Conditional Diffusion Model (CDiff) Overview

• CDiff is a class of diffusion-based generative models that conditions both the forward and reverse processes on auxiliary information such as noisy signals or coarse observations.
• The approach adapts conditioning mechanisms across diverse domains—from speech enhancement to graph synthesis—by modifying mathematical formulations and leveraging architectures like U-Net and transformers.
• Training and sampling in CDiff use noise-prediction objectives with tailored conditional strategies, resulting in improved fidelity, robustness, and performance across various empirical applications.

Searching arXiv for the provided paper and closely related conditional diffusion work to ground the article. Conditional Diffusion Model (CDiff) denotes a class of diffusion-based generative models in which sampling or denoising is explicitly conditioned on auxiliary information. In the literature represented here, the term is used in several related but domain-specific senses: as a generalized diffusion probabilistic model for speech enhancement that incorporates the observed noisy speech signal into both the diffusion and reverse processes (Lu et al., 2022); as a conditional generative model that reconstructs fine-grained environment-and-channel fingerprints maps from coarse observations (Jin et al., 12 May 2025); as a receiver-side conditional decoder for wireless semantic communications (Letafati et al., 26 Sep 2025); and as a broader design pattern in applications such as longitudinal MRI generation (Dao et al., 2024), cortical thickness trajectory prediction (Xiao et al., 2024), target speech extraction (Kamo et al., 2023), electrical impedance tomography (Shi et al., 10 Jan 2025), graph synthesis (Tsai et al., 2023), and planning in offline meta-reinforcement learning (Ni et al., 2023). Across these uses, CDiff replaces unconditional generation with generation under side information, either by modifying the forward diffusion process, the reverse denoising process, or both.

1. Conceptual scope and defining characteristics

In standard diffusion probabilistic modeling, a forward process gradually corrupts data and a reverse process learns to denoise it. The conditional variant augments this construction with side information such as noisy observations, class labels, speaker clues, spatial context, temporal encodings, or task embeddings. The resulting model learns a conditional distribution rather than an unconditional one. For speech enhancement, this is formalized as a “conditional diffusion probabilistic model” that incorporates the observed noisy speech signal $y$ into both the diffusion and reverse processes, with the stated goal of adapting to non-Gaussian real noises in the estimated speech signal (Lu et al., 2022).

A recurring distinction in the literature is where conditioning enters the model. In some formulations, conditioning appears only in the reverse process, as in the high-resolution EnvCF reconstruction model that learns $$p(\boldsymbol{\mathsf{F}} \mid \boldsymbol{\dot{\mathsf{F}}})$$ using a conditional denoising network (Jin et al., 12 May 2025). In others, conditioning appears in both the forward and reverse processes, as in the speech-enhancement CDiff and the semantic communications decoder, where the noising trajectory itself is altered by the condition (Lu et al., 2022, Letafati et al., 26 Sep 2025). A further variation is sparse or indirect conditioning: conditional sampling can be performed using an unconditional diffusion model and limited explicit guidance, with model predictive control used to approximate unavailable guidance at unguided steps (Shen et al., 2022).

This diversity suggests that “CDiff” is best understood as a family of conditional generative diffusion constructions rather than a single canonical architecture. A plausible implication is that the most informative taxonomy separates methods by conditioning locus: forward-conditioned, reverse-conditioned, and guidance-conditioned variants.

2. Mathematical formulations

For speech enhancement, CDiff is introduced as a generalized formulation of the diffusion probabilistic model. The conditional forward process is

$$q_{\text{cdiff}}(x_t|x_0, y) = \mathcal{N}\big(x_t; (1-m_t)\sqrt{\bar{\alpha}_t} x_0 + m_t\sqrt{\bar{\alpha}_t} y, \delta_t I \big) \tag{8}$$

with

$$\delta_t = (1-\bar{\alpha}_t) - m_t^2 \bar{\alpha}_t \tag{9}$$

where $m_t$ smoothly transitions from $0$ to $1$, so that early steps are more influenced by $x_0$ and later steps by $y$ (Lu et al., 2022). The corresponding conditional reverse process is

