Conditional DDPM: Denoising Probabilistic Model

• DDPM is a latent-variable generative model that employs a fixed Gaussian forward diffusion and learns a reverse denoising chain conditioned on auxiliary inputs.
• It supports classical sample generation and conditional density estimation, making it applicable to tasks like image translation, anomaly detection, and surrogate modeling.
• The method combines empirical convergence guarantees with robust conditioning strategies, demonstrating practical effectiveness across domains such as medical imaging and remote sensing.

Searching arXiv for the specified and closely related DDPM papers to ground the article in current references. In the literature summarized here, DDPM denotes a Denoising Diffusion Probabilistic Model: a latent-variable generative model that defines a fixed Gaussian forward diffusion from data to near-isotropic noise and learns a reverse denoising chain that reconstructs samples from that noise. In its conditional form, the reverse kernel is conditioned on auxiliary information, so the model represents a conditional distribution such as $p(x_0 \mid c)$ rather than only an unconditional data prior. This framework supports both classical sample generation and conditional density estimation, including a Transformer-based density emulator that conditions on arbitrary subsets of inputs and produces probabilistic, potentially non-Gaussian and multimodal outputs (Leung et al., 2024). Recent theory further gives near-optimal convergence guarantees for DDPM under general distributional assumptions and clarifies its relation to score-based sampling and DDIM (Jiao et al., 31 Oct 2025).

1. Core probabilistic formulation

A DDPM specifies a forward noising Markov chain on $x_0 \in \mathbb R^d$ by

$$q(x_t \mid x_{t-1}) = \mathcal N\bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I\bigr),$$

with $\alpha_t \equiv 1-\beta_t$ and $\bar\alpha_t \equiv \prod_{s=1}^t \alpha_s$. This yields the closed-form marginal

$$q(x_t \mid x_0)=\mathcal N\bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\bigr),$$

or equivalently

$$x_t=\sqrt{\bar\alpha_t}\,x_0+\sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$

As $T \to \infty$ and $\bar\alpha_T \to 0$, the terminal state $x_T$ approaches $x_0 \in \mathbb R^d$0 (Leung et al., 2024, Jiao et al., 31 Oct 2025).

The learned reverse process is a Markov chain with conditional Gaussian transitions

$x_0 \in \mathbb R^d$1

where $x_0 \in \mathbb R^d$2 denotes conditioning information. A common parameterization writes the mean through either a score estimate or a noise-prediction network. In the score-based form summarized in the convergence analysis,

$x_0 \in \mathbb R^d$3

while in the noise-prediction form the model learns $x_0 \in \mathbb R^d$4 and reconstructs the posterior mean from that estimate (Jiao et al., 31 Oct 2025).

Training is usually presented as variational inference on an ELBO. For the standard Gaussian choices, the objective reduces, up to constants, to the simplified DDPM loss

$x_0 \in \mathbb R^d$5

This is the squared-error noise-matching objective implemented by the Transformer-plus-diffusion density estimator, where the diffusion head is described simply as a denoising diffusion head trained with the standard DDPM squared-error loss (Leung et al., 2024).

2. Conditionalization strategies

The most common conditional construction leaves the forward diffusion unchanged and injects the condition only into the reverse model. In image-to-image settings, the condition is typically concatenated channel-wise with the noised state $x_0 \in \mathbb R^d$6 at the U-Net input, and time embeddings are injected into residual blocks. This pattern appears in thermal-to-visible face translation, CBCT-to-sCT synthesis, seismic ground-roll attenuation, and hyperspectral–multispectral fusion, with conditioning supplied respectively by thermal images, CBCT slices, noisy seismic gathers, and paired $x_0 \in \mathbb R^d$7 observations (Nair et al., 2022, Peng et al., 2023, Li et al., 2024, Shi et al., 2023).

A more structured conditional mechanism appears in the Transformer-based density emulator. The input is a sequence of $x_0 \in \mathbb R^d$8 pairs with one distinguished “request” token analogous to $x_0 \in \mathbb R^d$9. The final hidden state of that request token, $$q(x_t \mid x_{t-1}) = \mathcal N\bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I\bigr),$$0, is passed to the diffusion head, while the remaining tokens act as known or masked conditions. The denoiser embeds the time index $$q(x_t \mid x_{t-1}) = \mathcal N\bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I\bigr),$$1 and injects $$q(x_t \mid x_{t-1}) = \mathcal N\bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I\bigr),$$2 at every layer, for example by FiLM or by concatenation to the per-step time embedding, thereby turning an unconditional $$q(x_t \mid x_{t-1}) = \mathcal N\bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I\bigr),$$3 into a conditional $$q(x_t \mid x_{t-1}) = \mathcal N\bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I\bigr),$$4 (Leung et al., 2024).

