Diffusion-Classifier Synergy (DCS)

• Diffusion-Classifier Synergy (DCS) is a family of methods integrating diffusion models and classifiers to mutually guide generative and discriminative outcomes.
• It employs various strategies such as classifier-guided sampling, diffusion-as-classifier techniques, and shared-backbone architectures to optimize performance.
• Empirical studies show DCS improves generative quality and classification accuracy across modalities while addressing challenges like computational overhead and domain specificity.

Diffusion-Classifier Synergy (DCS) denotes a family of methods in which diffusion models and classifiers are coupled tightly enough that discriminative information changes generative dynamics, generative structure becomes a discriminative signal, or both. Across the cited literature, this coupling appears in several distinct forms: standard classifier-guided sampling through the conditional correction $\nabla_{x_t}\log p_t(y\mid x_t)$ (Kawar et al., 2022), zero-shot or few-shot classification by comparing denoising errors across candidate conditions (Li et al., 2023), repurposing a pretrained classifier as the backbone of a diffusion generator (Makarov et al., 18 Jun 2026), joint generative-discriminative training with shared representations (Deja et al., 2023), training-time regularization that mines classifier-like signals from diffusion models themselves (Jha et al., 2024), and coupled inference procedures that jointly diffuse signals and logits (Nurko et al., 17 Feb 2026). The term is also used explicitly for a reward-aligned mutual-boosting loop in few-shot class-incremental learning, where classifier state defines the reward that steers diffusion sampling and generated images in turn refine the classifier (Wu et al., 4 Oct 2025).

1. Conceptual scope and historical development

The narrowest meaning of DCS is the standard classifier-guided diffusion construction: a diffusion model estimates the unconditional score, while a classifier supplies the conditional correction that steers reverse-time sampling toward a target class. In this formulation, class-conditional generation is obtained by augmenting the unconditional score with the classifier gradient, so the classifier functions as a control field over noisy intermediate states rather than merely as a terminal predictor (Kawar et al., 2022).

The broader literature expands this idea substantially. In one direction, diffusion models themselves are treated as classifiers: denoising error under different candidate conditions is used as a surrogate for conditional log-likelihood, yielding “Diffusion Classifier” style zero-shot recognition without additional classifier training (Li et al., 2023). In another direction, classifiers cease to be external modules and instead become structural parts of the generator, as in the frozen-backbone speech system where a pretrained noise-conditioned classifier is repurposed as the score model’s encoder and only a lightweight score-prediction decoder is trained (Makarov et al., 18 Jun 2026).

A further expansion is from sampling-time guidance to training-time co-adaptation. “Learning Data Representations with Joint Diffusion Models” shares a UNet encoder between denoising and classification and trains both objectives end-to-end (Deja et al., 2023), while “Class-N-Diff” couples a latent DiT generator and a ResNet-101 classifier through a joint loss in dermoscopic imaging (Munia et al., 19 Oct 2025). Other works move the synergy into continual learning, robustness, or test-time refinement, suggesting that DCS is better understood as a general design principle for sharing information between discriminative and diffusion processes than as a single algorithmic template (Jha et al., 2024).

2. Mathematical foundations

A central identity for DCS is the decomposition of the class-conditional score: $$\nabla_{X_t}\log p_t(X_t\mid y)=\nabla_{X_t}\log p_t(X_t)+\nabla_{X_t}\log p_t(y\mid X_t).$$ In standard classifier guidance, the first term is estimated by a diffusion score model and the second by a noise-conditioned classifier, giving a guided score of the form

$$\tilde s(X_t,t,y)=s_\theta(X_t,t)+\gamma\,\nabla_{X_t}\log p_{\phi,t}(y\mid X_t),$$

with guidance strength $\gamma\ge 0$ (Makarov et al., 18 Jun 2026). The same decomposition underlies robust-classifier guidance, which argues that the practical bottleneck is not whether the classifier predicts the right label, but whether its gradients at noisy diffusion states are semantically meaningful (Kawar et al., 2022).

