- This topic has 18 replies, 4 voices, and was last updated 11 months ago by
.
-
Posts
-
Since I haven't posted in a while 😅 I might as well post this.
I hope you enjoy it as much as I did making it (some Photoshop added to fix hands stuffs):
Lovely work Legolas 💕.... and nice to see a non-DS image in the art showcase. I've been Lora training lately and it's as temperamental as working in SD. Lots of HQ images can sometimes produce poor Loras and fewer medium quality ones can turn out good using the same settings. It can be quite frustrating 😤.
@Frank22 Thank you, Frank. And, I I agree, we need a few more non-daz images here, I'll try to add more (I am loving the Anime/Graphic Novel styles).
As for LORAS, I have never trained one, and I agree that Loras can be really hard to make properly and, of the tons of them I tried using, only a very small number of them I found useful (one for composition, and one for hands, and a few others for style).
I also read a new technique about LORAS training of complex images that seems pretty interesting (careful, it's NSFW):
https://civitai.com/articles/3870/guided-training-with-color-associations-for-complex-concepts
This one is about how to better use Loras for image generation (this one is SFW):
https://civitai.com/articles/4072
which refers to this extension for more LORA control (also worth a read):
https://github.com/cheald/sd-webui-loractl
Hopefully, you'll be able to find some useful information there to help you out.
Thanks for those links, I have seen some of them.
With character Loras it's really important to know which chpt they were trained with as that makes all the difference. Many on Civitai don't provide this info so I don't bother with them. With styles or clothing Loras that's less important.
My go-to ones I always use are (I'm still using 1.5)-
Detail Tweaker
Bad-Hands-5
EasyNegative
Beyond Negative@Frank22 I think quite some people are still using 1.5 mainly because of the massive GPU power needed for SDXL. 🙁
Thankfully, I was able to fully transition into it, and the amount of inpainting I have to deal with to get good results is really minimal. Also, now SDXL has matured quite a lot. Maybe next year with the new RTX5090 we will get a new gen of image generators beyond SDXL 🤩.
I've got a decent rig now with a 4070 but the reason I stay with 1.5 is that I have a large amount of 1.5 resources that would be useless with SDXL. I also don't like how Loras for SDXL are often huge files often 10x that of 1.5. Civitai is gradually moving over to SDXL so I might have to reconsider at some point.
I've never had much luck with inpainting so if the hand/s are screwed up I usually just re-run the seed at a slightly different variation strength. ADetailer seems not work as good as it used to after updating it. Don't know why that would be.
You should try Lora training as while it can be frustrating it can be good fun. Once you get the hang of it it's relatively easy. If you want to try I can give you some tips on Kohya, cheers.
@Frank22 Thanks for the offer, but there are so many LORAS to play with, it's crazy. 😁
As for inpainting, I meant inpainting the whole image, but in small batches. I used to do that with SD1.5 and it would really crank up the quality of the images, but it would really take a very long time to do, since I was creating 4-5 variants for every detail and only keep the better one.
With SDXL I am just doing it for a few details here and there, since the High-res fix makes things really good anyway; i am using:
8x_NMKD-Superscale_150000_G
As for 1.5, I used to have a billion LORAs there too, but i just deleted them all with time. Thankfully, I have some pretty spacious HDD so the SDXL Loras are not a problem for me (nowadays they have mostly decreased in size anyway, starting from 200mb).
This text to video model is quite impressive - Sora.
@Frank22 Yes, I saw that, and it looks amazing. Hopefully we'll get a free one soon too 🙂
Also, there is already the next gen text2img from the same makers of SDXL, called Stable Cascade.
Apparently, it's faster, follows prompts better, and can do better hands (but it's only a preview, for now) so it's very limited in what it can do. Overall, I'm impressed and hopeful to see how it will perform when fully released and with all the bells and extras from the people from CivitAI.
Here are a few images I made with the preview model (unfortunately, it doesn't run on Automatic1111 yet, so I run it on Pinokio).
These were made at 1024x1024 (in seconds), seem to have less quality:
These ones I made natively at 1536x1536 (in less than 1 minute):
You may already know this but I came across the DWPose extension which is great for posing from photographs or DS renders.
@Frank22 I have been using it for months 😅 and I think I recommended it a while ago too.
@Frank22 I also tried the new Forge UI by the same guy that made ControlNet...it looks exactly like Automatic1111, and it is compatible with all its stuff, but it has DW integrated, as well as ControlNet stuff, and even has Stable Video Diffusion (which is not even available on A1111).
On top of that, it auto-detects settings of your PC (so, no weird suffixes to add to run the program) and it is quite faster too AND can already run Cascade Diffusion! (I tried installing it on A1111 and it wouldn't work).
Update: I managed to install all my extensions/loras/etc and it seems to work pretty well. I didn't encounter any problems yet.
As for following prompts, that also depends on the model being used. The cpt I usually use has no concept of certain objects as it's not been trained with them... record player, clothes hoist, electric organ etc. Mostly old/retro stuff. I'm probably expecting too much.
@Frank22 For better prompting (and better quality images) I am using the DPO lora (has both SDXL and 1.5 models):
https://civitai.com/models/242825/dpo-direct-preference-optimization-lora-for-xl-and-15-openrail
Granted, some things really need to be trained into the model, or they'll just not appear, sadly. That's why you have been training loras?
For me, I typically do just something with cool magics or some pretty girl, and for that there is definitely plenty of stuff on civitAI.
- You must be logged in to reply to this topic.
