Sequential Posterior Sampling with Diffusion Models

Tristan S. W. Stevens, Oisín Nolan, Jean-Luc Robert, Ruud J. G. van Sloun

Introduction

Diffusion Models (DMs) have established themselves as the gold standard for high-fidelity image generation and solving complex inverse problems. However, their iterative nature remains a significant bottleneck for real-time applications. Generating a single image typically requires dozens or hundreds of sequential steps, making them impractical for domains like medical ultrasound imaging, which demand high frame rates. This paper introduces a novel framework to accelerate diffusion-based posterior sampling by exploiting the inherent temporal correlations in video sequences.

Methodology: Leveraging Temporal Correlations

The core innovation lies in an autoregressive "warm start" initialization. Rather than initiating the reverse diffusion process from pure Gaussian noise (step $T$), the authors propose starting at a significantly lower noise level ($\tau' \ll T$). This approach effectively "jumpstarts" the model closer to the target posterior mean.

Two variants are proposed:

  1. SeqDiff: This method uses the reconstructed output of the previous frame as the starting point for the current frame. It is highly effective in scenarios with high frame rates and minimal inter-frame motion.
  2. SeqDiff+: To handle rapid motion, this version employs a Video Vision Transformer (ViViT). The ViViT analyzes a window of past frames to predict the next frame's state, providing a more accurate initialization that accounts for transition dynamics.
Comparison of sampling strategies: (a) Vanilla DPS starting from noise, (b) SeqDiff using the previous frame, and (c) SeqDiff+ using a learned transition model.

Experimental Evaluation

The methods were evaluated on the EchoNet-Dynamic dataset, which contains approximately 7,000 cardiac ultrasound (echocardiography) sequences. The researchers simulated a "scan-line undersampling" task—a common compressed sensing challenge in ultrasound aimed at reducing data acquisition time. Images were mapped to a polar grid and resized to 128x128 for processing.

Key Results and Discussion

The proposed framework demonstrates that temporal priors are remarkably valuable, often outweighing the benefits of raw iteration count. Key findings include:

  • Massive Speedup: The method achieves a 25x speedup over full diffusion trajectories, reducing the required iterations to as few as $N'=4$.
  • Superior Accuracy in Motion: During high-velocity phases of the heart cycle, SeqDiff+ improved the Peak Signal-to-Noise Ratio (PSNR) by up to 8% compared to standard initialization methods.
  • Qualitative Fidelity: Visual analysis confirms that SeqDiff+ preserves fine structural details in heart walls that are typically lost or blurred by "vanilla" fast diffusion methods.
Reconstruction results for cardiac ultrasound. SeqDiff+ with only 4 steps provides high-fidelity results comparable to or better than 100-step Vanilla DPS.

Conclusion and Future Work

By utilizing the "noise manifold" shortcut, this research moves Diffusion Models significantly closer to the sub-millisecond inference times required for live medical imaging. While the autoregressive nature of the method carries a theoretical risk of error accumulation, the continuous integration of new measurements ($y^t$) serves as an effective stabilizer. Future research will focus on lighter-weight transition models and testing the generalizability of the "warm start" approach in other high-speed domains such as automotive radar.

Summary updated

BibTeX
@inproceedings{stevens2025sequential,
  author = {Tristan S. W. Stevens and Oisín Nolan and Jean-Luc Robert and Ruud J. G. van Sloun},
  title = {{Sequential Posterior Sampling with Diffusion Models}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2025},
  doi = {10.1109/icassp49660.2025.10889752},
}
Figure