A second mathematical route reverses the direction of dependence and uses diffusion as a classifier. In “Your Diffusion Model is Secretly a Zero-Shot Classifier,” the approximate conditional likelihood is taken to be proportional to the negative expected denoising error,

$$\log p_\theta(\mathbf{x}\mid \mathbf{c}) \approx - \mathbb{E}_{t,\epsilon}\left[ \|\epsilon - \epsilon_\theta(\mathbf{x}_t,\mathbf{c})\|^2 \right] + C,$$

which induces a posterior over candidate conditions by Bayes’ rule (Li et al., 2023). “Diffusion Classifiers Understand Compositionality, but Conditions Apply” keeps the same basic likelihood-proxy view but emphasizes timestep weighting, writing class posteriors as weighted sums of per-timestep denoising errors and showing that classification performance depends strongly on how those timesteps are aggregated (Jeong et al., 23 May 2025).

A third route interprets classifier guidance as reward-guided posterior sampling. “Efficient Controllable Diffusion via Optimal Classifier Guidance” introduces a binary event with

$p(y=1\mid x)=\exp(\eta r(x)),$

so that the KL-regularized optimum

$p^\star(x)=\frac{1}{Z}q_0(x)\exp(\eta r(x))$

becomes exactly a classifier-guided posterior, with guidance derived from a learned distribution over terminal rewards conditioned on an intermediate diffusion state (Oertell et al., 27 May 2025). This extends DCS from semantic class control to general reward alignment.

The literature also clarifies when the coupling is theoretically justified. “Provably Reliable Classifier Guidance through Cross-entropy Error Control” shows that, under smoothness assumptions, small conditional KL error in the classifier induces small mean-squared error in the guidance vector, whereas without smoothness one can have vanishing cross-entropy error and still obtain unusable guidance gradients (Sahu et al., 29 Jan 2026). The practical implication is that DCS depends on gradient quality, not only label accuracy.

3. Recurrent forms of coupling

A useful synthesis of the cited literature is that DCS appears in several recurring regimes rather than one canonical architecture.

Regime	Core coupling	Representative papers
External guidance	Reverse-time score is corrected by classifier gradients	(Kawar et al., 2022, Javid et al., 8 Nov 2025, Sahu et al., 29 Jan 2026, Oertell et al., 27 May 2025)
Diffusion as classifier	Denoising error or diffusion classifier scores define posteriors over classes or prompts	(Li et al., 2023, Jeong et al., 23 May 2025, Wang et al., 15 Aug 2025, Jha et al., 2024)
Shared backbone or joint model	Generator and classifier share encoder features, losses, or backbone structure	(Makarov et al., 18 Jun 2026, Deja et al., 2023, Munia et al., 19 Oct 2025)
Diffusion-assisted discriminative training	Diffusion data or representations regularize a classifier	(Huang et al., 23 Feb 2026, Li et al., 2024)
Coupled semantic-state diffusion	Signal, logits, or labels are jointly denoised	(Nurko et al., 17 Feb 2026, Belhasin et al., 25 Nov 2025)

In external-guidance systems, the classifier remains modular, but the coupling can still be deep. Robust classifier guidance replaces vanilla noisy-state classifiers with adversarially robust ones because standard cross-entropy leaves input gradients underconstrained; adding an arbitrary scalar function to every logit preserves softmax predictions while changing gradients arbitrarily (Kawar et al., 2022). More recent work further treats calibration and diversity preservation as parts of the coupling problem, adding Smooth ECE fine-tuning, entropy regularization, tilted sampling, and $f$-divergence guidance to avoid early overconfidence and vanishing gradients during denoising (Javid et al., 8 Nov 2025).

In diffusion-as-classifier systems, the classifier is implicit rather than separately trained. The original Diffusion Classifier extracts class scores directly from denoising error (Li et al., 2023), later work studies when this confers compositional discrimination and when domain shift breaks it (Jeong et al., 23 May 2025), and NoOp shows that the injected noise itself is a major hidden variable in this coupling: matching noise improves class separability and reduces the need for multi-noise ensembling (Wang et al., 15 Aug 2025).

In shared-backbone systems, the boundary between classifier and generator becomes structural. The speech work freezes a pretrained noise-conditioned classifier and trains only a decoder-like Score Subnet on top of its internal features and JEM-style gradient taps (Makarov et al., 18 Jun 2026). Joint diffusion models share a UNet encoder between denoising and classification heads (Deja et al., 2023), while the FSCIL DCS framework defines reward functions from the current classifier state and uses them to guide diffusion sampling in a mutual boosting loop (Wu et al., 4 Oct 2025).