Several papers modify the forward process itself rather than only the reverse chain. In wireless reconstruction, the receiver-side degraded observation $$q(x_t \mid x_{t-1}) = \mathcal N\bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I\bigr),$$5 is injected into a conditional forward noising law

$$q(x_t \mid x_{t-1}) = \mathcal N\bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I\bigr),$$6

with a time-dependent interpolation coefficient $$q(x_t \mid x_{t-1}) = \mathcal N\bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I\bigr),$$7 rising from approximately $$q(x_t \mid x_{t-1}) = \mathcal N\bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I\bigr),$$8 to $$q(x_t \mid x_{t-1}) = \mathcal N\bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1}, \beta_t I\bigr),$$9 (Letafati et al., 2023). In ShiftDDPM, each condition $\alpha_t \equiv 1-\beta_t$0 is assigned a time-dependent shift trajectory $\alpha_t \equiv 1-\beta_t$1 and the forward kernel becomes

$\alpha_t \equiv 1-\beta_t$2

with $\alpha_t \equiv 1-\beta_t$3. That construction is explicitly motivated by the claim that reverse-only conditioning limits condition modeling and generation to a small time window (Zhang et al., 2023).

A further variant is multichannel conditioning for anomaly detection. MCDDPM forms a context representation from a clean image and a bridge-network latent code, then injects that context through cross-attention in the U-Net bottleneck. The model uses two noisy views of the same healthy image, one fully noised and one patch-noised, to enrich the conditional signal while keeping computational cost and memory on par with DDPM, pDDPM, and mDDPM (Trivedi et al., 2024).

3. Transformer-based conditional density estimation

The Transformer-plus-diffusion model of “Estimating Probability Densities with Transformer and Denoising Diffusion” is motivated by a limitation of regression-trained Transformers: they do not estimate the probability density distribution, even though many scientific tasks require full probabilistic outputs whose distributions may be non-Gaussian and multimodal. The paper therefore attaches a denoising diffusion head to a Transformer encoder and uses the resulting model as a conditional density estimator for arbitrary combinations of inputs and outputs (Leung et al., 2024).

The architecture is explicitly BERT-style. The Transformer comprises two multi-head attention layers, with $\alpha_t \equiv 1-\beta_t$4 heads and $\alpha_t \equiv 1-\beta_t$5 dimensions per head, giving $\alpha_t \equiv 1-\beta_t$6-dimensional hidden vectors. The embedding size for each token/value pair is also $\alpha_t \equiv 1-\beta_t$7. The request token’s final state is routed to the diffusion head, and the target being denoised is a one-dimensional vector rather than an image; accordingly, there is no U-Net or spatial hierarchy in the denoiser. Instead, the DDPM head is an MLP conditioned on the Transformer representation. The full model has $\alpha_t \equiv 1-\beta_t$8 M parameters, of which $\alpha_t \equiv 1-\beta_t$9 M belong to the DDPM head (Leung et al., 2024).

The diffusion process uses $\bar\alpha_t \equiv \prod_{s=1}^t \alpha_s$0 steps and a sigmoid noise schedule over $\bar\alpha_t \equiv \prod_{s=1}^t \alpha_s$1,

$\bar\alpha_t \equiv \prod_{s=1}^t \alpha_s$2

with $\bar\alpha_t \equiv \prod_{s=1}^t \alpha_s$3, $\bar\alpha_t \equiv \prod_{s=1}^t \alpha_s$4, and $\bar\alpha_t \equiv \prod_{s=1}^t \alpha_s$5 mapping $\bar\alpha_t \equiv \prod_{s=1}^t \alpha_s$6 uniformly into $\bar\alpha_t \equiv \prod_{s=1}^t \alpha_s$7. The reverse variances are fixed rather than learned, a single denoising network is shared across all timesteps, and optimization uses AdamW with CosineAnnealingWarmRestarts over $\bar\alpha_t \equiv \prod_{s=1}^t \alpha_s$8 epochs. Sampling starts from $\bar\alpha_t \equiv \prod_{s=1}^t \alpha_s$9 and iteratively applies

$$q(x_t \mid x_0)=\mathcal N\bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\bigr),$$0

to generate samples from $$q(x_t \mid x_0)=\mathcal N\bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\bigr),$$1 (Leung et al., 2024).

