(imported from civitai, original page) USE AT YOUR OWN RISK ON THE ON-SITE GENERATOR HERE. I'm not sure what's different here but it doesn't seem to work the same on here. This was tested locally on Forge UI, 35 steps, 5 CFG.

While working on another series of light trails loras (upcoming), and my Conceptual Photography series, and testing hundreds of epochs with LONG, COMPLEX PROMPTS (chatgpt is very helpful with this!), I'd come across epochs that were interesting and thought I'd share them here and call them WEIRD SCIENCE. These have prompt adherence but will disobey in creative ways and/or have concept bleed. They also have maximum SEED VARIETY. Run at higher strength at your own risk, or use the recommended. These will have sameface due to the small(ish) dataset size, just fix with your favorite character loras or default flux with inpainting/adetailer. These are style loras and trained on a wide variety of images and are not intended to reproduce any likeness.

TRAINED ON REAL PHOTOGRAPHY. Too many loras use ai to train ai imo, and it shows.

I have presented example images as UNEDITED AND NON-CHERRYPICKED GRIDS to show seed variety/creativity. Each pic in the grid uses the same exact prompt.

Personally I love these types of loras--often flux is way too literal and boring with its interpretation of a prompt, and seeds are WAY too similar. When I run a bunch of seeds on a prompt I want variety!

These are also trained at a high dim in an attempt to override flux's biases, so the file size is large. Someone more skilled and experienced could probably figure out a way to do this without having such large files, but thats my experience level at the moment. Tested on my 3090 locally with Forge UI with Euler/Simple and most around 5(distilled)CFG, on FLUX DEV FP8.

Settings for each are in the metadata of the images. See image gallery for examples of what these can do.