$$p_{\text{cdiff}}(x_{t-1} | x_t, y) = \mathcal{N}(x_{t-1}; \mu_\theta(x_t, y, t), \tilde{\delta}_t I) \tag{12}$$

with

$$p(\boldsymbol{\mathsf{F}} \mid \boldsymbol{\dot{\mathsf{F}}})$$0

where the network $$p(\boldsymbol{\mathsf{F}} \mid \boldsymbol{\dot{\mathsf{F}}})$$1 predicts the noise, including non-Gaussian components (Lu et al., 2022).

The associated conditional ELBO is given as

$$p(\boldsymbol{\mathsf{F}} \mid \boldsymbol{\dot{\mathsf{F}}})$$2

which leads to a conditional noise-prediction loss,

$$p(\boldsymbol{\mathsf{F}} \mid \boldsymbol{\dot{\mathsf{F}}})$$3

so the target contains both standard Gaussian noise and the actual noise in $$p(\boldsymbol{\mathsf{F}} \mid \boldsymbol{\dot{\mathsf{F}}})$$4 (Lu et al., 2022).

Other CDiff instances retain the DDPM objective but condition only the reverse model. In EnvCDiff, the forward process is Gaussian,

$$p(\boldsymbol{\mathsf{F}} \mid \boldsymbol{\dot{\mathsf{F}}})$$5

with

$$p(\boldsymbol{\mathsf{F}} \mid \boldsymbol{\dot{\mathsf{F}}})$$6

while the conditional reverse process is

$$p(\boldsymbol{\mathsf{F}} \mid \boldsymbol{\dot{\mathsf{F}}})$$7

and training minimizes

$$p(\boldsymbol{\mathsf{F}} \mid \boldsymbol{\dot{\mathsf{F}}})$$8

(Jin et al., 12 May 2025).

In wireless semantic communications, the conditional diffusion decoder uses a forward process conditioned on semantic information,

$$p(\boldsymbol{\mathsf{F}} \mid \boldsymbol{\dot{\mathsf{F}}})$$9

with the reverse update

$$q_{\text{cdiff}}(x_t|x_0, y) = \mathcal{N}\big(x_t; (1-m_t)\sqrt{\bar{\alpha}_t} x_0 + m_t\sqrt{\bar{\alpha}_t} y, \delta_t I \big) \tag{8}$$0

and

$$q_{\text{cdiff}}(x_t|x_0, y) = \mathcal{N}\big(x_t; (1-m_t)\sqrt{\bar{\alpha}_t} x_0 + m_t\sqrt{\bar{\alpha}_t} y, \delta_t I \big) \tag{8}$$1

(Letafati et al., 26 Sep 2025).

These formulations show two common mathematical motifs. First, conditioning may shift the mean of the forward kernel. Second, the denoiser or score estimator typically takes the noisy state, the condition, and the timestep as inputs, with training reduced to a noise-prediction or denoised-sample regression objective.

3. Conditioning mechanisms and architectural patterns

The conditioning variable can be an observed corrupted signal, a coarse-resolution field, a clue identifying a target source, a baseline scan, a temporal interval, or a task representation. In the speech-enhancement CDiff, $$q_{\text{cdiff}}(x_t|x_0, y) = \mathcal{N}\big(x_t; (1-m_t)\sqrt{\bar{\alpha}_t} x_0 + m_t\sqrt{\bar{\alpha}_t} y, \delta_t I \big) \tag{8}$$2 is the observed noisy speech signal, and the model is designed so that the latent variable trajectory traverses between clean and observed noisy data (Lu et al., 2022). In EnvCDiff, the condition is the low-resolution EnvCF map $$q_{\text{cdiff}}(x_t|x_0, y) = \mathcal{N}\big(x_t; (1-m_t)\sqrt{\bar{\alpha}_t} x_0 + m_t\sqrt{\bar{\alpha}_t} y, \delta_t I \big) \tag{8}$$3, provided at every denoising step as side information to ensure that hallucinated fine details are consistent with known coarse structure (Jin et al., 12 May 2025). In the semantic communications decoder, the condition is the noisy semantic latent received through the channel, incorporated to “steer” the decoding process toward the intended semantics (Letafati et al., 26 Sep 2025).

