if u get ur settings just right when using 'reference_adain+attn' - u can create something very close to ur input image, i havnt managed to perfectly replicate faces but u can get a close likeness
Thanks for showing the reference feature, yet another feature to explore! Maybe you tried them and preferred without (the results were certainly nice), but the model creator suggests orangemix, kl-f8-anime2 or Blessed2 vae's, so worth a look if you didn't. For initial generations I find restore faces generally worth enabling too.
I mean...nice result but it didn't even get the poster right. It ignored the hand on the face completely...not sure what were you trying to showcase. Still, thanks.
@@lemonscript it was a good showcase, most people want a bit of diversity and different ideas anyways. If commenter wanted to keep the hand on the face, they can also use a 2nd control net preprocessor like OpenPose
Hey, thank you! To be honest im not to sure what clip skip exactly does, but I feel like im getting consistently better results with clip skip set to 2. Less wonky eyes etc. About the blurryness, im not to sure if clip skip can help there, did you try out the preprocesser I showed in the video? Maybe that might help. Also try playing with the style fidelity To have the clip skip bar at the top, go to settings -> user interface and search for [info] Quicksettings list. Inside there type CLIP_stop_at_last_layers. Hit save and restart UI
Clip skip removes last X iterations of sampling of image. Different X gives really different results. Often the realistic models batter with clip skip 1, when more animated models are batter with x = 2. Anyway, I suggest to try it by yourself to fill on fingertips.
Good point, I tried using realsiticvision, but the results were kinda messy when trying out this preprocessor, similar but werid faces, weird hands etc. So I just gave up on that Idea and went with anime style, but yes you are right, going with a realistic style would be better
@@lemonscript tbh I’ve noticed the reference pre processors don’t work too well at recreating people. Sometimes they get close but you need to run through a ton of batches to find ones that really do it. Maybe for animated characters it works better though.
you fooled me with the intro thinking it's going to go from anime to photorealistic and now I'm disappointed so now you owe us a video actual doing that ;)
Hahahaha I can certainly see why you got fooled 😂 Not a bad idea, Im getting lots of recommendations of people take lets say ruffy from one piece and displaying them in a different style e.g realistic or studio ghibli. I will play around with it and maybe release a video if I manage to do it!
Hey, I did not aim for 100% similarity. I'm only showcasing that with the new preprocessor, just providing a single short sentence like "woman with brown hair" is enough to get a pretty similar results. You can of course improve the result by a ton by providing a better description of what you want. Event better, if you try to mimick a picture that you previously generate (say you have the original prompt) you can just take it, plug it in again, change it, and have almost the same character in lets say a different scene
Automatically subscribed when I heard you be the only StableDiffusion RUclipsr to pronounce "Euler" correctly
if u get ur settings just right when using 'reference_adain+attn' - u can create something very close to ur input image, i havnt managed to perfectly replicate faces but u can get a close likeness
can you share your setting?
yes please share
Share it
what does Clip Skip do? I keep seeing people use it but haven't heard anyone say what it does
Check the comments under the comment of @alonso garrote in this video, @Артем Суворов explained it pretty well!
what is the theme you are using on your interface? it looks so clean!
Hey im not using a specific theme, just basic control net in dark mode
Can you share what your pc specs are?
Thanks for showing the reference feature, yet another feature to explore! Maybe you tried them and preferred without (the results were certainly nice), but the model creator suggests orangemix, kl-f8-anime2 or Blessed2 vae's, so worth a look if you didn't. For initial generations I find restore faces generally worth enabling too.
Hey and thank your for the advice! I will certainly try and experiment more 💛
thank you! I have been using reference_only but this one has helped me!
I'm glad I could help you out! 💛
Nice one. Thnx for the highlight.
You're welcome! Always happy to be helpful !💛
I mean...nice result but it didn't even get the poster right. It ignored the hand on the face completely...not sure what were you trying to showcase. Still, thanks.
I was trying to showcase that with the new preprocessir its possible to get good results with very short prompts, nevertheless thank you!
@@lemonscript it was a good showcase, most people want a bit of diversity and different ideas anyways. If commenter wanted to keep the hand on the face, they can also use a 2nd control net preprocessor like OpenPose
I like your enthusiasm, just like a child playing with his new toy
Cheers bro
Hi there, nice video!, what does that clip skip do?
i tried ref only on fantasy cars but could only get blurry nonsense, any ideas?
Hey, thank you! To be honest im not to sure what clip skip exactly does, but I feel like im getting consistently better results with clip skip set to 2. Less wonky eyes etc. About the blurryness, im not to sure if clip skip can help there, did you try out the preprocesser I showed in the video? Maybe that might help. Also try playing with the style fidelity
To have the clip skip bar at the top, go to settings -> user interface and search for [info] Quicksettings list. Inside there type CLIP_stop_at_last_layers. Hit save and restart UI
Clip skip removes last X iterations of sampling of image. Different X gives really different results. Often the realistic models batter with clip skip 1, when more animated models are batter with x = 2. Anyway, I suggest to try it by yourself to fill on fingertips.
thanks bro. keep it going!
A bit confused why you would demonstrate with animated characters rather than demonstrating by creating photos more similar to the original
Good point, I tried using realsiticvision, but the results were kinda messy when trying out this preprocessor, similar but werid faces, weird hands etc. So I just gave up on that Idea and went with anime style, but yes you are right, going with a realistic style would be better
@@lemonscript tbh I’ve noticed the reference pre processors don’t work too well at recreating people. Sometimes they get close but you need to run through a ton of batches to find ones that really do it. Maybe for animated characters it works better though.
You never actually added the negative prompts
You should try it with no prompts and let it use the source image as your template.
you fooled me with the intro thinking it's going to go from anime to photorealistic and now I'm disappointed so now you owe us a video actual doing that ;)
Hahahaha I can certainly see why you got fooled 😂 Not a bad idea, Im getting lots of recommendations of people take lets say ruffy from one piece and displaying them in a different style e.g realistic or studio ghibli. I will play around with it and maybe release a video if I manage to do it!
I didn’t quite get the purpose, you used a photo to generate a painting. Good video nevertheless
sorry but i dont see the similarity
Hey, I did not aim for 100% similarity. I'm only showcasing that with the new preprocessor, just providing a single short sentence like "woman with brown hair" is enough to get a pretty similar results. You can of course improve the result by a ton by providing a better description of what you want. Event better, if you try to mimick a picture that you previously generate (say you have the original prompt) you can just take it, plug it in again, change it, and have almost the same character in lets say a different scene
@@lemonscript okay cool. But you can create similarity. I tried it. It's really cool. Much like image prompt in MJ
No similarity. It's a bite.
Hey, thanks for the reply. I think the results are pretty similar, considering I only provided the positive prompt "woman with brown hair"
i got bored with ai so fast...
Why is that? I find it fascinating