conditional samples at several noise levels with the analytic $q(x_0|y_\\sigma)$; if the sample histograms deviate measurably at converged loss, the premise that the NCVSD objective drives the generator to the denoising posterior is refuted.","supporting_citations":[{"cited_title":"ProlificDreamer : High-fidelity and diverse text-to-3D generation with variational score distillation","cited_arxiv_id":null,"evidence_quote":"Supplies the Variational Score Distillation gradient that NCVSD extends by conditioning on noisy data."},{"cited_title":"Elucidating the design space of diffusion-based generative models","cited_arxiv_id":null,"evidence_quote":"Provides the noise schedule $\\sigma_t = t$, the Tweedie-based score link, EDM-style preconditioning, and the design-space analysis used for the teacher and parameterization."},{"cited_title":"Tweedie’s formula and selection bias","cited_arxiv_id":null,"evidence_quote":"Tweedie's formula is the basis of Proposition 1, bridging the posterior mean and the conditional score."},{"cited_title":"Split-and-augmented Gibbs sampler -- Application to large-scale inference problems","cited_arxiv_id":null,"evidence_quote":"The Split Gibbs Sampler is the framework that gives PnP-GD its asymptotic exactness guarantee."},{"cited_title":"T., Klasky, M","cited_arxiv_id":null,"evidence_quote":"DPS is a representative approximate-conditional-score inverse solver used as a baseline and point of contrast."},{"cited_title":"Diff-Instruct : A universal approach for transferring knowledge from pre-trained diffusion models","cited_arxiv_id":null,"evidence_quote":"Diff-Instruct shows VSD applied to image distillation and motivates estimating the model score with an online score model."},{"cited_title":"Generative adversarial nets","cited_arxiv_id":null,"evidence_quote":"The GAN objective underlies the adversarial Jensen–Shannon loss used to train the discriminator."},{"cited_title":"Analyzing and improving the training dynamics of diffusion models","cited_arxiv_id":null,"evidence_quote":"EDM2 provides the teacher models, training-dynamics analysis, and magnitude-preserving network designs reused in the parameterization."}],"review_version":1}