Hey this is actually a problem I've been having. Thanks for mentioning it. I just learned I need to put it in the VAE folder to make it show up on the list.
btw, have you tried using --no-half-vae? it helped me to get rid of black pictures when generating using novelai, anything, etc.
I wouldn't recommend using the Anything VAE since it would cause some images to be black. Most of the time they would be fine but once every, say, 50-60 images I would get a completely black square.
Switching to vae-ft-ema-560000-ema-pruned.ckpt fixed the issue for me.
Hey this is actually a problem I've been having. Thanks for mentioning it. I just learned I need to put it in the VAE folder to make it show up on the list.
I wouldn't recommend using the Anything VAE since it would cause some images to be black. Most of the time they would be fine but once every, say, 50-60 images I would get a completely black square.
Switching to vae-ft-ema-560000-ema-pruned.ckpt fixed the issue for me.
Interesting. I made the exact same model, with an identical hash but with absolutely identical settings and prompt I get not even close to the same art as in the post.
Also, there seems to be some problem with the colors
I can't get the result you're getting. Even though hashes match, models might differ due to the weird way they are calculated. In my experience the hash stays the same no matter what weight you use with the add difference method. So either one of us might have made a mistake there (wouldn't be the first time I messed up writing instructions for a mix). You can download the model I'm using from https://mega.nz/folder/XMUzWIAL#i52o1QYOx7j1neujUJzfWw as mix__6.
I'm also using the latent upscaler for the highres fix. You probably won't get the exact same image because of my --xformers but you should get close.
As for the colors, I think used the vae-ft-mse-840000 one. As redjoe said, it should fix your color problems.
Recently seen some models mention a 'needed' use of different clip skip variables and I wanted to know more about them. Happened to come across this example, thanks 👌
Glad I could help. If you're going to be working with CLIP a lot, I recommend adding it to your main interface. Go into WebUI settings| User interace| Quicksettings list: Add CLIP_stop_at_last_layers (put a comma between each argument here). Refresh and it'll be up top next to your model selector.
Looks great! Would you be willing to share the embedding? How many images did you use to train it? I want to train a Yoimiya embedding, and I selected around 70 images. I wonder if that's enough.
Recently seen some models mention a 'needed' use of different clip skip variables and I wanted to know more about them. Happened to come across this example, thanks 👌
Interesting. I made the exact same model, with an identical hash but with absolutely identical settings and prompt I get not even close to the same art as in the post.
Also, there seems to be some problem with the colors
This is a problem known as "bruising" (the little purple spots here and there). To fix it, go to WebUI settings| Stable Diffusion tab| SD Vae. Set it to anything-v3.0.vae or nai.vae (I'm pretty sure these are identical). I have no idea if this will make your image identical to AA's, but it will fix the bruising and desaturation.
Interesting. I made the exact same model, with an identical hash but with absolutely identical settings and prompt I get not even close to the same art as in the post.
Also, there seems to be some problem with the colors
... I don't like ads too, but it's technically not against the rules as far as I know. (I agree we could add large watermarks to the prohibited content list.)
It's not against the rules, but I'd agree it probably should be.
Posts like the two from this user so far just seem like advertisements for their patreon and people don't come here to look at stuff that only serves to promote something like that.
I'll probably get downvoted for saying this, but I do have a fanbox where people do support me. I'm sure there are some links to it even on this site and yet I don't really feel like a scumbag T_T Calling out aggressive advertisement like in this post is fine. I don't like ads too, but it's technically not against the rules as far as I know. (I agree we could add large watermarks to the prohibited content list.)
I meant the model. You tagged this one as anydream when I'm fairly certain you meant to tag anything
Jesus, I didn't even catch that. Indeed, I did mean anything, because this image (most of it) was made in the same batch as the others I uploaded a few days ago! Thanks for pointing that out
I was about to say that you should be able to check the metadata to see the tags used for making it but it seems I forgot to give the tags for the non-ai scaled version, because I'm certain I had no extra hands but oh well.
Okay I just checked and I did tag no extra hands but didn't emphasize it so maybe that's why
I meant the model. You tagged this one as anydream when I'm fairly certain you meant to tag anything
I suspect that you have mistagged the model you used on this one
I was about to say that you should be able to check the metadata to see the tags used for making it but it seems I forgot to give the tags for the non-ai scaled version, because I'm certain I had no extra hands but oh well.
Okay I just checked and I did tag no extra hands but didn't emphasize it so maybe that's why
oh come on, the floating plates may be unrealistic, but at least they are well made, no anatomy errors or anything like that. I could say that a girl flying is unreal and that alone is also a bad image, come on, that's stupid.
if it is well done, without serious anatomical errors, it deserves a second chance.
There's a difference between disinterest & poor quality.
oh come on, the floating plates may be unrealistic, but at least they are well made, no anatomy errors or anything like that. I could say that a girl flying is unreal and that alone is also a bad image, come on, that's stupid.
if it is well done, without serious anatomical errors, it deserves a second chance.
There were a few things in the original that I wanted to touch up. Wasn't thrilled with how her right hand ended up looking and that it wasn't holding anything. I also didn't care that whatever it was she was eating in the original kinda looked like meat which in this context doesn't make sense since she is trying and failing to make cookies. So I did some touch ups, a few more img2img runs and got a result I'm very happy with.
There were a few things in the original that I wanted to touch up. Wasn't thrilled with how her right hand ended up looking and that it wasn't holding anything. I also didn't care that whatever it was she was eating in the original kinda looked like meat which in this context doesn't make sense since she is trying and failing to make cookies. So I did some touch ups, a few more img2img runs and got a result I'm very happy with.
Mostly a test of the new openOutpainting extension in A1111. Turns out it's well worth playing with; this didn't take long at all, and I can imagine it being easy to put multiple (specific) characters in a scene this way.
@nsokitesu@nohand@aireo2 model is available on the Zeipher discord after joining press this To make it clear: It is free. I use resources other people put hard work into, I can't sell this nor would I.
Wow, this is amazing! unfortunately Zeipher shut down his discord, so there no way for me to get the model. Is there any way you could share the model or your procedure an how to make a similar one?
