Abstract
Generative models like Diffusion Models have revolutionized image synthesis but suffer from slow inference due to iterative sampling. Consistency Models (CM) and Poisson Flow Consistency Models (PFCM) address this by enabling single-step generation. However, PFCMs previously required distillation from a pre-trained "teacher" model, imposing significant computational costs. This paper introduces Poisson Flow Consistency Training (PFCT), the first framework for training PFCMs in isolation. By leveraging physics-inspired perturbation kernels and optimized noise scheduling, PFCT achieves high-fidelity image denoising in a single step without a teacher. We demonstrate its efficacy on Low-Dose CT (LDCT) scans, achieving perceptual quality comparable to existing consistency-based methods.
Introduction
The primary bottleneck in deploying generative models for real-time applications, such as medical imaging, is the high number of function evaluations (NFE) required during inference. While Consistency Models map noise directly to data, their physics-based extension—the Poisson Flow Consistency Model—has historically been tethered to the distillation paradigm. Training a PFCM required first training a complex Poisson Flow Generative Model (PFGM++), which serves as a teacher.
PFCT removes this dependency, allowing for the direct training of PFCMs. This "teacher-free" approach is particularly valuable in domains where training a high-quality teacher is computationally prohibitive or where data constraints make distillation less effective.
Methods
PFCT bridges the gap between independent training and the electrostatic physics of Poisson Flow. The core innovation lies in how the model learns the trajectory of data points in a high-dimensional space ($N+D$ dimensions).
Isolated Perturbation Kernel
In the PFGM++ framework, data points are treated as charges. To train without a teacher, the authors developed a method to sample adjacent noisy data points by varying only the radial component of the noise while keeping the angular component constant. This allows the model to internally supervise the "flow" of the trajectory, enforcing consistency across different noise levels.
Sinusoidal Discretization and Beta Noise Distribution
Standard diffusion models often use exponential noise schedules. PFCT introduces a sinusoidal discretization schedule, which provides a smoother transition between noise levels, enhancing training stability. Furthermore, a Beta noise distribution is employed to prioritize high-noise scenarios during training, ensuring the model captures the global structure of the data before refining local details.
Architecture
The model utilizes a U-Net architecture integrated with Weighted Attention Gates. This design is optimized for conditional generation, where the model uses a noisy input (e.g., a low-dose CT scan) to predict the underlying clean image.
Experimental Setup
The proposed method was evaluated on the Low-Dose CT (LDCT) Image Denoising and Projection Dataset. The task involves transforming noisy, low-radiation CT scans into high-quality images equivalent to full-dose scans.
The training set consisted of 11,646 image crops (128x128), while testing was performed on full-resolution 512x512 images. The hyperparameter $D$, representing the additional dimensions in the Poisson flow, was set to 2048.
Results
PFCT demonstrates strong performance across perceptual and pixel-level metrics while maintaining extreme efficiency.
As shown in the table above, PFCT (D=2048) achieves an LPIPS of 0.018, a significant improvement over the raw LDCT (0.075). It performs comparably to the standard Consistency Model (CM) which has an LPIPS of 0.016. In terms of pixel accuracy, PFCT reached a PSNR of 42.07 dB and an SSIM of 0.96. Crucially, PFCT achieves these results in a single step (NFE=1), whereas diffusion-based models like EDM or PFGM++ require 79 steps to reach similar or slightly better quality.
Discussion and Limitations
While PFCT successfully breaks the reliance on a teacher model, a performance gap remains compared to distillation-based PFCMs (LPIPS 0.009). The teacher model provides a refined signal that the current isolated training objective does not fully replicate. Future work will explore optimizing the hyperparameter $D$ and investigating whether more complex noise distributions can further close this gap. Despite this, PFCT represents a major step toward making physics-inspired generative models more accessible and computationally efficient.
Summary updated