Continuous Adversarial MeanFlow Transfer
This paper introduces MeanFlow-Transfer and Continuous Adversarial MeanFlow (CAMF), a unified framework that adapts heterogeneous pretrained flow models to new domains with limited data while simultaneously accelerating generation to few-step sampling, achieving state-of-the-art quality with up to 125× fewer neural function evaluations.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of artificial intelligence, computers have learned to create images that look startlingly real, from sunsets over mountains to portraits of people who never existed. These systems, often called generative models, work by learning the statistical patterns of millions of photographs. They start with random noise and gradually refine it, step by step, until a clear picture emerges. For years, this process has been like a slow, careful sculpting session, requiring hundreds of tiny adjustments to get the details right. While the results are beautiful, the time and computing power needed to make them have been a major bottleneck. Researchers have been trying to speed this up, hoping to create high-quality images in just a few steps, but they have faced a stubborn problem: the tools that make these images fast are often locked into specific formats. A model trained to predict one type of pattern cannot easily be taught to predict another, and a model designed for one subject, like cats, struggles to adapt to a new one, like cars, without losing its speed or quality.
A team of researchers at ÉTS Montreal has developed a new approach that solves both of these problems at once. They created a method that can take any of these existing, slow image generators—regardless of how they were originally built—and quickly adapt them to create new images of different subjects, all while making the process hundreds of times faster. Their system, which they call MeanFlow-Transfer, acts as a universal translator. It takes the output of a slow, pre-trained model and converts it into a shared language that any fast generator can understand. This allows the system to start with a model trained on a massive dataset of general images and instantly teach it to create high-quality images of specific things, like birds or cars, using only a small amount of new data. The result is a generator that can produce sharp, detailed images in just a handful of steps, a task that previously required hundreds.
To ensure these fast images didn't lose their fine details in the rush, the researchers added a second stage to their process. They introduced a refinement step that uses a learning technique called adversarial training. In this stage, a second neural network acts as a critic, comparing the images produced by the fast generator against real photos. Instead of just checking if the overall shape is right, this critic looks at the subtle changes between the starting noise and the final image, checking if the journey the image took to get there makes sense. This helps the system recover tiny details that are often blurred out when trying to speed things up. By combining the translation method with this critical eye, the team found that their system could match or even beat the quality of the original, slow models, but with up to 125 times fewer computational steps.
The researchers tested this method by taking four different types of pre-trained models, each built with a different internal logic, and adapting them to five distinct new domains, including images of birds, cars, and artwork. In every case, the system successfully transferred the knowledge from the original model to the new subject. When they measured the quality of the images using standard metrics, the new system produced results that were as good as, or better than, the original models that had been fine-tuned for the new task. Perhaps most impressively, it achieved this quality while using a fraction of the computing power. For instance, a model that originally needed 250 steps to create a good image of a bird could now do it in just four steps without losing clarity.
The team also discovered that trying to force the old models to follow a new schedule of steps, a method some other researchers had tried, actually made the training unstable and the results worse. Instead, their success came from simply mapping the different ways the models "thought" about the image into a single, common way of describing movement. They proved that this approach works because it respects the underlying physics of how the images are formed, rather than trying to force them into a rigid new pattern. Furthermore, they showed that their refinement technique, which checks the entire journey of the image creation, is a natural extension of older methods that only checked the final moment. As the time steps between checks get smaller, their method smoothly turns into the older, simpler version, proving that their new approach is a more powerful and flexible version of what came before.
This work matters because it removes the barriers that have kept fast image generation locked to specific types of models or subjects. Before this, if you wanted a fast generator for a new type of image, you often had to start from scratch or accept a drop in quality. Now, a single framework can take a wide variety of existing, powerful models and turn them into fast, specialized tools for any new task. The researchers demonstrated that by using their method, it is possible to create a high-quality generator for a new domain using a tiny amount of data, making advanced image synthesis much more accessible and efficient. The findings suggest that the future of generative AI may not be about building bigger, slower models, but about finding smarter ways to adapt and accelerate the ones we already have.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.