Architecturally, several patterns recur. A U-Net backbone is used in precipitation downscaling, with the concatenated low-resolution input, topographic information, and noise-corrupted high-resolution precipitation image fed into the network at each timestep (Lyu et al., 2024). An attention-based 3D UNet is used for longitudinal MRI generation, where the noisy image, conditioning MRI, and time-visit encoding are combined through latent summation,

$$q_{\text{cdiff}}(x_t|x_0, y) = \mathcal{N}\big(x_t; (1-m_t)\sqrt{\bar{\alpha}_t} x_0 + m_t\sqrt{\bar{\alpha}_t} y, \delta_t I \big) \tag{8}$$4

(Dao et al., 2024). In cortical thickness trajectory prediction, baseline CTh, demographics, diagnosis, and time difference are concatenated along the channel dimension of a 1D Attention U-Net input tensor (Xiao et al., 2024). In tabular building-energy imputation, a single MLP takes noisy target features, condition features, masks, and timestep embedding, with condition features kept fixed and visible at all timesteps (Sinha et al., 4 Nov 2025). In wireless channel identification, a transformer conditions its attention and MLP mechanisms on scenario label $$q_{\text{cdiff}}(x_t|x_0, y) = \mathcal{N}\big(x_t; (1-m_t)\sqrt{\bar{\alpha}_t} x_0 + m_t\sqrt{\bar{\alpha}_t} y, \delta_t I \big) \tag{8}$$5 and timestep $$q_{\text{cdiff}}(x_t|x_0, y) = \mathcal{N}\big(x_t; (1-m_t)\sqrt{\bar{\alpha}_t} x_0 + m_t\sqrt{\bar{\alpha}_t} y, \delta_t I \big) \tag{8}$$6 to capture hidden channel features in multiple latent noise spaces (Li et al., 14 Jun 2025).

A notable alternative to explicit conditional training is training-free energy guidance. FreeDoM uses off-the-shelf pre-trained networks to construct time-independent energy functions, estimates a clean image from the current noisy state, and applies an energy gradient step during sampling,

$$q_{\text{cdiff}}(x_t|x_0, y) = \mathcal{N}\big(x_t; (1-m_t)\sqrt{\bar{\alpha}_t} x_0 + m_t\sqrt{\bar{\alpha}_t} y, \delta_t I \big) \tag{8}$$7

which turns an unconditional diffusion model into a conditional one without extra training (Yu et al., 2023). Relatedly, “Conditional Diffusion with Less Explicit Guidance via Model Predictive Control” studies conditional sampling using an unconditional diffusion model with limited explicit guidance, approximating missing guidance by differentiable forward simulation and backpropagation (Shen et al., 2022).

4. Training, sampling, and inference workflows

The generic training pattern in CDiff methods is to sample a clean datum and condition, choose a timestep, corrupt the datum according to the forward process, and optimize a denoising objective. In speech enhancement, training samples $$q_{\text{cdiff}}(x_t|x_0, y) = \mathcal{N}\big(x_t; (1-m_t)\sqrt{\bar{\alpha}_t} x_0 + m_t\sqrt{\bar{\alpha}_t} y, \delta_t I \big) \tag{8}$$8, $$q_{\text{cdiff}}(x_t|x_0, y) = \mathcal{N}\big(x_t; (1-m_t)\sqrt{\bar{\alpha}_t} x_0 + m_t\sqrt{\bar{\alpha}_t} y, \delta_t I \big) \tag{8}$$9, and $$\delta_t = (1-\bar{\alpha}_t) - m_t^2 \bar{\alpha}_t \tag{9}$$0, generates $$\delta_t = (1-\bar{\alpha}_t) - m_t^2 \bar{\alpha}_t \tag{9}$$1 using the interpolated conditional forward process, predicts combined noise with $$\delta_t = (1-\bar{\alpha}_t) - m_t^2 \bar{\alpha}_t \tag{9}$$2, and minimizes $$\delta_t = (1-\bar{\alpha}_t) - m_t^2 \bar{\alpha}_t \tag{9}$$3 (Lu et al., 2022). In speech-enhancement inference, sampling begins with