4. Architectures and training patterns

The shared-backbone speech model is among the clearest architectural realizations of DCS. It operates on log-Mel filterbank features with FFT size $1024$, hop length $256$, and $$\nabla_{X_t}\log p_t(X_t\mid y)=\nabla_{X_t}\log p_t(X_t)+\nabla_{X_t}\log p_t(y\mid X_t).$$0 Mel bins. A noise-conditioned classifier $$\nabla_{X_t}\log p_t(X_t\mid y)=\nabla_{X_t}\log p_t(X_t)+\nabla_{X_t}\log p_t(y\mid X_t).$$1 is first trained with cross-entropy on noisy inputs, then frozen; a lightweight decoder-style “Score Subnet” is attached and trained under a Denoising Score Matching objective. The Score Subnet consumes two types of taps from the frozen classifier: forward feature taps $$\nabla_{X_t}\log p_t(X_t\mid y)=\nabla_{X_t}\log p_t(X_t)+\nabla_{X_t}\log p_t(y\mid X_t).$$2 and gradient taps

$$\nabla_{X_t}\log p_t(X_t\mid y)=\nabla_{X_t}\log p_t(X_t)+\nabla_{X_t}\log p_t(y\mid X_t).$$3

where the marginal log-density is derived from a JEM-style log-sum-exp over classifier logits. Each tap is RMS-normalized, projected, fused with cross-attention, and decoded coarse-to-fine to predict a one-channel score estimate in log-Mel space (Makarov et al., 18 Jun 2026). The classifier is still used for standard classifier guidance at sampling time, but it is no longer a standalone auxiliary network; it is the generator backbone itself.

Joint end-to-end models adopt a different pattern. “Learning Data Representations with Joint Diffusion Models” decomposes the denoiser into a shared encoder $$\nabla_{X_t}\log p_t(X_t\mid y)=\nabla_{X_t}\log p_t(X_t)+\nabla_{X_t}\log p_t(y\mid X_t).$$4, a diffusion decoder $$\nabla_{X_t}\log p_t(X_t\mid y)=\nabla_{X_t}\log p_t(X_t)+\nabla_{X_t}\log p_t(y\mid X_t).$$5, and a classifier head $$\nabla_{X_t}\log p_t(X_t\mid y)=\nabla_{X_t}\log p_t(X_t)+\nabla_{X_t}\log p_t(y\mid X_t).$$6. The classifier acts on average-pooled multiscale encoder features, and the overall objective is the sum of standard diffusion denoising terms and a standard cross-entropy loss, optionally augmented with classification losses on noisy inputs at selected timesteps (Deja et al., 2023). “Class-N-Diff” instead works in latent space: a VAE encodes dermoscopic images, a DiT denoiser produces denoised latents, and a convolutional front-end plus ResNet-101 classifier consumes either the clean latent or the denoised latent under a curriculum controlled by a gating variable $$\nabla_{X_t}\log p_t(X_t\mid y)=\nabla_{X_t}\log p_t(X_t)+\nabla_{X_t}\log p_t(y\mid X_t).$$7. The joint objective is

$$\nabla_{X_t}\log p_t(X_t\mid y)=\nabla_{X_t}\log p_t(X_t)+\nabla_{X_t}\log p_t(y\mid X_t).$$8

so the classifier regularizes the diffusion model through latent discriminability rather than through sampling-time guidance (Munia et al., 19 Oct 2025).

Some DCS systems operate entirely at training or test time without shared weights. In “Expanding the Role of Diffusion Models for Robust Classifier Training,” a frozen diffusion encoder provides representations $$\nabla_{X_t}\log p_t(X_t\mid y)=\nabla_{X_t}\log p_t(X_t)+\nabla_{X_t}\log p_t(y\mid X_t).$$9 that a robust classifier aligns to through a cosine-similarity regularizer, while diffusion-generated synthetic data are used simultaneously in adversarial training. The diffusion model is frozen throughout, and the classifier is trained with a TRADES-style objective plus a representation-alignment term (Huang et al., 23 Feb 2026). DBMEF is even more modular: a pretrained discriminative classifier first produces top-$$\tilde s(X_t,t,y)=s_\theta(X_t,t)+\gamma\,\nabla_{X_t}\log p_{\phi,t}(y\mid X_t),$$0 candidates and a confidence score, a “Confidence Protector” decides whether rethinking is needed, and a pretrained diffusion model then re-ranks only the uncertain cases by denoising-error verification over the candidate set, using positive and negative text conditions fused by a guidance-style interpolation (Li et al., 2024).

