This lora is intended to recreate stylized Widowmaker's LIKENESS thus it is not a widowmaker (character lora), it's stylized Widowmaker's LIKENESS!!!!
The lora has an issues: first is that the head is turning away and the second one is that with weight above 0.70 images that generated are getting very close up and might be cropped!
CIVITAI'S GENERATOR IS NOT TESTED WITH THIS LORA AND THE LORA IS NOT INTENDED TO BE USED WITH CIVITAI'S GENERATOR!!! USE AT YOUR OWN BUZZ BALANCE(hehe)
Have fun!
Description
Comments (1)
@vassid
1. first is that the head is turning away - all images in dataset depict that character is turning head
2. second one is that with weight above 0.70 images that generated are getting very close up and might be cropped! - this is exactly what is in Your data set, all imgs are close ups and LORA did learned that and that's why @ str 1 You get this behaviour.
So lora was trained correctly according to your data set. Simple rule of thumb is: garbage in => garbage out.
Read around how to create datasets, few pointers:
1. diverse data set, larger != better
2. multiple shots of subject with different angles
3. different environment and lightning conditions
4. mix of full body and close ups
5. no duplicates or very similar images
6. images don't have to come from same character, i.e. head from widow and body from different character if You can't find whole body shoots
7. dont mirror head, mirror if using just half of the head as training image
8. check multiple checkpoints, last one is not necessary best one (could be overbaked)
Hope this helps a little and good luck.