Fast-DDPM: Accelerated Diffusion Models

• Fast-DDPM is an architecture for denoising diffusion models that uses a matched set of 10 discrete timesteps to significantly reduce computational overhead in high-dimensional medical imaging.
• The method aligns both training and inference processes, reducing training time to 20% and sampling time to 1% compared to standard 1,000-step DDPMs.
• Empirical results in volumetric super-resolution, denoising, and translation show superior PSNR/SSIM performance and drastic speedup over traditional GAN and CNN based architectures.

Fast-DDPM is an architecture and sampling protocol for Denoising Diffusion Probabilistic Models (DDPMs) specifically engineered to overcome the significant computational bottlenecks inherent in standard diffusion-based generative methods, especially for high-dimensional, medical image-to-image generation. Fast-DDPM departs from the canonical 1,000-step training and sampling paradigm by aligning both training and inference to a drastically reduced, matched, highly optimized 10-step schedule. This approach achieves state-of-the-art generation fidelity (as measured by metrics such as PSNR/SSIM) while cutting training time to approximately 20% and sampling time to 1% of the standard DDPM baseline, as validated across volumetric super-resolution, denoising, and translation tasks in 3D/4D medical imaging (Jiang et al., 2024).

1. Motivation and Core Principles

The principal challenge addressed by Fast-DDPM is the prohibitive cost of applying standard DDPMs to medical volumes, which are often three- or four-dimensional and require days to weeks for training and minutes to hours for sampling a single image. These constraints are largely a result of the 1,000-step Markov diffusion/sampling chains, which are not coordinated between training and inference: standard approaches train across all 1,000 noise scales but often use only a small subset during sampling, leading to severe computational waste and suboptimal step utilization (Jiang et al., 2024).

Fast-DDPM solves this by:

• Restricting both training and sampling to a common set of 10 discrete timesteps.
• Designing two noise schedulers (uniform and non-uniform over noise level) for versatility.
• Directly aligning the denoiser’s capacity with the actual inference trajectory, eliminating the training/sampling mismatch prevalent in DDIM/PLMS/PNDM/DDPM-Solver approaches that rely on post-training schedule subsampling.

2. Architecture and Conditioning Schemes

The denoiser $\epsilon_\theta$ in Fast-DDPM retains the established U-Net backbone, adopting modality-appropriate dimensionality:

• In 2D, each block uses Conv2D + GroupNorm + ReLU.
• For 3D or 4D (e.g. temporal) data, these become Conv3D/Conv4D, with GroupNorm and ReLU or Swish activations.
• Down-/up-sampling blocks concatenate conditional images or feature maps (such as additional slices or multi-contrast channels) in the channel dimension and can incorporate FiLM-style scaling for richer conditioning.
• For volumetric input, skip connections propagate fine-structure through encoder-decoder levels, with each stage concatenating conditional features channel-wise (Jiang et al., 2024).

Normalized input volumes (e.g., $256\times256\times N_{\mathrm{slices}}$ into $[-1,1]$) and flexible condition fusion strategies ensure architectural adaptability across MRI, CT, and other domain-specific flows.

3. Diffusion Process, Noise Schedule, and Loss

Fast-DDPM defines its noise schedule by subsampling a smooth “master” $\alpha^2(t)$ curve, itself defined as

$$\alpha^2(t) = \prod_{j=1}^{1000} \left(1 - \beta(j/1000)\right), \text{ with } \beta(u) = 0.0001 + (0.02-0.0001) u$$

Evaluated at ten points $t_i$ (either $t_i = i/10$ for uniform or denser at high noise for non-uniform), the schedule yields

$$\alpha_i = \sqrt{\alpha^2(t_i)}, \quad \sigma_i = \sqrt{1-\alpha^2(t_i)}$$

The one-step forward kernel and marginal transitions are

$$q\left(x_i|x_{i-1}\right) = \mathcal N\left(\sqrt{1-\beta_i} x_{i-1}, \beta_i \mathbf{I}\right), \;\; \beta_i = \alpha^2_{i-1} - \alpha^2_i$$

$$q(x_i \mid x_0) = \mathcal N\left(\sqrt{\bar\alpha_i} x_0, (1-\bar\alpha_i) \mathbf{I}\right), \;\; \bar\alpha_i = \prod_{j=1}^i \alpha^2_j$$

as in canonical DDPMs, but operating solely on these 10 selected scales.

