I retrained the LoRA with upscaled images and revised the trigger words since "splash" gave it a strong bias towards water. Now the triggerword is "gs2", but the LoRA also works well without the trigger.
The LoRA is now more flexible and easily steered towards the desired result. Faces are still wonky, but you can fix them with inpainting (or automate the process in Automatic1111 with the extension "after detailer")
You can still use words like "swirling" "dynamic" or similar to give it more dynamism.
I retrained the LoRA with upscaled images and revised the trigger words since "splash" gave it a strong bias towards water.
Now the triggerword is "gs2", but the LoRA also works well without the trigger.
The LoRA is now more flexible and easily steered towards the desired result.
Faces are still wonky, but you can fix them with inpainting (or automate the process in Automatic1111 with the extension "after detailer")
You can still use words like "swirling" "dynamic" or similar to give it more dynamism.