Don't forget the negative embeddings or your images won't match the examples
The negative embeddings go in your embeddings folder inside your stable diffusion webui folder
(if you have a different install I'm not sure but someone will know online)
To use the negative embeddings properly they need to be added to your negative Prompt only
Also this is trained on 768 resolution images so it won't do well with 512 resolutions, different ratios do work but are hit or miss.
4 months, 700,000 images, and a very high electric bill and the model is finally in the "Good Enough" stage.
My goal of this model was for it to generate highly detailed and realistic images without the need of those tags in the prompt and it works most of the time, but not every time so that's were the negative embeddings came in making the model about 50% better in most cases.
There is another version of this model the pushes the realistic style even harder, but loses some coherency for it as they're just not enough differing images of that type that at least I could find, so I haven't released it. If you guys want it I will release it in it's current state but I'll probably never get to finishing it as it is now Summer time and I'm tired of it being 105 degrees in my room all day long, so this might be my last model at least until SD LX comes out. Might do some Loras if people want them, but for this model there not really need.
Description
Don't forget the negative embeddings or your images won't match the examples
The negative embeddings go in your embeddings folder inside your stable diffusion webui folder
(if you have a different install I'm not sure but someone will know online)
To use the negative embeddings properly they need to be added to your negative Prompt only
Also this is trained on 768 resolution images so it won't do well with 512 resolutions, different ratios do work but are hit or miss.
4 months, 700,000 images, and a very high electric bill and the model is finally in the "Good Enough" stage.
My goal of this model was for it to generate highly detailed and realistic images without the need of those tags in the prompt and it works most of the time, but not every time so that's were the negative embeddings came in making the model about 50% better in most cases.
There is another version of this model the pushes the realistic style even harder, but loses some coherency for it as they're just not enough differing images of that type that at least I could find, so I haven't released it. If you guys want it I will release it in it's current state but I'll probably never get to finishing it as it is now Summer time and I'm tired of it being 105 degrees in my room all day long, so this might be my last model at least until SD LX comes out. Might do some Loras if people want them, but for this model there not really need.
FAQ
Comments (50)
This new update is something else, wow! Do you have any socials so I can follow you?
I don't really, I'm not much of a social person. posting on here is about a social as I get :)
Would love this! But where's the actual V2 checpoint? Was it removed again?
I only see a 162Kb *.zip file called "training data"
It's there on my end? try logging out than back in
I got it via your link in this comment thread, thanks :)
It's still not working on the page though, did logout and back in. The other versions still work
Maybe I'm a newb but wanted to post this for visibility. Wasn't to clear from the OP.
To anyone else, download the negative embeddings in the description, put them in embeddings folder.
Example: C:\Program Files\Git\stable-diffusion-webui\embeddings
There's four safetensor/checkpoints to dl and put there. To use each of those four, make your negative description include:
bwu dfc ubbp updn,
YOU MUST DO THIS, and should be able to recreate the sample OP images or something similar following their descriptions/seeds if you did it right.
You may get interesting or funky/broken results if you add ((emphasis)) to these or end up using them as a lora in the lora folder by accident like I did at first.
yeah they make a big difference, they are pretty strong strength so ((emphasis)) might be too much if anything using [[less emphasis]] might help in some cases and no I wouldn't say your a newb I should have made it clearer what negative embeddings and how to use they, I forget not everyone watches every single SD news video like I do ฅ^•ﻌ•^ฅ
do i have to convert them to a .ckpt file like i have to do for the main model?
@infotron you shouldn't have to, but I don't know what your using to run SD.
@sopenit494 im using stable diffusion webui, idk what safetensors are but they dont work. i have to convert all models to a ckpt file for SD to recognize it
@infotron That's weird, are you using a very old version of the webui? Safetensors have been supported for a verry long time. There basically a model format that doesn't allow any extra or malicious code to be added to the model file..
@sopenit494 how do i know what version im on? I guess ill just get the latest version from github anyway
@sopenit494 WOAH THERES SO MUCH NEW STUFF
i checked when the files were created and it said like mid 2022 btw
The V2 model was trained with Clip Skip 1, right?
yep
Hate to be a bother and love the model itself as I'm making good use of it. But I'm more curious about the Negative Imbeds, they seem to work for pretty much anything in a very good way (great push towards realism and detail).
