This was trained on Sabu's recent art. I tried to delineate between his monochromatic, his greyscale, and his more fully colored art, which is why I put three trigger phrases.
Training params for version 3: { "lr": 0.00015, "engine": "ai-toolkit", "epochs": 30, "ecosystem": "zimageturbo", "keepTokens": 0, "networkDim": 32, "numRepeats": 8, "resolution": 1024, "lrScheduler": "constant", "minSnrGamma": null, "noiseOffset": null, "networkAlpha": 32, "optimizerType": "adamw8bit", "shuffleTokens": false, "textEncoderLr": null, "flipAugmentation": false, "trainTextEncoder": false }
I had real trouble getting a lora of the same quality as the version 2. I'm trying to understand what makes a good style lora, and the default settings really made terrible loras. So I kept fiddling with the learning rate, the number of repeats, but I still feel like I'm shooting in the dark :/
Description
This is the 25th Epoch, the preview images led me to believe that it had become seriously overtrained, or regressed, but I noticed the pictures looked the same as Epochs 1 - 8 so I saved it anyway to make sure, that's what I'm doing now.
FAQ
Comments (2)
The newer version is significantly better on faces, and still pretty good at tits. Still getting some watermarks/signatures, and ZTurbo itself still limits the effectiveness of NSFW stuff, but overall: a great improvement!
I'm happy with the output, despite a few places it still struggles. Thank you for the support!