Realistic Character Lora training for Chroma?
I want to train a LoRa for my AI Influencer model with chroma, has anyone done that before with good results?
If yes, what tool did you use and what settings + how many images?
ζδΉζηΈεηθ―·ζγ
Ostris AI toolkit
-anywhere from 4 to 20+ images of decent quality. (obviously better quality images produce better results)
-I haven't used any captioning.
-I've pretty well stuck with the default settings, deviating a little here and there but nothing major worth writing about.
-I train with sampling off
-save every 100 steps
-good results come early, don't be afraid of trying loras as early as step 400...
-I've noticed some loras produce horizontal banding up and down the image at higher step counts, especially in conjunction with long prompting... more below on that
--probably over training, I'm not sure :)
Almost all of my character loras have turned out really, really good. I'd suggest saving your dataset, retraining loras when you update the model you're using. the toolkit will auto-download the latest model for you when you train your lora. It will use the standard "uncalibrated" version! I have not tested whether or not training on the "detail-calibrated" versions is better or not.
-the loras work with other flux models well too :D
When prompting
-I've noticed SHORTER prompts=better results (for me at least)...
-longer prompts tend to produce images with a severe banding effect up and down the image (horizontal banding)
--I've had some success counteracting this by playing with the t5 pad node, but mostly from condensing my prompt
-I use 45 steps and 1.00 lora weight
-"amateur photograph", "dim lighting", "simple background", "realism", "[insert your preference] scene"
-I use minimal negative prompting.
--"disfigured"," fake", "animated", "unrealistic"