v2:REBIRTH

dont blame me for this. set primary model type in your profile settings

@s6yx, my settings already are... but maybe the problem is in the site itself. Even the first link in files section gives me 5.1Gb file.

@Devilcraft again the gripe is on the website not me

@s6yx ok, that wasn't my goal to blame you or someting like this, just the problem you could help with perhaps. Thanks for great model!

@Devilcraft only way to fix it would be to delete fp32 model

it was taken off huggingface, i use the one i have locally

@s6yx Has anyone reuploaded it, to like google drive or mega.nz? or can you upload it so people who were late to get it, download it?

use 1.2.2

@s6yx thanks, using it and I really like it, but I wish I could figure out what it is with 1.2.1 cause I'd like to try it.

weights and proper prompting: [large breasts], (small breats:1.2) negative prompts: (large breasts:1.2)

@s6yx pro tip. use tiny if small is not good enough

@IgorGu Thank you!

i cant do anything about that, The site regulates it

that's true, people complain that my arts always have the same face, I try to vary the hair style, color and etc.

The good side is that the face is beautiful lol

While I can use embeddings or loras to get different faces and hair to get other likenesses, they ALL have the same super pointy chin. I havent figured out how to get a different chin..

You need to use proper prompting to change it. For example, Having multiple characters at different weights works extremely well and should look similar to this "(looks like x:0.7), (looks like y:0.8), (looks like z:1.0)". You have to check each one individual first in order to see if the model actually knows that character; real or fictional. You can also just change nationality/ethnicity; instead of "cute woman" try "(cute Turkish woman:1.4).

If you really, reeeealy want a very specific face then LORAs, and photoshop's liquify tool (*cough* VPN, *cough* 1337x) will work fine. I also think I heard somewhere about face and expression features being made for controlnet, though I might be wrong on that one.

working on this, I recreated basil_mix recipe trying to introduce a more diverge range. Will see after testing

You have to throw every LoRA you have at it, some of them can magically work - but yeah, it downright ignores most of them.

Depends on your prompts. Give an example on what you put in your positives and negatives.

Yeah, I run 2 tabs at the same time with nearly identical prompts, and one of them will simply be thick with flowers all over. I guess I could nail down what is triggering it if I went through the trouble, but there are very few words different between both prompts. It's cute, actually.

Put "flowers:1.2" in negatives

should

@s6yx with 1.22 I can only install the inpainting model, it keeps saying:

">> Optimizing revAnimated_v121 (30-60s)

| Loading diffusers VAE from stabilityai/sd-vae-ft-mse

| global_step key not found in model

| Using replacement diffusers VAE

** Conversion failed: Error(s) in loading state_dict for CLIPTextModel:

Missing key(s) in state_dict: "text_model.embeddings.position_ids"."

I will try an earlier version.

Also, with 1.22 in SD I can see rendering progress, when the image is finshed I often get a black or pink images which also is downloaded to the HDD. Any ideas?

Besides that, as far as I could see, this is very amazing and I'd like it so much to have a stable version running, especially on Invoke. Thanks for all the effort.

1231231123

stable-diffusion-webui/embeddings

might have to just photoshop it on the pics you save. Be pretty easy, just crop the eyebrows to new layer and hue shift it to the color you want them to be.

Maybe this extension will help?

https://github.com/hnmr293/sd-webui-cutoff

@dimsum88 Hmm, that looks intresting. I haven't tried it. I also am a really big fun of photoshop though lol. It looks like the extension could work though. Don't know how you would prompt it in with the tokens. I might experiment with it.

try moving the words for the eyebrows at the start, then give them more weight something like this (blonde eyebrows:1.4)

you guys never heard of inpainting?

Yeah use it a lot on hands and feet even with embeddings

you're free to use another model instead

They are a crapshoot - and they are with every model out there. Some models work harder to hide hands, this one doesn't. Just gen in bigger batches - the ones that work are very much worth it. Again, show me a model that does 2.75D as good as this that makes better fingers. I'll wait.

They're all terrible with that.

Either you have settings very out of whack possibly not covered by the prompt info, or you are missing something you have done in your settings. This model is extremely robust - I run up to 5 LoRA's and fight with things like odd postures, but the faces are rock solid. And trust me, my prompts are a mess.

