Fan Art Summertime Saga - Fan Art

5.00 star(s) 1 Vote

Microtom

Well-Known Member
Sep 5, 2017
1,152
4,245
Any thoughts on creating unified LORA with all characters combined ?
The base SS Lora is that. There's a problem with it however. The Lora is trained for free on civitai AI, and they have a 10k step limit for training. The more images your dataset contains, the more steps you'll need.

Including all or most of the characters causes the training to require more than 10k steps. As it is, the current Lora is undertrained.

To solve this, we can purchase GPU time to do our training somewhere else than civitai, or train locally with a beefy GPU.

When SD3 comes out, I expect to purchase GPU time to create a new and improved lora. However, it firsts needs to come out. We also might need a sexual base checkpoint like pony diffusion or we'll waste our time. And the full tech update has to release so we have access to high resolution images.
 

Wardonn

Member
Jul 21, 2017
267
4,125
Any thoughts on creating unified LORA with all characters combined ?
I thought about it, but I don't know if I have enough experience. Maybe I will try it sometime in the future, but I have to learn some things. I'm not sure that one big model will be better than the individual small models. And I will probably come across the limits of civitai. Well, we will see.

Now I want to complete the female characters. There are not many of them anymore. And then I want to relax a little. I've been doing it for almost two months at a time every day. It climbs my brain a little. I want to do something else too.
 
  • Like
Reactions: k4me and Amnesion

phoenixman123

Member
Feb 10, 2019
478
4,215
The base SS Lora is that. There's a problem with it however. The Lora is trained for free on civitai AI, and they have a 10k step limit for training. The more images your dataset contains, the more steps you'll need.

Including all or most of the characters causes the training to require more than 10k steps. As it is, the current Lora is undertrained.

To solve this, we can purchase GPU time to do our training somewhere else than civitai, or train locally with a beefy GPU.

When SD3 comes out, I expect to purchase GPU time to create a new and improved lora. However, it firsts needs to come out. We also might need a sexual base checkpoint like pony diffusion or we'll waste our time. And the full tech update has to release so we have access to high resolution images.
In other words, be happy with what we have now as it’s a big task to do something that we can do currently do with good results.
 
  • Like
Reactions: Minecraftnoob

Wardonn

Member
Jul 21, 2017
267
4,125
The base SS Lora is that. There's a problem with it however. The Lora is trained for free on civitai AI, and they have a 10k step limit for training. The more images your dataset contains, the more steps you'll need.

Including all or most of the characters causes the training to require more than 10k steps. As it is, the current Lora is undertrained.

To solve this, we can purchase GPU time to do our training somewhere else than civitai, or train locally with a beefy GPU.

When SD3 comes out, I expect to purchase GPU time to create a new and improved lora. However, it firsts needs to come out. We also might need a sexual base checkpoint like pony diffusion or we'll waste our time. And the full tech update has to release so we have access to high resolution images.
Techupdate has a greater resolution, but with Upscale I will achieve quite good results. If you want to try the new better model and are interested, I would share my datasets. That took me the most work and time.

New Tech Update:

new_1.png


My upscale from old art:

upscale_1.jpeg
 

Microtom

Well-Known Member
Sep 5, 2017
1,152
4,245
Techupdate has a greater resolution, but with Upscale I will achieve quite good results. If you want to try the new better model and are interested, I would share my datasets. That took me the most work and time.

New Tech Update:

View attachment 3698195


My upscale from old art:

View attachment 3698202
Yeah sure, I'd like to see one of your lora dataset. It's true that in this example, DC's art looks much blurrier compared to your upscale. The lines however have a lighter interior in your upscale version. Here, for example:
Untitled-2.png

We can probably fix that with img2img or manually by combining the two images, or manually drawing it.
 
5.00 star(s) 1 Vote