The empirical demonstration uses a large dataset of astronomical observations and measured labels of stars within the Galaxy. The model is applied to a variety of inference tasks, and the paper states that it can infer labels accurately with reasonable distributions. Taken together, the contribution is to graft an off-the-shelf DDPM onto a Transformer encoder so that any subset of known inputs can define a conditional density for a requested output (Leung et al., 2024).

4. Architectural patterns and objective design

For spatial data, the dominant denoiser remains the time-embedded U-Net. In T2V-DDPM, the denoiser follows the same U-Net backbone as “Improved DDPM,” with four levels of downsampling and upsampling, self-attention at $$q(x_t \mid x_0)=\mathcal N\bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\bigr),$$2 resolution, sinusoidal time embeddings, and thermal conditioning projected by a $$q(x_t \mid x_0)=\mathcal N\bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\bigr),$$3 convolution and concatenated to feature maps (Nair et al., 2022). In CBCT-based synthetic CT generation, the noise predictor is a 2D U-Net with residual blocks, group normalization, SiLU activation, multi-head self-attention at intermediate resolution, and sinusoidal timestep embeddings mapped to a $$q(x_t \mid x_0)=\mathcal N\bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\bigr),$$4-dimensional vector and then to scale-shift parameters for normalization layers (Peng et al., 2023). In DDPM-Fus, the U-Net includes several down-sampling and up-sampling stages, residual convolutional blocks, self-attention modules at intermediate scales, skip connections, and a final $$q(x_t \mid x_0)=\mathcal N\bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\bigr),$$5 convolution to predict noise with the same shape as the high-resolution HSI (Shi et al., 2023).

For non-image or non-grid targets, conditional DDPMs are adapted to the geometry of the target space. The Transformer density emulator uses an MLP denoiser because the target is a one-dimensional vector rather than a spatial field (Leung et al., 2024). The transonic wing-pressure surrogate works in PCA-modal space, using a full-rank PCA basis as a non-truncated, reversible linear reparameterization of the pressure field and an “MLP-U-Net” with residual fully connected layers, LayerNorm, ELU, self-attention blocks, and FiLM-style injection of time and label embeddings (Francés-Belda et al., 13 Apr 2026).

Although the canonical training criterion is the simplified noise-prediction loss, several papers alter it in problem-specific ways. DDPM-Fus derives the ELBO reduction to a weighted noise-prediction loss but reports that its experiments use the $$q(x_t \mid x_0)=\mathcal N\bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\bigr),$$6 norm rather than $$q(x_t \mid x_0)=\mathcal N\bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\bigr),$$7 (Shi et al., 2023). The wing-pressure surrogate introduces a signal-aware weighting

$$q(x_t \mid x_0)=\mathcal N\bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\bigr),$$8

and trains with

$$q(x_t \mid x_0)=\mathcal N\bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\bigr),$$9

to improve fidelity in regions with strong pressure gradients (Francés-Belda et al., 13 Apr 2026). The wireless reconstruction model uses a conditional loss in which the target combines a residual term proportional to $$x_t=\sqrt{\bar\alpha_t}\,x_0+\sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$0 and a Gaussian-noise term, reflecting the fact that the denoiser must jointly explain measurement degradation and diffusion noise (Letafati et al., 2023). MCDDPM augments the reconstruction objective with a bridge-network term $$x_t=\sqrt{\bar\alpha_t}\,x_0+\sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$1 and uses a single-timestep inference mode for anomaly reconstruction, reducing inference from $$x_t=\sqrt{\bar\alpha_t}\,x_0+\sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$2 to $$x_t=\sqrt{\bar\alpha_t}\,x_0+\sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$3 U-Net forwards per slice (Trivedi et al., 2024).

5. Representative application domains

Conditional DDPMs are used across density estimation, inverse problems, image translation, physical surrogates, and unsupervised detection. The reported tasks differ markedly in conditioning modality, architecture, and interpretation of stochastic outputs.