A distinct architectural branch couples semantic state variables directly. “Joint Enhancement and Classification using Coupled Diffusion Models of Signals and Logits” keeps the classifier frozen and builds two interacting diffusion processes, one over the input signal $$\tilde s(X_t,t,y)=s_\theta(X_t,t)+\gamma\,\nabla_{X_t}\log p_{\phi,t}(y\mid X_t),$$1 and one over the classifier logits $$\tilde s(X_t,t,y)=s_\theta(X_t,t)+\gamma\,\nabla_{X_t}\log p_{\phi,t}(y\mid X_t),$$2. The paper develops Parallel, Alternating, and Nested strategies, all of which let improving signal estimates refine logits and improving logits guide signal enhancement toward discriminative regions of the manifold (Nurko et al., 17 Feb 2026). “Advancing Image Classification with Discrete Diffusion Classification Modeling” pushes the same idea into label space: it diffuses class posteriors or discrete labels rather than images, modeling classification as reverse denoising from a uniform label distribution to $$\tilde s(X_t,t,y)=s_\theta(X_t,t)+\gamma\,\nabla_{X_t}\log p_{\phi,t}(y\mid X_t),$$3 (Belhasin et al., 25 Nov 2025).

5. Empirical evidence across domains

The empirical record shows that DCS is not confined to a single modality or objective. In speech generation, the shared-backbone Score Subnet achieves the same conditional FID $$\tilde s(X_t,t,y)=s_\theta(X_t,t)+\gamma\,\nabla_{X_t}\log p_{\phi,t}(y\mid X_t),$$4 as a conventional U-Net plus classifier system while using $$\tilde s(X_t,t,y)=s_\theta(X_t,t)+\gamma\,\nabla_{X_t}\log p_{\phi,t}(y\mid X_t),$$5M total parameters, $$\tilde s(X_t,t,y)=s_\theta(X_t,t)+\gamma\,\nabla_{X_t}\log p_{\phi,t}(y\mid X_t),$$6M trainable parameters, and $$\tilde s(X_t,t,y)=s_\theta(X_t,t)+\gamma\,\nabla_{X_t}\log p_{\phi,t}(y\mid X_t),$$7 GMACs/step, compared with $$\tilde s(X_t,t,y)=s_\theta(X_t,t)+\gamma\,\nabla_{X_t}\log p_{\phi,t}(y\mid X_t),$$8M total parameters and $$\tilde s(X_t,t,y)=s_\theta(X_t,t)+\gamma\,\nabla_{X_t}\log p_{\phi,t}(y\mid X_t),$$9 GMACs/step for the two-model baseline at $\gamma\ge 0$0. On unconditional generation it matches the U-Net on FID $\gamma\ge 0$1 while using only $\gamma\ge 0$2M trainable parameters instead of $\gamma\ge 0$3M (Makarov et al., 18 Jun 2026).

In image generation, the quality of the classifier side of the coupling matters measurably. Robust classifier guidance on ImageNet $\gamma\ge 0$4 improves FID from $\gamma\ge 0$5 with vanilla classifier guidance to $\gamma\ge 0$6, with precision increasing from $\gamma\ge 0$7 to $\gamma\ge 0$8, while human raters prefer robust-guided outputs at $\gamma\ge 0$9 significance (Kawar et al., 2022). More recent calibration- and divergence-aware guidance reaches FID $$\log p_\theta(\mathbf{x}\mid \mathbf{c}) \approx - \mathbb{E}_{t,\epsilon}\left[ \|\epsilon - \epsilon_\theta(\mathbf{x}_t,\mathbf{c})\|^2 \right] + C,$$0 on ImageNet $$\log p_\theta(\mathbf{x}\mid \mathbf{c}) \approx - \mathbb{E}_{t,\epsilon}\left[ \|\epsilon - \epsilon_\theta(\mathbf{x}_t,\mathbf{c})\|^2 \right] + C,$$1 with a ResNet-101 classifier and no diffusion-model retraining; in that study, Smooth ECE fine-tuning alone improves FID from $$\log p_\theta(\mathbf{x}\mid \mathbf{c}) \approx - \mathbb{E}_{t,\epsilon}\left[ \|\epsilon - \epsilon_\theta(\mathbf{x}_t,\mathbf{c})\|^2 \right] + C,$$2 to $$\log p_\theta(\mathbf{x}\mid \mathbf{c}) \approx - \mathbb{E}_{t,\epsilon}\left[ \|\epsilon - \epsilon_\theta(\mathbf{x}_t,\mathbf{c})\|^2 \right] + C,$$3, but Jensen–Shannon-regularized guidance produces the strongest overall result (Javid et al., 8 Nov 2025).