Could you give a definition as to what each one does or do they have to be used together?
sure, bwu- stands for blurry, watermark, unrealistic, dfc- stands for dull, flat, color, ubbp- stands for unbelievably bad body parts, updn- stands for ultra-saturated, painting, drawing, "n" I can't remember made this 1 first :|
Thanks, appreciate the information. :)
What doe each of the negative embeddings do?
bwu
dfc
ubbp
updn
bwu- stands for blurry, watermark, unrealistic, dfc- stands for dull, flat, color, ubbp- stands for unbelievably bad body parts, updn- stands for ultra-saturated, painting, drawing, "n" I can't remember made this 1 first :|
Is it possible to have furry but more in the style of ‘monster girls’? ie basically human or mixed morph faces? Ty, awesomelora tho !
probably, with a lora, or mix the model with a more human focused model, and/or go to e621 and find the type of image your looking for and try use some of those tags
the best model on s.b so far! I've tested like 8 models and this is perfect!
SDXL came out a couple days ago. Any possibility of seeing a tune of SDXL some time soon?
not for awhile, SDXL takes much longer to train than 1.5 and need to do a lot of testing to really dial down what works best training wise.
@sopenit494 oof... Let me know. I'm looking forward to needing fewer inpaints and using descriptive language!
@sopenit494 hope you're still looking into an sdxl version, i like how this detailer does fur.
Downloaded the training images zip, but the archive is tiny and doesn't seem to have any images in it. Was assuming I'd see the 700,000 images. Was an incorrect file uploaded by mistake or am I just misunderstanding what I should be seeing?
Those are the negative embeddings and for some reason this site labels zip files as training images, I would love to upload the dataset I used, but it's over 600gb and would take me a month to upload
@sopenit494 Makes sense, thanks a lot!
@sopenit494 would you be willing to do a sneakernet distro? I'd buy or ship a drive over if it would come back with the annotated dataset.
Maybe you could spin torrents and I could seed them with those hash-matched files. Or several of us could seed. This is a solvable problem.
Annotated files can as easily be used to train an annotation system!
Please ignore/report anyone posting shady links!
Do a merge with AlphaTango
Just fantastic!
I've read about the BWU, but does anyone know exactly what words are are hidden in this negative embedding?
This was mentioned somewhere (copy&pasted):
Credit to ShisoFox for going through the effort and asking the creator what they meant.
bwu - blurry, watermark, unrealistic
dfc - dull, flat, color
ubbp - unbelievably bad body parts
updn - ultra-saturated, painting, drawing, not realistic
dont know how muh you update the checkpoint but im starting to get worst results now, if you still have an older version i could use that would be lovely
haven't touch it in 9 months, maybe whatever ui your using might have updated something to change the results 🤷♂️
good model but need a gooood big update
awesome models, but the tags list link is missing
Yeah those links were only good for a year, I would update them but I don't have the text files anymore 😥
@sopenit494thank you for your reply
I restored it trough archived copy. This model is legend. https://paste.ee/p/cQeCz
@mrtpatata800 How do those work? They're just numbers
@sneedo this is model learning reference - sorted how many images, used for learning, has specific tag. This means: if tag exists, model definitely can understand it. The less number - the weaker tag strength. This means, if you want low frequent tag feature persist on every generated image, you need to strengthen it by multiplicative factor, as well strong ones can be lowered for more creative generation.
This model has very rich variety of it, even half of it barely obvious, that exist in model without this reference file.
IMHO every model should have such file, in terms of understand what its capable of, and which features need fine tuning for needed prompt generation result.
@mrtpatata800 Hey, paste.ee is down. Do you still have these tags and can you share them another way?
@Mcracken, https://drive.google.com/file/d/12HANFcMH_CcB2nMDBVCeuglx9BMEpzKX/view
hope this wont die. I'm impressed, this model is still popular. Pony models is way better at present time, but this one still best at flexibility. I think good workflow in ComfyUI can push it to limits with upscalers and other improving stuff.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.