
Cyb3rJuice Lora!
When i trained my first Lora, it was clear i was going to use it together with the incredible JunkJuice models (especially PsychoNex).
But i made a mistake (several actually):
-
The generated dataset contained too much cyberpunk background (it could be a benefit in some case, but it overpowers too much other intended styles as you can see in last preview)
-
I used PonyXL for base, i should have used directly a JunkJuice model...
So, i went back to training and did it again:
-
Generated dataset does not have a background => focus on face as intended!
-
Used AnyGhostLora + Cyb3rlad1es as base, with a bit of Hyperrealistic Lora to "unflatten" the initial style to better fit the base model (SilkySauce)
-
A smidge of SilkySauce as a refiner (at step 0.9)
-
Flunked the Kohya_ss parameters but it still did his job ^_^;
Now, it works as intended with this model ♥.
I did some testing (as attached) with original images generated by skumerz and i feel it works best in the 0.8-0.6 range.
Check original images for prompts (and tips!)
-
UbeSauce: https://civitai.com/images/24604847
-
SilkySauce: https://civitai.com/images/24409810
-
PsychoNex: https://civitai.com/images/21027647
That's all folks!
PS: i retip any and all buzz i receive to the original contributors of the resources i used, thanks if you did tip me :D
描述:
训练词语: cyb3rJuice
名称: cyb3rJuice.safetensors
大小 (KB): 41864
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success
名称: dataset.zip
大小 (KB): 71674
类型: Training Data
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success