Imagine being able to train SDXL with native FLUX VAE – it’s a possibility that has excited many in the community. And now, thanks to some experimentation, we can confirm that it’s indeed possible.
The journey wasn’t easy, though. It took 10 hours of training on a 4060ti to get even close to the baseline output. But the results are promising, and the potential for further improvement is huge.
With this breakthrough, we can explore new possibilities like fine-tuning CLIPs for anime and converting the model to Rectified Flow from the get-go. The community can now take this forward and experiment with larger data for VAE.
To support this effort, a goal of $5000 has been set up on ko-fi, which will account for trial runs and experimentation. The community’s involvement can take this project to new heights.
This achievement is a testament to the power of community-driven innovation. With collaboration and perseverance, we can push the boundaries of what’s possible in AI and machine learning.
—
*Further reading: [Stable Diffusion subreddit](https://www.reddit.com/r/StableDiffusion/)*