This is my very first attempt at training.
If you're curious, I used Kohya SS to train. I've included in the files section, to the right over there, the settings file I used for Kohya as well as every training image I used. Literally just load that setting file, point to those images, click start. You'll have this exact LoRA. Help me improve my skills.
Description
1.1
added more training images (58 total)
improved training prompting (i.e. "smoke, grey smoke that looks like a man kicking a soccer ball, white background")
ability to define smoke and/or background color
ability to define smoke source
FAQ
Comments (7)
Oh, good Rola,I love this style, it's great to do it for the first time,
A tip I got when creating my first LoRAs was to not use actual words as triggerword, since they are already present in the neural network. The training and outcome might be effected by this. However, the effect is nice. I left a review, too. ;)
I figured this would be used to "grow" its understanding of the trigger words. Bad assumption?
@magikman74 In a way it does. I would say it replaces the understandig of smoke, with a new concept. I made the experience, that training with an abstract triggerword works a bit better. You could try to change it to sm0k3 or something and train again, to see, if the effect is understood in a similar way.
@DonMischo great idea, thank you.
This is quite amazing!! Can you share a bit of your workflow on what kind of images these were trained on and how you captioned them? These types of none existing styles are the most impressive!
Check out my latest version. I give it all, including the training images.