Hi all,

I was curious what would happen if I mixed some of my flux lora files in different ratios and just asked it to create me an image of 'an attractive woman covered in gunge'. These were some of the results and I thought they made for quite an interesting mixed set of images.

Hope you enjoy.

Some great outputs, I'm seeing similar results when combining Loras either with different concepts or substances. Flux's learning with Loras is a dark art. The T5 encoder is hugely flexible and you can either let it go it's own way with minimal captioning and retain a lot of the underlying model or caption more details at a risk of collapsing the main model. Either way, you don't always know how the Lora will interpret and learn the concept.
Stacking a pie Lora with a gunge or custard lora give some great results as generally the pie images were less messy but the combination didn't add gunge to the image, just the more messy concept resulting in more cream and filling being smeared.

Thanks, I have two main lora files I generally use. One is very flexible and seems to better understand the concept and the other tends to give better mess but imposes more of the style of it's training data on the output. I find the Lora files do pick up quite a lot of style from their training images that can be quite overpowering to different concepts or things you want the scene to be. I generally run the lora through quite a number of tests along similar lines to these tests below:

- different outfits with specified colour (lots of outfits)
- different art styles (realistic, 2.5D, anime)
- different environments and settings (gameshow, on a bed, in a garden, playing tennis, as a waitress, in a bath of gunge)
- tests with no models in them at all to have it show 'gunge' in different places and on different objects

The best ones are flexible enough to make these without just throwing me out images that looked like the training data and understanding what 'gunge' actually is. The best ones I have made have needed a high number of steps and a higher network dimension in the training. They result in much bigger lora files (175mb or so is my average now) and take a long time to train but seem very flexible to apply to lots of situations. They do still pick up some of my style though as they tend towards asian women and more model type poses but they can be prompted out of that. I am keen to fine tune a whole model to see the advantages over a lora.

Wonderful images, man. A lot of these generations look like the old messy models asia ladies (naomi, arisa, and koko)

haha, yeah definitely. With a really limited prompt like this it is going to go closer to it's training images. My original lora files were all trained off my old messy models asia images. Newer models I have trained off my own AI images .... but it all has it's roots in the original images.