- This topic has 33 replies, 9 voices, and was last updated 1 year, 2 months ago by
.
-
Posts
-
@Davizca87 - "As someone who is training daily models...."
So can I pick your brains then?
Is the Kohya-ss Extra Networks script necessary to use Kohya trained Loras in a1111? I don't want to install a script that might break my SD installation if I don't have to. Can I just put the Lora made with Kohya into the SD Lora folder or won't it work?You can PM me if you don't want to reply in this thread.
@Davizca87 I have to agree with you. Although, in my renders I am using a combo of img2img and ControlNet (Canny).
I have tried a text2img with Canny (0.7 seems to give me best results)....and wow! it looked pretty good 😍
@frank22 The problem is that the extension of Kohya-SS Extra Networks is deprecated and caused conflicts in the past cause used a certain type of inference on the seed generation of LyCoris that actually is a no-anymore.
In fact it had a problem I addressed some months ago with not LORA's but yes with LyCoris and in specific with LoCon files.
That's why Automatic1111 fixed it so you theorically won't need anything other than to put your LORA's in models LORA (same as LyCoris types too)
And as recommendation to anyone using SD or SDXL, go for ComfyUI or Fooocus (this last is very useful if you don't want to mess with nodes). It's more difficult at first but what you can do there surpass by far what you can do in Automatic1111. Insane control and making your own workflows is totally useful.
Hope this help!
@Legolas18: Yeah, and nowadays with ComfyUI + ControlNet IP Adapter with clipvision you can do crazy things. There are a lot of useful ControlNets.Wow, this ThrEaD has gone deep.
Will have to revisit all of this if/when getting local install working. (favorited)
And in meantime it will prob all change 10 times over again.
(But then, that's for the better with AI as it is truly constantly improving.)Example, no matter what you think of AI, the images @legolas18 has posted in here (and elsewhere, and others) are dang gorgeous in their own right.
As Always, i am a bit slower then the rest 🙂
Tried to copy what Legolas did. I managed in the end, but it was still trial and error. But the only way to learn is to do, and....try to remember what the #*&@! I did
Anyway, AI versus DS
linkSTill think there is more "character" in the DS version, but the hair is much better in SD. Time these two merge ...
Thanks @Davizca87 for that answer. I asked because I was looking at this tute and it said I had to use the Extra Networks extension. 18 mins into this video, stamped for you.
@gaver I had another go at using SD on your image. It uses a combination of ControlNet's Canny and Softedge_hed. Here is a very in-depth ControlNet guide, btw:
@frank22: Yeah, that video is way too old nowadays we have better ways to training and in fact using Automatic1111 is not giving good results. We use Kohya_SS, Everydream, etc. I'd reccomend Kohya_SS because it's easier to get started (although it has insane amount of settings, optimizers, schedulers and variables so you need a lot of experimentation).
By the way, you can always downgrade from one commit (ver.) to another. I did that when Auto was breaking extensions I was using.
@Legolas18: You can combine several ControlNet, one popular is OpenPose to control the pose + Depth/Canny. The more you use ComfyUI, better you will understand the inferencing although the seed generations differs from A1111 though.Thanks @Davizca87. So if I use Kohya to train a Lora do I need to install the Extra Networks extension in SD a1111 to use it? Or not?
@frank22 You went and linked a 8 old month video from youtube on AI 🙁
That's like ancient 😅
Go to the same user (Aitrepreneur) and watch his last 3 videos. They are all about how to train LORAs 😁👍
@frank22 I have no idea, I never did any LORA training 🙁
- You must be logged in to reply to this topic.