$$\delta_t = (1-\bar{\alpha}_t) - m_t^2 \bar{\alpha}_t \tag{9}$$4

and iteratively computes

$$\delta_t = (1-\bar{\alpha}_t) - m_t^2 \bar{\alpha}_t \tag{9}$$5

until producing the clean estimate $$\delta_t = (1-\bar{\alpha}_t) - m_t^2 \bar{\alpha}_t \tag{9}$$6 (Lu et al., 2022).

In conditional tabular imputation, the sample is partitioned into observed variables $$\delta_t = (1-\bar{\alpha}_t) - m_t^2 \bar{\alpha}_t \tag{9}$$7 and unobserved variables $$\delta_t = (1-\bar{\alpha}_t) - m_t^2 \bar{\alpha}_t \tag{9}$$8. Diffusion is applied only to $$\delta_t = (1-\bar{\alpha}_t) - m_t^2 \bar{\alpha}_t \tag{9}$$9, while $m_t$0 remains fixed and conditions each reverse step. Dynamic masking randomizes the observed/target split during training, and inference begins from random noise on masked features before iterative denoising conditioned on the fixed observed part (Sinha et al., 4 Nov 2025).

In target speech extraction, the forward SDE corrupts clean target speech $m_t$1 toward the mixture $m_t$2,

$m_t$3

and the reverse SDE samples target speech conditioned on mixture $m_t$4 and clue $m_t$5,

$m_t$6

(Kamo et al., 2023). That work also introduces ensemble inference, which repeats sampling multiple times with different random seeds and averages the outputs to reduce extraction errors caused by the diffusion process (Kamo et al., 2023).

In the EIT reconstruction framework CDMVC, the condition is an initial pre-imaged conductivity reconstruction $m_t$7, concatenated channel-wise with the noisy conductivity sample at each step. Sampling includes an additional optimization stage every several steps to enforce voltage consistency using a forward voltage constraint network,

$m_t$8

(Shi et al., 10 Jan 2025). This illustrates a broader pattern in CDiff applications: conditional generation is often combined with domain-specific constraints during sampling rather than relying on denoising alone.

5. Empirical behavior across domains

The empirical claims in the literature emphasize improvements in fidelity, robustness, and generalization under the relevant task-specific metrics.

For speech enhancement on VoiceBank + DEMAND, the implemented CDiff variant, CDiffuSE, outperforms its unconditional Diffusion baseline DiffuSE by a clear margin in an ablation study, with PESQ improving from $m_t$9 for the Large model, and perceptual metrics such as CSIG and COVL also improving (Lu et al., 2022). The same study reports that compared with SEGAN, DSEGAN, and SE-Flow, CDiffuSE performs better, especially in PESQ, and that under unseen noise, trained on VoiceBank-DEMAND and tested on CHiME-4, CDiffuSE degrades less and achieves best scores on all metrics among the compared methods (Lu et al., 2022).

For sparse-guidance conditional sampling, model predictive control approximations to guidance have high cosine similarity to real guides, above $0$0 even when $0$1 steps out of $0$2, and above $0$3 even for $0$4. Baselines using image-level guidance gradients such as CLIP are near $0$5, almost orthogonal to the true conditional guide (Shen et al., 2022). In Stable Diffusion experiments with only 5 explicit guidance steps, adding 3 MPC-approximated guidance steps reduces FID to Reference from $0$6 to $0$7, and FID to Gold Standard from $0$8 to $0$9 (Shen et al., 2022).

