klein9b
Use change into shexyo style
Lower strength keeps face shapes closer to original
Works on anything including lineart
qwen-edit
Changes images-lineart to shexyo stylechange photo into Shexyo style color lineart into Shexyo style
Known issues : adds steam , doesn't details certain stuff
Reduce strength add other loras + promp words to controls what gets added
Description
FAQ
Comments (10)
amazing lora! can you make a any photo to dragon ball super anime style lora as well?
I've heard it's notoriously difficult to train Edit 2509 for styles. This result is impressive. Way to go.
Ty :)
Was curious as well , seems to be easier than flux kontext from what I heard.
@blo01 Would you mind sharing your training settings and experiences? I wanted to train for 2509, but I was being convinced not to do it by people who talk about how hard it is. Anyway, I know it's a lot of info to share, and I understand if you don't feel like it. I'm just impressed and would like to give it a go too.
@Jellai Yeah people sleeping on it , but it is a bit of trial and error,still testing .
If a few words this is what I did:
1 collect 20-30 images
2 convert to realistic style , train at 512 res for 35 epoch
3 convert to lineart initial dataset , resume to train on top at 512 res in reaching 70 epoch
4 take both datasets train on top at 1024 res reaching 80-90ep
Mixing both messes them up at step3 , but merging datasets and training at higher res kinda helps. Idk style already appeared even after step2 at 20ep , then you could skip step3 altogether.
I just wanted to see if showing it lineart would improve anything.
You can start at 256 as well and bump to 768 , qwen does not crop image at higher res and I did see quality bump after step4 , so it likes higher res data.
I will maybe add face detailing on top this one in next ver.
@blo01 Oh good. This is exactly what I was already doing, up to a point. So, let me get this straight...
1. You collect the original art.
2. you converted line art to realistic. That's what I've done.
3. But then you convert the realistic conversions back to line art? Or are you saying that this is just how you test the training? Meaning you compare to the originals to judge how well it has trained.
If it's not just for testing purposes, why convert back to line art? Don't you already have the line art you need? It's possible I misunderstand what you're saying.
@Jellai
Original > Realistic - trained , tested , it worked fine on photos but didn't do that well on flat/anime/lineart .
Original > Lineart - trained on top step 1 so it could work on both.
Most likely you could just do it in one step and would get same result.
@blo01 Oh I see. Smart. I'll give it a go. Thank you.
awesome. hoping to see more stylized styles !
so this lora makes everyone sweat and tremble? nice)