fm-boosting
FMBoost is an open-source research project presented at ECCV 2024 that enhances Latent Diffusion Models by integrating Flow Matching for high-resolution image synthesis. Developed by the CompVis Group at LMU Munich, the software addresses the trade-off between image fidelity and generation speed. It employs a two-stage pipeline: first, a small diffusion model generates a low-resolution latent representation leveraging the diversity of stochastic diffusion; second, a Coupling Flow Matching model regresses a vector field to directly map this low-resolution latent to a high-resolution latent within a continuous probability path. This approach allows users to cascade models, transforming low-resolution outputs (e.g., 128x128 pixels) into high-fidelity images at 1024x1024 or 2048x2048 pixels with minimal computational cost. The project demonstrates that this method achieves exceptionally fast synthesis times, averaging 0.347 seconds for 1024x1024 images, outperforming standard high-resolution models like LCM-SDXL