When diffusion itself is used for classification, the coupling is competitive but conditional. Diffusion Classifier obtains $$\log p_\theta(\mathbf{x}\mid \mathbf{c}) \approx - \mathbb{E}_{t,\epsilon}\left[ \|\epsilon - \epsilon_\theta(\mathbf{x}_t,\mathbf{c})\|^2 \right] + C,$$4 on ImageNet and $$\log p_\theta(\mathbf{x}\mid \mathbf{c}) \approx - \mathbb{E}_{t,\epsilon}\left[ \|\epsilon - \epsilon_\theta(\mathbf{x}_t,\mathbf{c})\|^2 \right] + C,$$5 on ObjectNet with Stable Diffusion 2.0, and reaches a Winoground average of $$\log p_\theta(\mathbf{x}\mid \mathbf{c}) \approx - \mathbb{E}_{t,\epsilon}\left[ \|\epsilon - \epsilon_\theta(\mathbf{x}_t,\mathbf{c})\|^2 \right] + C,$$6, exceeding OpenCLIP ViT-H/14’s $$\log p_\theta(\mathbf{x}\mid \mathbf{c}) \approx - \mathbb{E}_{t,\epsilon}\left[ \|\epsilon - \epsilon_\theta(\mathbf{x}_t,\mathbf{c})\|^2 \right] + C,$$7 on that benchmark (Li et al., 2023). The later compositionality study sharpens this picture: diffusion classifiers are especially strong on spatial and relational tasks and can be extremely accurate on the self-generated Self-Bench domain, but cross-domain drops are substantial, and SD3-m, despite stronger generative compositionality, attains lower discriminative accuracy ($$\log p_\theta(\mathbf{x}\mid \mathbf{c}) \approx - \mathbb{E}_{t,\epsilon}\left[ \|\epsilon - \epsilon_\theta(\mathbf{x}_t,\mathbf{c})\|^2 \right] + C,$$8) than earlier versions ($$\log p_\theta(\mathbf{x}\mid \mathbf{c}) \approx - \mathbb{E}_{t,\epsilon}\left[ \|\epsilon - \epsilon_\theta(\mathbf{x}_t,\mathbf{c})\|^2 \right] + C,$$9) (Jeong et al., 23 May 2025). NoOp shows that some of the variance arises from the forward corruption itself: on ImageNet, a 5-noise ensemble costs $p(y=1\mid x)=\exp(\eta r(x)),$0 hours of inference on 32 V100 GPUs, whereas NoOp keeps single-sample inference at $p(y=1\mid x)=\exp(\eta r(x)),$1 hours and improves accuracy from $p(y=1\mid x)=\exp(\eta r(x)),$2 to $p(y=1\mid x)=\exp(\eta r(x)),$3 in the 4-shot setting (Wang et al., 15 Aug 2025).

Training-time DCS also yields consistent gains. In continual personalization of text-to-image diffusion models, adding diffusion-classifier scores to EWC and distillation produces clear improvements over C-LoRA and non-DC baselines; on Custom Concept (6 tasks), DSC EWC DC reaches KID $p(y=1\mid x)=\exp(\eta r(x)),$4, $p(y=1\mid x)=\exp(\eta r(x)),$5, and $p(y=1\mid x)=\exp(\eta r(x)),$6, compared with C-LoRA’s KID $p(y=1\mid x)=\exp(\eta r(x)),$7, $p(y=1\mid x)=\exp(\eta r(x)),$8, and $p(y=1\mid x)=\exp(\eta r(x)),$9 (Jha et al., 2024). In robust classifier training, adding diffusion representation alignment on top of diffusion-generated synthetic data improves ImageNet ConvNeXt-B from $p^\star(x)=\frac{1}{Z}q_0(x)\exp(\eta r(x))$0 clean/AutoAttack to $p^\star(x)=\frac{1}{Z}q_0(x)\exp(\eta r(x))$1, and ViT-B/16 from $p^\star(x)=\frac{1}{Z}q_0(x)\exp(\eta r(x))$2 to $p^\star(x)=\frac{1}{Z}q_0(x)\exp(\eta r(x))$3 (Huang et al., 23 Feb 2026). DBMEF, a purely test-time enhancement framework, improves ResNet-50 by $p^\star(x)=\frac{1}{Z}q_0(x)\exp(\eta r(x))$4 on ImageNet and $p^\star(x)=\frac{1}{Z}q_0(x)\exp(\eta r(x))$5 on ImageNet-A through confidence-gated diffusion re-ranking (Li et al., 2024). Label-space diffusion classification likewise improves most strongly when uncertainty is highest; under Strong Aug on ImageNet with $p^\star(x)=\frac{1}{Z}q_0(x)\exp(\eta r(x))$6 training data at resolution $p^\star(x)=\frac{1}{Z}q_0(x)\exp(\eta r(x))$7, DiDiCM-CP raises top-1 accuracy from $p^\star(x)=\frac{1}{Z}q_0(x)\exp(\eta r(x))$8 to $p^\star(x)=\frac{1}{Z}q_0(x)\exp(\eta r(x))$9 and top-5 from $f$0 to $f$1 (Belhasin et al., 25 Nov 2025).