Domain	Conditioning signal	Stated outcome
Astronomical conditional density estimation (Leung et al., 2024)	Arbitrary combinations of tokenized inputs; request token hidden state	Infers labels accurately with reasonable distributions
Wireless image reconstruction (Letafati et al., 2023)	Degraded receiver-side image $$x_t=\sqrt{\bar\alpha_t}\,x_0+\sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$4	At $$x_t=\sqrt{\bar\alpha_t}\,x_0+\sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$5 with $$x_t=\sqrt{\bar\alpha_t}\,x_0+\sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$6, PSNR $$x_t=\sqrt{\bar\alpha_t}\,x_0+\sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$7
Thermal-to-visible face translation (Nair et al., 2022)	Thermal image $$x_t=\sqrt{\bar\alpha_t}\,x_0+\sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$8	VIS-TH Rank-1 $$x_t=\sqrt{\bar\alpha_t}\,x_0+\sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$9; standard DDPM $T \to \infty$0/image vs coarse-init $T \to \infty$1/image
CBCT-to-sCT generation (Peng et al., 2023)	CBCT slice $T \to \infty$2	Brain test MAE $T \to \infty$3 HU vs raw CBCT $T \to \infty$4 HU
HSI–MSI fusion (Shi et al., 2023)	LrHSI $T \to \infty$5 and HrMSI $T \to \infty$6	Reverse process generates fused HrHSI from test HrMSI and LrHSI
Ground-roll attenuation (Li et al., 2024)	Noisy seismic gather $T \to \infty$7	DDPM-2c clean prediction: MAE $T \to \infty$8, SSIM $T \to \infty$9, PSNR $\bar\alpha_T \to 0$0
Wing pressure surrogate (Francés-Belda et al., 13 Apr 2026)	Mach number, angle of attack, four control-surface deflections	Reduces mean absolute error; GRI has $\bar\alpha_T \to 0$1 with full-field MAE
Brain MRI anomaly detection (Trivedi et al., 2024)	Healthy-image context via multichannel latent code and cross-attention	Promising performance on BraTS20 and BraTS21

These examples show that the conditional DDPM formalism is not confined to unconditional image synthesis. It has been used for reliability-aware surrogate modeling on unstructured aerodynamic surfaces, restoration under channel impairments, multimodal remote-sensing fusion, and medical-image reconstruction where the generated sample is explicitly compared with an input to localize anomalies or improve quantitative image quality (Francés-Belda et al., 13 Apr 2026, Letafati et al., 2023, Shi et al., 2023, Peng et al., 2023, Trivedi et al., 2024).

6. Theory, interpretation, and recurrent misconceptions

A theoretical account of DDPM convergence under general distributional assumptions is given in “Optimal Convergence Analysis of DDPM for General Distributions.” The analysis assumes bounded second moment, a non-uniform high-probability Lipschitz condition on forward scores, and a weighted $\bar\alpha_T \to 0$2 score-estimation error

$\bar\alpha_T \to 0$3

Under these assumptions, the authors state the informal upper bound

$\bar\alpha_T \to 0$4

With perfect scores, this yields

$\bar\alpha_T \to 0$5

and the paper gives a matching lower bound $\bar\alpha_T \to 0$6 for Gaussian data under a standard schedule, showing the rate is tight up to logarithmic factors for a wide class of targets (Jiao et al., 31 Oct 2025).

This theory directly addresses a recurrent misconception that DDPM is inherently slower than DDIM in dimensional scaling. The convergence paper states that DDPM and DDIM share the same dependence on $\bar\alpha_T \to 0$7, and that the new DDPM analysis matches the $\bar\alpha_T \to 0$8 dependence in total variation while improving the smoothness dependence from a global Lipschitz constant $\bar\alpha_T \to 0$9 to the non-uniform quantity $x_T$0 (Jiao et al., 31 Oct 2025). A distinct misconception concerns conditioning: conventional conditional DDPMs often place all condition information in the reverse network, but ShiftDDPM argues that this concentrates condition modeling in a narrow timestep window and therefore introduces condition-dependent forward shifts so that trajectories are separated throughout diffusion time (Zhang et al., 2023).

Interpretation of stochastic outputs is another point of caution. In the wing-pressure surrogate, repeated conditional generations are used to define a Local Reliability Index and a Global Reliability Index. The reported correspondence between sampling-induced spread and reconstruction error supports its use as a qualitative reliability indicator rather than calibrated uncertainty quantification (Francés-Belda et al., 13 Apr 2026). This distinction is significant for scientific applications: conditional DDPM samples can reveal ambiguity structure and failure modes without automatically providing statistically calibrated posterior uncertainty.

A plausible implication of the combined empirical and theoretical record is that conditional DDPMs are best understood as a broad family of conditional distribution learners whose practical behavior depends as much on conditioning design and geometry of the target space as on the underlying diffusion formalism. The surveyed papers support three stable conclusions: the Gaussian forward/reverse machinery is shared across highly heterogeneous tasks, conditioning can be attached at the reverse process or baked into the forward trajectory itself, and diffusion outputs are most valuable when the task genuinely requires distributions rather than only point predictions (Leung et al., 2024, Zhang et al., 2023, Jiao et al., 31 Oct 2025).