This LoRA generates human figures in states of motion such as swimming, dancing, etc. Please see the image gallery for prompt templates and ideas. It is also possible to generate landscape and other types of images, but that is not the focus of the LoRA.
Due to the training material, by default figures will be rendered with little or no clothing, so put something on them in your prompt if you want SFW images 😅.
Known problem: the LoRA has the tendency to produce extra limbs, which is probably due to images in the dataset where some limbs are closely intertwined. I suppose I can train again with those images taken out, but the resulting LoRA would probably be less "creative". For now, you can try a different seed, lowering the weight (if you want to keep the composition), or try an earlier epoch such as epoch 8.
Description
This is epoch 12. Other epochs can be found at tensor.art/models/852787139383961903
FLUX.1 - dev-fp8
Trigger: painting
Repeat: 20 Epoch: 4 (Trained on 36 512x512 images for 2880 total steps)
Then trained for another 8 epochs with 42 images for 9600 total steps
(* 20 8 42)=6720 (+ 6720 2880)=9600
Unet LR: 0.0005 Scheduler: cosine Optimizer: AdamW
Network Dim: 8 Alpha: 4
Epoch Loss
1 0.360
2 0.362<
3 0.349
4 0.350<
painting. A woman and a polar bear are swimming underwater.
5 0.358
6 0.346
7 0.339
8 0.325
painting. A woman and a tiger are ballet dancing.
9 0.336
10 0.326
11 0.309
12 0.300<-lowest
painting. A man and a goat are ballet dancing.