
My first model ever was a LoRa called Fruitpunk which I subsequently took down. Some people really liked it but I honestly felt it was miserably overtrained and what they REALLY liked were the original images I made that trained it.
I've rectified that issue in this version. The images used here for comparison are from juggernaut, with the exact same prompt and seed used to generate the comparison images, so you can see how it is affected. You can get MUCH stronger effects and style than this by using words like "fruitman" or even just referencing fruits or fruit tattoos.
There WILL still be some bias and overtraining, this was made with just 15 images, my main goal was to ensure the tattoos in the images it received were coherent, and that the prompts used would push it in the correct direction.
I expect this to improve greatly on the next version.
描述:
THIS VERSION IS HIGHLY BIASED FOR CHARACTER PORTRAITS.
This version has better prompt adherence in most scenarios and is in general more flexible than its predecessor while maintaining the preferences. You may need to use some prompting, like adding the word "fruit" to previous generations to get the same effect, but in general it basically works the same way, and is similar stylistically.
Alternately, you can just increase the cfg past 1.0 and you'll get very similar images to your previous gens, with a bit better adherence, just be sure to turn down your overall cfg slightly when doing so to tune it such that you don't get that nasty artifacted look.
训练词语:
名称: Fruitpunk_4-000201.safetensors
大小 (KB): 223102
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success