For precipitation downscaling, the bias-informed conditional diffusion model achieves highly accurate results in an 8 times downscaling setting, outperforming previous deterministic methods (Lyu et al., 2024). The reported table gives RMSE $1$0, Corr $1$1, and Bias $1$2 for the full model, compared with RMSE $1$3, Corr $1$4, and Bias $1$5 for SRCNN (Lyu et al., 2024).

For longitudinal MRI generation, the conditional diffusion model reaches FID $1$6 and SSIM $1$7, compared with FID $1$8 and SSIM $1$9 for Med-DDPM and FID $x_0$0 and SSIM $x_0$1 for VAE (Dao et al., 2024). For cortical thickness trajectory prediction, CDiff yields MAE $x_0$2 over all subjects, compared with $x_0$3 for GRUD, $x_0$4 for LSTM-T, and $x_0$5 for cFSGL, while Bland-Altman analysis shows near-zero bias with narrow 95% confidential interval compared to ground truth in 6–36 months (Xiao et al., 2024).

For target speech extraction on Libri2mix, Diff-TSE-MT+Ens. reports PESQ $x_0$6, ESTOI $x_0$7, and SI-SDR $x_0$8, improving over the discriminative TSE baseline at PESQ $x_0$9, ESTOI $y$0, and SI-SDR $y$1 (Kamo et al., 2023). For EIT reconstruction, CDMVC attains RE $y$2, SSIM $y$3, PSNR $y$4, and DR $y$5 in simulation data, with an ablation showing RE decreasing from $y$6 to $y$7 and SSIM increasing from $y$8 to $y$9 after adding voltage consistency during sampling (Shi et al., 10 Jan 2025).

For dual-conditional graph synthesis, CDGraph reports validity $$p_{\text{cdiff}}(x_{t-1} | x_t, y) = \mathcal{N}(x_{t-1}; \mu_\theta(x_t, y, t), \tilde{\delta}_t I) \tag{12}$$0 on Facebook and BlogCatalog in the summary table, with edge error $$p_{\text{cdiff}}(x_{t-1} | x_t, y) = \mathcal{N}(x_{t-1}; \mu_\theta(x_t, y, t), \tilde{\delta}_t I) \tag{12}$$1 and clustering MMD $$p_{\text{cdiff}}(x_{t-1} | x_t, y) = \mathcal{N}(x_{t-1}; \mu_\theta(x_t, y, t), \tilde{\delta}_t I) \tag{12}$$2, outperforming SPECTRE, GSM, EDGE, and DiGress on the reported metrics (Tsai et al., 2023). For offline meta-reinforcement learning, MetaDiffuser reports $$p_{\text{cdiff}}(x_{t-1} | x_t, y) = \mathcal{N}(x_{t-1}; \mu_\theta(x_t, y, t), \tilde{\delta}_t I) \tag{12}$$3 on Cheetah-Dir versus $$p_{\text{cdiff}}(x_{t-1} | x_t, y) = \mathcal{N}(x_{t-1}; \mu_\theta(x_t, y, t), \tilde{\delta}_t I) \tag{12}$$4 for Prompt-DT and $$p_{\text{cdiff}}(x_{t-1} | x_t, y) = \mathcal{N}(x_{t-1}; \mu_\theta(x_t, y, t), \tilde{\delta}_t I) \tag{12}$$5 for CORRO (Ni et al., 2023). These results suggest that conditional diffusion is not confined to perceptual synthesis; it also functions as a structured conditional generator for trajectories, graphs, and inverse problems.

6. Relation to adjacent conditional diffusion paradigms

Several papers clarify what CDiff is not, or what alternatives exist. ShiftDDPMs argues that most existing conditional DDPMs relate conditions only to the reverse process and fit it to the reversal of an unconditional forward process, which limits condition modeling and generation to a small time window (Zhang et al., 2023). Its proposal is to introduce conditions into the forward process by allocating an exclusive diffusion trajectory for each condition using shifting rules, thereby dispersing condition modeling across all timesteps (Zhang et al., 2023). This provides a useful conceptual contrast with reverse-only conditioning.

