firefly Flowing Fireflies "Collapse: Starchaser"

The training set mainly consists of Danbooru, webpage screenshots, and derivative works (ai paintings), including a few game screenshots.

Therefore, it is not able to restore the images in the game very well, but it has high versatility and playability. It is recommended to use it with cetus-MIX WhaleFall2 model.

Weight recommendation 0.6~0.8, hope everyone can contribute more images and provide valuable feedback. I will also actively make corrections in the future.

Note: The resolution of the training set is high, recommended for high-resolution play. Parameters can refer to my example images.

V2 version updates:

Added some in-game screenshots and other derivative images.

Enhanced training on facial features.

Performs well on data outside the training set (generalization). Will provide some performances not in the dataset for comparison in the comments section.

As I personally believe that achieving generalization is the ultimate goal of training, my training leans towards this.

After all, direct image generation from the original image can also achieve restorability (not).

Most of the time, please try to avoid using negative embedding without special needs (it will fix your style in a specific direction). If interested, you can try it yourself, maybe you will get better results, resolution recommended above 2k.