You need to directly implant the full prompts out of a PNG Info drop, and then make sure you are using the same sampler, hires fix settings, CFG, noise, all that. Make sure to gen at a clean res like 512x768 or 768x512 (I scale that 1.5 with hires & 4X Ultrasharp but this model isn't picky about upsampler), I also find that steps work great even at 20-25 using DPM++ SDE Karras (both norm and hires).

ps. I run Clip Skip 2, but I don't know if I need to.

Weird red eye outlines can often be caused by using a VAE that is a bad match for the model.

@kaibosh thanks for the advise. I will check out what else I may have done wrong.

@OctopusHugger i didn’t realize. I’m using one of the recommended VAE but maybe I should try a different one. Thanks for the tip!

4xUltraSharp

4x_foolhardy_Remacri works well

I found ribs show up often with the prompt 'skinny' or 'slim'. Can add ribs to negative prompt to help with that. As for prominent muscles, negative prompt is again the best solution. I usually add something like "muscular, abs, six pack abs" to negative prompt. Can also add weight to the negative prompts if simply adding them doesn't give you the desired effect.

If I'm prompting "skinny" I get ribs a lot, too. "Smooth physique" seems to work pretty well. Also "slender thighs" or whatever.

the latter

@s6yx  thanks

And did you really read the description thoroughly? I assumed you did not.

when you use upsampler u must tone down the denoising strength to 0,1 or lower and not use latent. It would always rip the picture apart at the end

Not without permission from me

@s6yx hi, how to get your permission？

Thanks for the answer @s6yx , I would like to use your model on a private web app that do not charge for the service. Could I have your permission? Also I would suggest to change the license because the CreativeML Open RAIL-M does not include such restriction so if I didn't read your description in details I would have probably missed your point.

I can't seem to get it to generate anything much older than mid 20's kind of stuff. Very odd.

@Gray_Geist "Salt and pepper hair", "older man" helps me gray up some Daddies.

Oh please do. Your model is the most amazing one that I've used. Idk what kind of sorcery is that but it never fails to generate jaw-dropping results!

get a break, man. i love your models!!

me too bro

From what I've noticed its caused by many things. loras, Positive & negative prompts that change appearance, and smaller things that all add up. Something else that greatly contributes to deformed faces (especially eyes) is 1) multiple people 2) not looking at the viewer 3) anything that isn't a portrait,

The best way to fix imo is simply by inpainting the faces. you're going to have to mess with the settings yourself to get what best fits the over all image but heres some starting points:
Strip back positive prompt to the essential stuff
Mask blur: <5
Masked content: ORIGINAL
Inpaint area: ONLY MASKED
padding: 150~
steps: varies but I stick to 40-70
method: same as txt2img
cfg scale: 5~
denoising strength: 0.4-0.6 (You'll have to dial it in)
batch count: <10 till you find what works

Last thing I'll add is prompts that prevent bad faces are really inconsistent even at high weights, negative more so than positive.

@rshipleton501 Thank you for the great comment

@rshipleton501 learning inpainting for faces changes the game

nsfw, pecs, abs, midriff, bare chest, exposed chest

-neg:
shirt, covered chest

other negs if needed: vest, sleeves, t-shirt, undershirt, pajamas

I get pretty good results with: "shirtless, pectorals" and the negatives posted by @nashidaran

Error: 'Expecting property name enclosed in double quotes: line 2 column 4 (char 5)'. Check your prompts with a JSON validator please. Full error message is in your terminal/ cli.

Go to https://github.com/camenduru/stable-diffusion-webui-colab/issues and tell Camenduru.

no

@s6yx Thanks for the answer. Then I just wonder why some already exist, just like the print-on-demand ones? Looks like i'm the only one who asked for permission. 😅

@s6yx Just wondering if there's a reason. I don't have any desires or plans to make an NFT or do anything with crypto but I wouldn't even have thought to ask. So it makes me wonder is there a specific reason? Or is it just cause that's not something you felt like at the time of this person asking?

I am curious as well considering the difference in file size.

@lordakisho file size means nothing in relation to difference of the model. Difference is less issues with distorted screwed up faces when detailing style and composition in prompts. Also fixed some VAE issues that had been happening which made it give bad generations with some other VAEs

@s6yx Thank you for that information. I'm pretty new to SD, still learning.

(low quality, worst quality:1.4), disfigured, deformed, mutilated, mutated and try to change the size 800x1000 works best for me

512x768

512x512

768x512

@UniiiChan But if I change sizes it won't really work since I want to make the animated videos for social media which is 9:16 aspect ratio, unless I'm wrong?

