About :
Synthetic ArtStyles are cocktails of specific dosages of pre-trained ArtStyles all mixed with a crude LoRA made from generated images to stabilize into a new style
to put it simply this an experiment involving training then merging
Samples settings :
weight of 0.8 to 1
Checkpoints: Hassaku, AnyloraV5
Steps : 22-25
Sampler : DPM++ 2M
ClipSkip : 2
Description
FAQ
Comments (3)
Can you share a bit more info about your workflow to create this synthetic artstyles?
Train a lora for each of a few artists, then merge them, then...?
it starts with training a quick and crude LoRAs with a small dataset (dataset could be hand picked or CNET generated with ip adapters) these will serve as main ingredients.
these crude LoRAs when used alone they're overfit, strong and unstable, so we mellow them down with other stable style LoRAs at also while experimenting with different weights, until we settle on an interesting style .
@navimixu So, merging multiple overfit lora with a more stable one is a viable merging strategy. I'll have to try that.
The results speak for themselves, this is an amazing style lora.
Thanks for sharing!
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.