FreeDoM occupies another adjacent position. It is not a trained conditional diffusion model in the usual sense but a training-free conditional diffusion model that leverages time-independent energy functions constructed from off-the-shelf pre-trained networks (Yu et al., 2023). This framework shows that conditional behavior can be induced without retraining the generative model, although the paper notes limitations in inference cost, fine-structure control, and conflicting multi-condition guidance (Yu et al., 2023).

“Conditional Diffusion with Less Explicit Guidance via Model Predictive Control” addresses a separate misconception: that high-quality conditional diffusion necessarily requires explicit guidance at every step. The reported results indicate that conditional sampling can be achieved with minimal explicit guidance by simulating forward with unconditional diffusion and backpropagating explicit guidance at reachable timesteps (Shen et al., 2022). This suggests that the operational notion of a conditional diffusion model includes not only models trained directly with conditions but also systems that recover conditionality through guidance approximations.

A further extension appears in score-based continuous-time discrete diffusion, where diffusion is generalized to categorical data via a stochastic jump process and continuous-time Markov chain, with learning based on matching conditional marginal distributions (Sun et al., 2022). Although this paper is not framed around “CDiff” as a named model, it reinforces the broader point that conditional diffusion ideas are portable across continuous, discrete, and mixed-feature data regimes.

7. Applications, limitations, and interpretive issues

The applications represented here are heterogeneous: speech enhancement (Lu et al., 2022), target speech extraction (Kamo et al., 2023), precipitation downscaling (Lyu et al., 2024), longitudinal medical image generation (Dao et al., 2024), cortical thickness trajectory prediction (Xiao et al., 2024), electrical impedance tomography (Shi et al., 10 Jan 2025), environment-aware channel map super-resolution (Jin et al., 12 May 2025), wireless semantic communications (Letafati et al., 26 Sep 2025), building-energy data imputation (Sinha et al., 4 Nov 2025), channel scenario identification (Li et al., 14 Jun 2025), social graph synthesis (Tsai et al., 2023), and offline meta-RL planning (Ni et al., 2023). The unifying property is conditional generation under structured side information.

Several limitations are explicit. In the MPC-guidance framework, backpropagation through unconditional diffusion steps requires memory linear in the number of denoising steps (Shen et al., 2022). In FreeDoM, inference cost is somewhat higher than fully trained models, energy guidance may be coarse on large domains, and simple weighted sums of multiple conditions may yield imperfect results when conditions conflict (Yu et al., 2023). In graph synthesis, the method is specialized to dual-conditional social graphs and uses a classifier-guided denoising setup shaped by homophily and contagion assumptions (Tsai et al., 2023). In portfolio optimization, transaction costs can erode the advantage of diffusion-generated return estimates unless explicitly modeled in the optimization problem (Gao et al., 26 Sep 2025).

One common misconception is to treat CDiff as a single standardized algorithm. The surveyed work does not support that reading. Instead, “Conditional Diffusion Model” is a recurring designation for a modeling principle whose concrete realization depends on domain structure and conditioning semantics. Another misconception is that conditioning always enters through labels or prompts. In the speech-enhancement CDiff, the condition is the noisy observation itself (Lu et al., 2022); in EIT it is an initial reconstruction and voltage consistency constraint (Shi et al., 10 Jan 2025); in meta-RL it is a learned task context (Ni et al., 2023). This broader view is important for interpreting claims across papers that share the “CDiff” label but not a common architecture.

A plausible implication is that the main scientific contribution of CDiff-style models lies less in the diffusion backbone itself than in the way conditional information is aligned with the forward corruption law, reverse denoiser, and sampling-time constraints. In that sense, CDiff is best understood as a conditional generative interface between diffusion modeling and domain structure.