Faster Image Generation through Innovative Diffusion Models

The world of artificial intelligence (AI) is developing rapidly, and particularly in the field of image generation, there are constantly new advancements. Diffusion models have proven to be powerful tools for creating high-quality images, however, their application has been limited by a time-consuming, iterative process. This multi-step sampling process slows down generation and makes it difficult to use in interactive applications. New research approaches now promise a remedy by significantly increasing the speed of image generation.

From Multi-Step to One-Step: Distillation of Diffusion Models

A promising approach to accelerating image generation is the distillation of multi-step diffusion models into so-called one-step generators. Here, a complex, multi-step "teacher" model is used to train a simpler, one-step "student" model. The goal is to match the distribution of the images generated by the student model as closely as possible to the distribution of the teacher model. Previous methods use the reverse Kullback-Leibler (KL) divergence for this, a method for measuring the difference between two probability distributions. However, this method has the disadvantage that it tends to focus on the modes of the distribution, which can lead to a limited variety of generated images.

f-Distill: A New Approach to Divergence Minimization

A new research approach called f-Distill generalizes the previous approach by offering a more flexible framework for divergence minimization. This approach is based on the use of f-divergences, a class of functions that encompasses various divergence measures such as the reverse KL divergence, the forward KL divergence, and the Jensen-Shannon divergence. By selecting different f-divergences, an optimal compromise between mode coverage and training variance can be found.

The f-Distill approach calculates the gradient of the f-divergence between the teacher and student distributions. This gradient is expressed as the product of the score differences of the two distributions and a weighting function. The weighting function is determined by the density ratio of the two distributions and naturally emphasizes the regions with higher density in the teacher distribution, especially when using divergences that are less mode-seeking. Interestingly, it turns out that the common method of variational score distillation with reverse KL divergence is a special case of the f-Distill framework.

Empirical Results and Outlook

Empirical studies show that alternative f-divergences, such as the forward KL divergence and the Jensen-Shannon divergence, outperform previous methods of variational score distillation in various image generation tasks. In particular, with the Jensen-Shannon divergence, f-Distill currently achieves outstanding results in one-step generation on ImageNet64 and in zero-shot text-to-image generation on MS-COCO.

These advances open up new possibilities for the use of AI-based image generators in various application areas. The faster generation through one-step models enables interactive applications and facilitates integration into existing workflows. For companies like Mindverse, which offer AI-powered content solutions, these developments are of particular interest. The integration of faster and more efficient image generators into platforms like Mindverse can further optimize content creation and open up new possibilities for creative applications.

Bibliography: https://huggingface.co/papers/2502.15681 https://chatpaper.com/chatpaper/zh-CN/paper/110223 https://arxiv.org/abs/2311.18828 https://openaccess.thecvf.com/content/CVPR2024/papers/Yin_One-step_Diffusion_with_Distribution_Matching_Distillation_CVPR_2024_paper.pdf https://tianweiy.github.io/dmd/ https://openreview.net/forum?id=ogk236hsJM&referrer=%5Bthe%20profile%20of%20Guo-Jun%20Qi%5D(%2Fprofile%3Fid%3D~Guo-Jun_Qi1) https://neurips.cc/virtual/2024/poster/93608 https://www.researchgate.net/publication/384215207_One-Step_Diffusion_with_Distribution_Matching_Distillation https://arxiv.org/pdf/2502.08364 https://research.adobe.com/publication/one-step-diffusion-with-distribution-matching-distillation/