@s6yx  if I change sizes it won't really work since I want to make the animated videos for social media which is 9:16 aspect ratio, unless I'm wrong?

@TheEnigma  16:9 (1280x720) but most people use their smartphone upright so 4:5 is better or 1:1

Put "doppelganger" in your negative prompt. I learned this from @MakingPhoto on Youtube so credit to her. If more than one body is a big issue with a certain image that you're working on then move it up to being the first word of your negative prompt.

My personal experience is that the multiple bodies mostly happens when using hiresfix with a denoising strenght of 0.6 (and presumably above that). Decreasing it to 0.45 worked for me.

turn off restore faces and use neg prompts

Just switch to a model that can do eyes and use inpaint. No amount of neg seemed to do the trick. Synergix does ok though I am still looking for a better option.

I figured out that you can get pretty good eyes if you load up on negs and use After Detailer with a prompt like "detailed eyes, detailed pupils, round pupils" just FYI

There's nothing to stop you doing the inpaint with any other model.

if you use a normal checkpoint, it will leave blurs and lots of noise because the checkpoint isn't meant for the inpaint process, while the inpaint checkpoint has one clear image and one with noise. It gives more consistency in the generation and stops the blurs.

@Kuresuto Interesting. So instead of just noise results there are actual finished rendered images included in the model? But I think you could generate any different image from a seed depending on your prompts. If a rendered image is based on a seed, is it using it, or does it use some other rendered image in it's database that most closely matches what you have rendered?

@pkayze A seed only helps on keeping the cossistency of the colours, shapes and patterns. If you don't write the same prompt, you can have very different results on a same seed. You know how you have the denoise setting? that's mostly because there will always be noise in an image, and sd has an inbuilt system to filter/remove that noise. On a normal checkpoint the image comes inbuilt with noise for the imput on the denoise, on a inpaint checkpoint the images come separated, one clear one noised.
The only purpose is to have a more consistent generation, since the only thing we want is the masked area (normally a small part compared to the image), and we want the new generated area to be as similar with the rest as possible, without bluring.

@Kuresuto Thanks for the more detailed explanation! Interesting way to solve a problem!

got the same problem and upscaling does the trick. I can suggest to following video on youtube: https://www.youtube.com/watch?v=RiN-zrUlneQ

Is inpaint version just as good as original?

@akkg8668866 Inpaint is for inpainting and outpainting (original can't be used for outpainting and very limited in inpainting )

@HMS_Phone so it would generate worse when it comes to normal generations?

I think you can solve it with negative embeddings, that usually helps a lot - like "easynegative" or other

Depending on the severity of the issue you're encountering, maybe you could model "merge" solo and let the model "forget" weights around hair(use as many related tags you can and merge it with a similar model that handles your hair the way you prefer? It's a pain, yes, but might give you a "relatively simple" fix? This has occasionally worked for me, (but maybe I've just been lucky rather than actually accomplishing anything with the technique)

Either way, hope you manage to generate what you're looking for!

@taiconan I kind of got a work around by using Wildcards manager and created a negative file with an absolute ton of hairstyles and hair colours... then its a case of hit and miss, but the odds are better. about 1 in 3 now have this "pompadour/Anime/manga" style hair, where as before it was every image.

I do not do paid commission this is a project off of love of tech not for profit or gain at the moment. I appreciate the offer

Please tell me how do you obtain a "stable and pretty face", without using a LORA (or other external resources.)
What kind of words do you use in the prompt?

I second this request!

@s6yx Yes please

真鸡儿恶心啊。。。

PLIS GOD ...

Due to having a new job, time does not allow. Doesnt mean it wont happen in the future

What's so good about model 1.0?

I'm also curious since it's just a merge. seems to be just a happy accident that came out soo well

@LostAtSea Yeap happy accident + some better merging recipe later on

If time allows, Life comes first

@s6yx Of course life comes first, I was just wondering. Thanks for taking the time to reply, have a nice day.

Interesting take on a free product or service. I have a life outside the Stable Diffusion sphere. I recently landed a well paying job that eats up 80% of my time.

does SD models made produce money for me? - no
Even if I sold this off, would it be enough to sustain me? - no

This comment does not make me want to go back into diving hrs on testing and training on my 4060 ti.

Good day

This dude's name is WaifuRegressor. That's all you need to know.