Training minimizes the standard MSE in the predicted Gaussian noise: $256\times256\times N_{\mathrm{slices}}$0 This loss focuses network capacity exclusively on the actually utilized noise levels, maximizing training impact per step.

4. Sampling Procedure and Step Alignment

Sampling runs the following 10-step iterative scheme: $\alpha^2(t)$1 Because the network is trained and sampled on the identical collection $256\times256\times N_{\mathrm{slices}}$1, no noise levels are “wasted” on untrained regimes, and there is no mismatch—contrary to naive DDPM or mismatched DDIM/PLMS samplers. This principle underpins the extreme acceleration: sample generation converts from a $256\times256\times N_{\mathrm{slices}}$2-step, tens-of-seconds (to hours for 3D) process to a 10-step, second-scale operation with fidelity parity or improvement (Jiang et al., 2024).

5. Specialized Adaptations for High-Dimensional Medical Data

In 3D/4D instantiations:

• The U-Net blocks are reconfigured to Conv3D (or Conv4D) layers with $256\times256\times N_{\mathrm{slices}}$3 (resp. $256\times256\times N_{\mathrm{slices}}$4) kernels, stride-2 downsampling, and learned transpose-conv upsampling.
• Skip connections propagate all spatial resolutions.
• Conditional data—a stack of adjacent slices, modalities, or contrasts—is encoded as multi-channel input, concatenated at each architecture level.
• Input volumes are normalized slice-wise, and the model is trained to reconstruct highly structured anatomical details.
• For tasks such as volumetric super-resolution, image denoising, and translation, Fast-DDPM rigorously outperforms baseline convolutional and GAN-based architectures on both perceptual (SSIM) and distortion (PSNR) metrics (Jiang et al., 2024).

6. Empirical Efficiency and Benchmark Results

Reported benchmarks show, on multi-volume super-resolution:

• DDPM training $256\times256\times N_{\mathrm{slices}}$5 h $256\times256\times N_{\mathrm{slices}}$6 Fast-DDPM $256\times256\times N_{\mathrm{slices}}$7 h ($256\times256\times N_{\mathrm{slices}}$8).
• DDPM sampling $256\times256\times N_{\mathrm{slices}}$9 min/volume $[-1,1]$0 Fast-DDPM $[-1,1]$1 s/volume ($[-1,1]$2).
• CT denoising and MRI translation exhibit similar speedup ($[-1,1]$3 training, $[-1,1]$4 sampling).
• Across all tasks, Fast-DDPM achieves superior PSNR/SSIM and outperforms both classic and SOTA GAN/CNN methods (Jiang et al., 2024).

Task	Training Time	Sampling Time	PSNR/SSIM	Relative Speedup
Standard DDPM	136 h	3.7 min	SOTA, baseline	1×
Fast-DDPM	26 h	2.3 s	Higher	$[-1,1]$50.2× train/$[-1,1]$60.01× sample

7. Broader Impacts, Limitations, and Confirmatory Studies

The Fast-DDPM approach has catalyzed further work in fast, domain-adapted diffusion architectures:

• Lung-DDPM+ replaces the standard 1,000-step DDPM with domain-conditioned, high-order ODE solvers, achieving $[-1,1]$7 fewer FLOPs, $[-1,1]$8 lower memory, and $[-1,1]$9 faster sampling while preserving segmentation and visual metrics (Jiang et al., 12 Aug 2025).
• Minutes to Seconds applies a similar paradigm for 2D inpainting, combining reduced-parameter networks, skip-step DDIM sampling, and a two-stage (coarse-resolve/fine-refine) process for $\alpha^2(t)$0 acceleration without significant drop in LPIPS or SSIM (Zhang et al., 2024).
• These architectures validate the transferability of the Fast-DDPM principle—full-step schedule matching, ultra-low NFE, and network-targeted efficiency gains—to diverse generation and restoration domains.

A potential limitation is the focus on one specific set of 10 schedule points; while high performance is retained across tasks, extremely nonstationary noise characteristics or highly atypical conditioning may require reoptimization of the schedule. Nonetheless, Fast-DDPM sets a new standard for computationally efficient, high-fidelity generative modeling in high-dimensional spaces (Jiang et al., 2024).