Joint generative-discriminative models also show that DCS need not sacrifice either side outright. The shared-encoder joint diffusion model reaches $f$2 classification accuracy and FID $f$3 on CIFAR-10, outperforming earlier hybrid methods on both tasks (Deja et al., 2023). In FSCIL, the explicitly named DCS framework uses reward-aligned generation to improve both retention and new-class learning and is reported to achieve state-of-the-art performance on FSCIL benchmarks through its mutual boosting loop (Wu et al., 4 Oct 2025).

6. Limitations, misconceptions, and open directions

A persistent misconception is that DCS is synonymous with standard classifier guidance. The cited literature indicates a broader landscape: guidance is only one regime, alongside diffusion-as-classifier, shared-backbone generators, reward-aligned co-evolution, diffusion-assisted robust training, and coupled signal-logit or label-space diffusion (Makarov et al., 18 Jun 2026, Jha et al., 2024, Belhasin et al., 25 Nov 2025). A related misconception is that classifier accuracy alone is sufficient. Multiple papers reject this directly: standard cross-entropy does not constrain input gradients (Kawar et al., 2022), early overconfidence can make guidance gradients vanish during denoising (Javid et al., 8 Nov 2025), and even vanishing cross-entropy error need not imply accurate guidance vectors without smoothness assumptions (Sahu et al., 29 Jan 2026).

The main practical limitation is computational. Diffusion classifiers remain expensive because each candidate label requires denoising-error evaluation across timesteps, and prior methods often relied on many sampled noises to stabilize predictions (Li et al., 2023). Even after NoOp reduces ensemble dependence, the method still requires a forward pass for every candidate class and remains expensive relative to CLIP-like models (Wang et al., 15 Aug 2025). Coupled signal-logit diffusion improves noisy classification, but Parallel, Alternating, and Nested strategies substantially increase neural function evaluations relative to sequential enhancement (Nurko et al., 17 Feb 2026). Training-time DCS can also introduce constant-factor runtime overhead, as in replay-free continual personalization where DC-score computation is the main cost rather than storage (Jha et al., 2024).

Another limitation is domain specificity. The speech backbone-reuse model is validated only on SC09 in log-Mel space (Makarov et al., 18 Jun 2026). Class-N-Diff is specific to binary dermoscopic diagnosis, and its fairness claims are based on subgroup-performance tables rather than formal fairness criteria; improvements are mixed across accuracy, AUC, and sensitivity on DDI subgroups (Munia et al., 19 Oct 2025). Diffusion classifiers can be near-ceiling on self-generated domains yet degrade sharply under cross-domain transfer, especially for SD3-m, which indicates that generative competence and discriminative transfer are not equivalent (Jeong et al., 23 May 2025).

The current literature suggests several directions rather than a settled endpoint. Adaptive timestep weighting appears crucial for discriminative readout from diffusion models (Jeong et al., 23 May 2025). Noise optimization indicates that the forward corruption itself is a controllable interface for DCS rather than a fixed nuisance variable (Wang et al., 15 Aug 2025). Shared-backbone systems suggest that discriminative encoders may often already contain score-relevant structure (Makarov et al., 18 Jun 2026), while reliable-guidance theory suggests that future classifier designs should target smoothness and gradient quality explicitly, not only cross-entropy (Sahu et al., 29 Jan 2026). Taken together, these works suggest that the most mature form of DCS is neither “classifier plus diffusion” nor “diffusion replaces classifier,” but a spectrum of designs in which representation reuse, gradient quality, calibration, reward shaping, and timestep-dependent control are treated as parts of a single coupled system.