@s6yx I wanted to apologize, I didn't mean for it to sound so offhandedly. I genuinely think your work with this model has been an incredible contribution for the community and was just surprised how good it still is despite releasing so early in the Stable Diffusion timeline.

I completely understand work takes much higher priority over a free hobby. I've been on the other end before with my own fans making too many requests and kind of ruining what I originally found fun. I didn't consider I was basically doing the same thing to someone I thought did a cool thing too.

Sorry. I hope you find an inspiration again soon.

@s6yx dude do not care about ungrateful crybabies , enjoy your life and job and if you get a lot of free time consider monetizing your work

@s6yx damn bruh he was joking lol

@s6yx it sounded funny to me, why cry about such comments, i don't think that he was complaining about you, just talking facts, and those facts happened to be very epics. Like Saitama for example, he does the same thing.

That's going to be very difficult as this is a merge, not a trained checkpoint.

no time soon, I am not actively working in the sd scene anymore

try reducing width and when you get the girl you want, use inpaint
or you can use ((very close up face)) or (full body) for one

@Elaneor thank you, yeah I was doing a 16:9 resolution and noticed that I started getting 1 character when I did more square or narrow images. I'll try the prompts!

@ym1r this model works best with 3:2 and 5:4 as the description says. The first prompt I found using this model said just ... serious girl, long hair.. and was using 640 х 800

I am sad. Your model remains my favorite.

Also, Thank you for your good work. Your model is amazing!

It's a shame, in my opinion this is the best SD model, or at least my favorite.

This model is hands down my favourite one. Thank you for all your hard work on it. I am saddened by your departure, but appreciate all you have done. All the best to you!

Fair enough. It was a good run! People are moving on to SDXL at the moment, so seems like a good time to retire.

o7

Thank you for your amazing work! As you can tell, we are sad to see you go, and hope to maybe see some new work from you in the [not too] distant future.

As has been mentioned here, I'm sorry to see no more updates as my main model and was hoping for an SDXL one.

Good luck in you future endeavors.

Still hoping for a SDXL Update 😭😭😭

Will you be creating a completely new model?

@s6yx - Any plans on creating a new model/checkpoint? :)

@emmasteadman no

@Archivist no

@naughtyskynet unfortunately, I neither time or hardware to do so.  

@s6yx ok, thanks for the reply.
This is one of the best models so far, it's a pity for it being abandoned, but real life comes first, of course.

@s6yx Thank you for letting me know :) I hope you're doing good :)

This is the best model I've ever used. Thank you very much!!

RIP

@s6yx would be great to see the dataset and pipeline before seeing you go, hope all is well

s6yx answered this for someone already, they said "no time soon, I am not actively working in the sd scene anymore"

https://civitai.com/models/7371/rev-animated?commentId=218870&modal=commentThread

I was hoping Rev Animated could get an XL update, but in the meantime, you can check out Starlight XL as a substitute, which I find gives out similar outputs as Rev Animated.

the add-ons depends on your taste/preference. just download the latest model since it would be the most refined version out of them all.

Unfortunately the author has stepped away entirely. :(

@k21 Why did this message get 2 dislikes? There must be more to this story than what it seems to be imho

Read up on CC-BY-NC and CC-BY-NC-SA (since they lay the groundworks on collaborative non-commercial licenses), then get a lawyer if you can't think this through.

how can someone found you use rev anime for do what you do?

It's still my go go when I try to create landscape/cities or robots. Not sure why this model is so much better than all the others.

You can try to sell them if there is an interested buyer, but in terms of license I think that the AI picture has to be altered in a way so that you get the copyright for it. In general AI pictures can be used by anyone (so far as I know) and for everything.

OK thanks.

https://huggingface.co/spaces/CompVis/stable-diffusion-license

There's a link to this document in the licence for this model.
I'm not sure what applies, but this is probably a good starting point.

I would be aware of all models trained by individuals, as opposed to companies that vouch for the authenticity of the training data.

@almat993 As soon as you edit anything on the image and save it on photoshop for example as png or whatever, it removes any file residuals from it being generated.

It is the model that is protected by this terms, not the images you create.

The upscalers I normally use are:
RealESRGAN_x4plus_anime_6B

4x-UltraSharp

4x_fatal_Anime_500000_G

I think most of those are available on huggingface, but I don't remember for certain because I've been using them for so long. Those are the names of the files; hopefully a quick web search will help you find them <3