Ever noticed how hard was to train the 2.1 model and its derivates? It was because it had a heavily censored text encoder, which is the thing that makes pairs with images and roughly understands what's inside the pictures you provide to the model.
Well, there are better and uncensored text encoders compatible with the 2.0-2.1 models on huggingface, and once you get the model you want to train on in diffusers format, you can competely change the single parts of the model. Now, that's what I have done with this model and I will keep working on it.
I have trained the model on a large amount of women of any kind, using the waifu diffusion tagger, and by doing this the model captured the single features of each woman and mixed them together to get info for the celebrities that the encoder knows about but the unet is not trained on.
Currently the model has low capability on NSFW stuff, but I will try to make it work next, until then you can have some acceptable depictions of celebrities, I very well know these photos aren't perfect but it is a huge step forward on the basic model.
Ah, since it is not overfitting, it should be able to deal with other concepts, try yourself and, since it was smartly merged with Illuminutty you can try with their embeddings...
Description
First version, you might notice that the base model is 768x768 but it is supposed to work decently with 512x512 too
FAQ
Comments (16)
YOOOOOOO!!!!
WTF a dislike for what.
Gonna try this in a sec.
Not clear to me: if model 1.5 actually is uncensored and works for everibody (with 99% of LORA made for it), while model 2.1 was largely rejected by people... what advatages will give your model, once you will finish to uncensore and train it?
2.0 models have "better" prompting and usually understand better the things you ask them, don't underestimate the structure of that model.
Overall the output quality is usually better and the training is easier, it's just something "new", if people won't like it i really don't mind.
2.1 would have been an upgrade because base render resolution is 720x720. Due to it's censored nature it failed. If OP can manage to get it on the uncensored level of 1.5, it will open up a whole new world for Loras,TI and other checkpoints using this as base model. Seeing as SDXL is very resource heavy just to use the base model, it is excluding a lot of users due to hardware limitations on day to day use. So having a 2.1 option that is equivalent of 1.5 capabilities could be a good option for people with lower end GPUs.
Yes 1.5 is now the standard, but so was VHS for home video for a long time. When DVDs came out and was comparable and better quality than VHS, it eventually got replaced. We are in the same sort of space. 2.1 is better in quality overall if it wasnt for the extreme censorship. It's like if DVD format came out but was extremely censored, people would still carry on using VHS in that regard. SDXL is like BluRay in that regard but too many people dont have the hardware requirements to run it so having a better 2.1 would be a good alternative for eventual upgrade.
@spoffninja that's exactly my point, thank you for explaining it neatly! I really can not use SDXL on my computer, while I'm sure it's great, having a decent alternative with many updates in regards to 1.5 may be interesting for most people.
@Redigleb_Doppler2482 yeah Im in the same boat! XL looks fantastic but even the pruned version is a slog to load up once, then im on limited renders before vram memory limitation chop out. That's not even loading additional models like loras and controlnets. If GPU prices werent insane as it currently is, I would just upgrade my GPU. But no way do I have funds for the current prices a upgrade is sitting at. I assume tons of people here are in the same situation. So having something like a 2.1 uncensored base for more higher quality base renders would be awesome.
@spoffninja I'm also managing to train the model at 512*512, this makes it halfway between 768 and 512 with decent outputs for both resolutions (this particular version i just uploaded is mostly 768 though), the reason I'm doing this is, of course, for good usage even for the lowest end GPUs. I'm training the model in Colab because i have 4 gig VRAM, although i can generate renders at 768*768 it takes a lot with lowvram, and, well, it would be useless.
Y'all need to start upgrading to a card with at least 8GB of vram pronto before XL comes out. This effort here with 2.1 is commendable but will become irrelevant very soon.
@olternaut the cost of a 8 gig VRAM card is also in electricity, I understand that SD XL is better in everything but I'd rather spend some time for the people who can't afford for a better graphic card. Also, the evolution of all of this is way too fast and thinking that 2.1 is obsolete is not a good attitude to me.
@olternaut Yet.. like my example, DVD is still around and still popular for physical media. Because its more accessible to print stuff on that doesnt require a financial license fee. So to there are tons of people out there with lower spec cards that still will take part in the image generation revolution. Look at civitai, there are tons of SD1.5 models released every few minutes with the size of 1.99/4GB. Thats because those are super easy accesible. The 7/9gb models get posted much less so because they are mostly a walled off garden for some people, unless they also use google collan on top of their local installs. So no a lot of people wont be shoveling out a ridiculous price gauge from the card manufacturers. SDXL0.9/1 will be porsche or a lamb in the space. But wont be the norm till it's accessible. Until then, stuff like an improved 2.1 will be well received by the rest of the 10s or 100s thousand that don't have that gpu power
I'm speaking as someone who has a card bigger than 8gb
I'm not uploading an early access model anymore, I can't even see the comments on my ckpts. Well, anyway, in three hours it will be for anyone to use (keep in mind that this model has EXACTLY the same hash as the other one, the one not SFW... just saying)
To bad about the early access troubles, but the work you're doing is very much appreciated. Thank you.
Hello everybody looks like I managed to "fully uncensor" the model.
Despite I barely put any man in the training dataset, the text encoder seem to have caught most of the facial features from women celebrities and managed to mash them together to allow decent male-celebs renderings, tomorrow I will upload the best model yet.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.