G-U-N/Phased-Consistency-Model

Phased Consistency Models (PCM) partition the diffusion ODE trajectory into sub-trajectories, enabling efficient multi-step image generation with 2-4 inference steps while maintaining flexibility for classifier-free guidance and negative prompt conditioning. The approach distills pre-trained diffusion models (SD 1.5, SDXL, SD3) into lightweight LoRA adapters, addressing limitations of prior work (LCM) including stochastic sampling error accumulation and insensitivity to guidance parameters. Training requires O(N) objectives rather than CTM's O(N²), making it more practical for adapting existing text-to-image models on HuggingFace.

514 stars. No commits in the last 6 months.