Ready for an image upgrade? Best of Flux takes what you create and cranks it up a notch—making everything brighter, more colorful, and packed with detail. Don't believe me? Check out the before and after posts below.

After the success of Best of Pony, I’m back with Best of Flux—and trust me, it’s even better. I handpicked some of the most eye-catching, top-rated images on Civitai and turned them into a custom dataset. The dataset is based on a process I fine-tuned over months of research and testing and this LoRA represents much of that work. Built on Flux Dev 1, this LoRA keeps the magic of those standout generations but taps into Flux’s full power to make your images truly pop. Whether you want more vibrant colors or next-level complexity, Best of Flux brings out the best that Flux has to offer.

My notes are in the linked article if you are interested in the dataset-building process.

Version 3.0

Version 3.0 is much more stable than earlier versions. While the visuals are not as extreme as Version 2.0 and 2.2, this version produces intricate and enhanced details without the unpredictability of the earlier versions. Use at between 0.4 and 1 strength, depending on the level of enhancement you'd like to add. The trigger word is "best_of."

Version 2.2

Trained for an additional 200 steps, Version 2.2 can make even more striking transformations—but subtle enhancements will require a lower LoRA strength. At full strength, this LoRA can easily reinterpret a prompt, sometimes creating very different (but, in most cases, very striking) images. It's fun to play with and you'll get some very nice images that otherwise wouldn't be possible. The trigger word is still "best_of."

Version 2.0

A month ago, I made an earlier version of this LoRA that I was never quite happy with, but after experimenting with Flux for a while, I wanted to give it another go. This version does much more of what I wanted from the original. Use at full strength and with the trigger word "best_of" and watch the transformations take place. Based on the dataset, this version seems to do better with a digital art style and complex scenes that aren't close-ups of people.