Efficient Image Generation with Diffusion Models: ProReflow Optimizes the Process

Diffusion models have revolutionized image and video generation, but the high computational cost remains a challenge. Flow Matching offers a promising solution by transforming the diffusion process of diffusion models into a straight line for a few or even just one step. A new method called ProReflow now promises to further increase the efficiency of Flow Matching.

Challenges and Solutions

The current implementation of Flow Matching reaches its limits when it comes to optimally simplifying the complex diffusion process. ProReflow addresses this problem with two innovative techniques: progressive reflow and aligned velocity prediction (aligned v-prediction).

Progressive Reflow

Progressive reflow optimizes the Flow Matching process by gradually reflowing the diffusion models in local time steps until the entire diffusion process is simplified. This approach reduces the complexity of Flow Matching and enables more efficient computation.

Aligned Velocity Prediction

The second key component of ProReflow is aligned velocity prediction. This technique emphasizes the importance of the direction of the velocity in Flow Matching over the pure magnitude of the velocity. By prioritizing directional alignment, the accuracy of the generated image is improved.

Experimental Results

Tests with established diffusion models like SDv1.5 and SDXL demonstrate the effectiveness of ProReflow. For example, ProReflow achieved an FID (Fréchet Inception Distance) score of 10.70 with SDv1.5 on the MSCOCO2014 validation dataset with only 4 sampling steps. This result is close to the FID score of the original model (10.05) with 32 DDIM steps. This highlights the potential of ProReflow to significantly reduce computational costs without significantly impacting the quality of the generated images.

Outlook

ProReflow represents a significant advancement in the field of image generation with diffusion models. By combining progressive reflow and aligned velocity prediction, ProReflow enables more efficient and faster image generation. This development could further advance the use of diffusion models in various application areas, from art to research, and open up new possibilities for creative applications. The release of the code on Github will enable further research and development in this area and facilitate the integration of ProReflow into existing systems.