Fan Art Our Red String: Fan Art

5.00 star(s) 2 Votes
Discussion Reviews (2)
TIGER ACE

TIGER ACE

Member
Jan 9, 2019
126
514
Here is the final result, I would like if some of you who uses Stable Diffusion locally on your machine to try and post result and what you guys think, I dont want to put it on Civit Ai and put a new version there if its not needed, but once uploaded it will help the people who uses civitai to generate the images.

I think its better and doesnt have the problems of the last version. But its also over trained, you might have to turn down lora weight to 0.75 or 0.7 but start from 1.0

The faces arent always great on first run still, if its not pov or upclose stuff, but inpainting or upscaling does make it better, to me characters work really well, putting Louise will get you a red head girl with a choker everytime.

You must be registered to see the links


as I was writing I did discover the problem, I think there is an issue here, which is the girl names are acting like activation tags, if there is no keep token basically a fixed tag, every tag activates lora, except I needed to train the characters so I kept their name as keep token, it seems to me, that maybe each girl is acting like their own style? I dont know Minnerva and Holly seems to have suffered from this, but yea I will need to see what happens, maybe you guys can try and tell me. same with all the girl, in example put in Ivy and the image comes back purple, How did I forget I used 2 keep tokens last time, because I foresaw this, but forgot this time. As always, I am not an expert on anything, pretty bad on my part.
 
  • Like
Reactions: lostsoul76, Vyacheslav Grinko, lostnoir and 13 others
CaptainBlt

CaptainBlt

Member
Jun 10, 2019
245
5,019
TIGER ACE said:
Here is the final result, I would like if some of you who uses Stable Diffusion locally on your machine to try and post result and what you guys think, I dont want to put it on Civit Ai and put a new version there if its not needed, but once uploaded it will help the people who uses civitai to generate the images.

I think its better and doesnt have the problems of the last version. But its also over trained, you might have to turn down lora weight to 0.75 or 0.7 but start from 1.0

The faces arent always great on first run still, if its not pov or upclose stuff, but inpainting or upscaling does make it better, to me characters work really well, putting Louise will get you a red head girl with a choker everytime.

You must be registered to see the links


as I was writing I did discover the problem, I think there is an issue here, which is the girl names are acting like activation tags, if there is no keep token basically a fixed tag, every tag activates lora, except I needed to train the characters so I kept their name as keep token, it seems to me, that maybe each girl is acting like their own style? I dont know Minnerva and Holly seems to have suffered from this, but yea I will need to see what happens, maybe you guys can try and tell me. same with all the girl, in example put in Ivy and the image comes back purple, How did I forget I used 2 keep tokens last time, because I foresaw this, but forgot this time. As always, I am not an expert on anything, pretty bad on my part.
Click to expand...

I think the result is pretty good, thanks for a great job
You don't have permission to view the spoiler content. Log in or register now.
 
  • Like
  • I just jizzed my pants
  • Love
Reactions: LHDLLB, lostsoul76, dhuemnptai and 25 others
Emoon

Emoon

Member
May 4, 2017
146
431
TIGER ACE said:
Here is the final result, I would like if some of you who uses Stable Diffusion locally on your machine to try and post result and what you guys think, I dont want to put it on Civit Ai and put a new version there if its not needed, but once uploaded it will help the people who uses civitai to generate the images.

I think its better and doesnt have the problems of the last version. But its also over trained, you might have to turn down lora weight to 0.75 or 0.7 but start from 1.0

The faces arent always great on first run still, if its not pov or upclose stuff, but inpainting or upscaling does make it better, to me characters work really well, putting Louise will get you a red head girl with a choker everytime.

You must be registered to see the links


as I was writing I did discover the problem, I think there is an issue here, which is the girl names are acting like activation tags, if there is no keep token basically a fixed tag, every tag activates lora, except I needed to train the characters so I kept their name as keep token, it seems to me, that maybe each girl is acting like their own style? I dont know Minnerva and Holly seems to have suffered from this, but yea I will need to see what happens, maybe you guys can try and tell me. same with all the girl, in example put in Ivy and the image comes back purple, How did I forget I used 2 keep tokens last time, because I foresaw this, but forgot this time. As always, I am not an expert on anything, pretty bad on my part.
Click to expand...
What is the base model you train the Lora on? This can have a significant impact with certain Characters or Styles as keywords like 'Jessica' or 'Holly' for instance may trigger data in the base model. It's always wise to use unique tags for specific features in the lora, like J3ss1ca or H0lly and the like. This way you can tell the AI that you want to specifically trigger training data from the lora. It may also help you to work out specific characteristics, like adding 'mature female' to Minnerva tags and for others using 'young adult' or using tall man for Axel and chubby man for Stan and the like. You can utilize the tags from
You must be registered to see the links
entry to your advantage to work out differences.

Below are some examples done with AutismConfetti and no additional tools like face fixers or control net, just simple prompt -> sampler -> output. And for that they turned out pretty good. It sticks to the artstyle with a acceptable bias from the underlying model which can be easily fixed/reduced with a soft style transfer via control net. Base pony sucks at faces who are not upclose so you can only do so much through a lora to remedy this issue. It's better to let ADetailer or other face restoration tools deal with that, instead of focusing on your end on it. The only issue i noticed is the mistletoe with Holly, it happened for me and in CaptainBlt's image above as well.

But all in all you're doing a great job so far and i highly appreciate your efforts. I know what a time consuming process it is to create a lora! <3
 
Last edited:
  • Like
Reactions: LHDLLB, goldenboy812, dhuemnptai and 12 others
TIGER ACE

TIGER ACE

Member
Jan 9, 2019
126
514
Emoon said:
What is the base model you train the Lora on? This can have a significant impact with certain Characters or Styles as keywords like 'Jessica' or 'Holly' for instance may trigger data in the base model. It's always wise to use unique tags for specific features in the lora, like J3ss1ca or H0lly and the like. This way you can tell the AI that you want to specifically trigger training data from the lora. It may also help you to work out specific characteristics, like adding 'mature female' to Minnerva tags and for others using 'young adult' or using tall man for Axel and chubby man for Stan and the like. You can utilize the tags from
You must be registered to see the links
entry to your advantage to work out differences.

Below are some examples done with AutismConfetti and no additional tools like face fixers or control net, just simple prompt -> sampler -> output. And for that they turned out pretty good. It sticks to the artstyle with a acceptable bias from the underlying model which can be easily fixed/reduced with a soft style transfer via control net. Base pony sucks at faces who are not upclose so you can only do so much through a lora to remedy this issue. It's better to let ADetailer or other face restoration tools deal with that, instead of focusing on your end on it. The only issue i noticed is the mistletoe with Holly, it happened for me and in CaptainBlt's image above as well.

But all in all you're doing a great job so far and i highly appreciate your efforts. I know what a time consuming process it is to create a lora! <3
Click to expand...
Thanks for tips, I appreciate it. In this case I was lamenting me not putting an activation tag, I think my tagging works for the characters so not too unhappy about that, I removed the activation tag from last version and didnt use it just character tag. I was thinking that I should have used it to make it clear that for the model, as the character tags are the first tag and kept in first place during training, it can create the issue of becoming different style. I did it last time, but not this time.

I use base pony to train to keep it compatible with most models and other style loras.
 
  • Like
Reactions: Emoon
You must log in or register to reply here.
Top Bottom
5.00 star(s) 2 Votes