PeRFlow offers a novel approach to accelerate pre-trained diffusion models, enabling high-quality image generation in as few as 4 steps. Targeting researchers and developers working with Stable Diffusion, it provides a plug-and-play module that significantly reduces sampling time while maintaining or improving generation quality compared to existing methods like LCM.

How It Works

PeRFlow learns a piecewise linear probability flow by dividing the diffusion process into segments and applying a "reflow" operation to each. This "divide-and-conquer" strategy avoids the computationally expensive simulation of the entire ODE trajectory required by prior methods, allowing for efficient online training. The core innovation is the learned $\Delta W$ (difference in model weights), which can be fused with existing Stable Diffusion models as a universal accelerator.

Quick Start & Requirements

• Install dependencies with env/install.sh.
• Requires PyTorch and Hugging Face diffusers (v0.19.3 recommended for specific applications).
• GPU with CUDA is necessary for inference and training.
• Pre-trained delta-weights for SD-v1.5 and SDXL are available on Hugging Face.
• Official project page: https://piecewise-rectified-flow.github.io
• Tech Report: https://arxiv.org/abs/2405.07510

Highlighted Details

• Achieves high-fidelity image generation in 4 steps, outperforming LCM in diversity and quality.
• Fine-tuning converges significantly faster than InstaFlow (4k vs. 25k iterations).
• $\Delta W$ module is compatible with various SD pipelines (ControlNet, IP-Adapter, multi-view generation).
• Supports classifier-free guidance and negative prompts.

Maintenance & Community

• Active development with recent releases of SDXL and training scripts.
• Contributions are welcomed.
• Related projects: Rectified Flow, InstaFlow.

Licensing & Compatibility

• The repository itself does not explicitly state a license.
• Models released on Hugging Face are typically under permissive licenses (e.g., MIT for diffusers), but specific model licenses should be checked.
• Designed for compatibility with existing Stable Diffusion workflows and fine-tuned models.

Limitations & Caveats

The project is presented as a research artifact with NeurIPS 2024 publication. While delta-weights are provided, users may need to fuse them with their specific base models, requiring careful implementation. The "universal plug-and-play" claim is demonstrated across several applications, but extensive testing across all possible SD variants might be needed.