2000 steps trained on a100 28 images 1024x1024 1hr30m manual captioning natural language. Tested on flux dev fp8/fp16, lora workflow in image metadata.

please play around it is very experimental. i notice that flux lora seem to do better with higher weights try experimenting with higher strength. hopefully it will be improved when flux training is better

update: i just tested fp16, please use that instead of fp8 it is much better. the green field image are generated using fp16

trigger word: megumin

clothing: wearing a witch hat and a cape, dressed in a red long sleeved strapless dress with a brown belt wearing black fingerless gloves and a collar, she wears bandages on her left left and black thighhighs on her right leg,

Example
Positive prompt ✅:

still frame japanese anime screencap from 2019, megumin wearing a witch hat and a cape, dressed in a red long sleeved strapless dress with a brown belt wearing black fingerless gloves and a collar, she wears bandages on her left left and black thighhighs on her right leg, she is looking at the camera mouth opened, subtitles bottom of the screen that says "explosion!", outdoors green field day,

Negative prompt ❌:

useful link

https://github.com/bghira/SimpleTuner/blob/main/documentation/quickstart/FLUX.md

https://civarchive.com/articles/6678/training-a-flux-lora-using-simpletuner-on-runpod