MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) researchers have introduced a groundbreaking framework called Distribution Matching Distillation (DMD). This innovative approach simplifies the traditional multi-step process of diffusion models into a single step, addressing previous limitations.
Traditionally, image generation has been a complex and time-intensive process, involving multiple iterations to perfect the final result. However, the newly developed DMD framework simplifies this process, significantly reducing computational time while maintaining or even surpassing the quality of the generated images. Led by Tianwei Yin, an MIT PhD student, the research team has achieved a remarkable feat: accelerating current diffusion models like Stable Diffusion and DALL-E-3 by a staggering 30 times. Just compare the image generation results of Stable Diffusion (image on the left) after 50 steps and DMD (image on the right) after just one step. The quality and detail are amazing!
The key to DMD’s success lies in its innovative approach, which combines principles from generative adversarial networks (GANs) with those of diffusion models. By distilling the knowledge of more complex models into a simpler, faster one, DMD achieves visual content generation in a single step.
But how does DMD accomplish this feat? It combines two components:
    1. Regression Loss: This anchors the mapping, ensuring a coarse organization of the image space during training.
    2. Distribution Matching Loss: It aligns the probability of generating an image with the student model to its real-world occurrence frequency.
Through the use of two diffusion models as guides, DMD minimizes the distribution divergence between generated and real images, resulting in faster generation without compromising quality.
In their research, Yin and his colleagues demonstrated the effectiveness of DMD across various benchmarks. Notably, DMD showed consistent performance on popular benchmarks such as ImageNet, achieving a Fréchet inception distance (FID) score of just 0.3 – a testament to the quality and diversity of the generated images. Furthermore, DMD excelled in industrial-scale text-to-image generation, showcasing its versatility and real-world applicability.
Despite its remarkable achievements, DMD’s performance is intrinsically linked to the capabilities of the teacher model used during the distillation process. While the current version utilizes Stable Diffusion v1.5 as the teacher model, future iterations could benefit from more advanced models, unlocking new possibilities for high-quality real-time visual editing.