Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
XENIA11.png

If you know, you know.
 
RunPod and Google Colab among many others let you rent hardware and run your own models
do you have a guide for retards i can follow
i used colab in the past for koboldai but i just clicked a button and waited for the python script to finish
 
I like Podrun.io myself. An RTX4090 can be had for less than a dollar an hour.

If you are an absolute mad lad, you could try running a Tesla P40 with a modded fan shroud. $220 and you get 24GB of VRAM. It only runs fp32 though, so no training only image generation.

You would need p100s for that.

Edit: My dream setup for ML if I had 2k to spend would be an X99 chipset motherboard with a xeon with 40PCIE lanes. Loaded with 4 P40s and an NVME drive. Sure my basement would sound like a jet hanger but I would have 96GB of VRAM, and I could do anything I want.
 
Last edited:
During AMD's Computex livestream, Stability AI's CTO Christian Laforte announced that Stable Diffusion 3 ("medium" model) will be available for download on June 12.

View attachment 6049101
Well this is good news. At least we'll be able to get answers to a lot of what we are curious about. "Medium" means the 2B model I believe. It's a shame they're not releasing the 8B yet, I should be able to run that and would very much like to try it locally. (I'm guessing the API version is the 8B but I've not seen it stated explicitly, I think).

Whilst my interest isn't NSFW I am curious about the degree to which they've managed to censor it and prevent people retraining it to produce nudity. The ability to build censorship into models is an interesting and dangerous area. Getting the 2B model out there will help answer some of that.
 
Well this is good news. At least we'll be able to get answers to a lot of what we are curious about. "Medium" means the 2B model I believe. It's a shame they're not releasing the 8B yet, I should be able to run that and would very much like to try it locally. (I'm guessing the API version is the 8B but I've not seen it stated explicitly, I think).

Was SDXL the previous largest model at 3.5B? 2B isn't so bad, and optimization can make smaller models more capable than previous bigger ones. They've also changed the backbone for this one. Though it would be nice to have the 8B one available as well.

Whilst my interest isn't NSFW I am curious about the degree to which they've managed to censor it and prevent people retraining it to produce nudity. The ability to build censorship into models is an interesting and dangerous area. Getting the 2B model out there will help answer some of that.
Is it even possible to prevent nudity in retrained models/LoRAs by sufficiently mucking up the base model? I think all of Stability's efforts are focused on making sure the default products they offer aren't naughty and that they have plausible deniability in the name of "muh safety". I don't think anything they do is actually going to stop or slow down the coomers, unless they stop releasing models. I could be wrong.


Click on filters, and "SD 3". The category is empty, waiting to be filled. With cum.
 

Was SDXL the previous largest model at 3.5B? 2B isn't so bad, and optimization can make smaller models more capable than previous bigger ones. They've also changed the backbone for this one. Though it would be nice to have the 8B one available as well.
Yes. SDXL at 3.5B which was a significant jump. So this has backslid. I know that the number of parameters isn't the be-all and end-all by any means. But still feels wrong. I want 8B. I've got 20GB VRAM and I think that I can run it. Even if not I can use runpod.io to run it cheaply enough. And I want to see what others can do with it, too!

Is it even possible to prevent nudity in retrained models/LoRAs by sufficiently mucking up the base model?
I don't know - that's what I'm curious about. I can think of a few techniques you could do that might hinder retraining to add it. I'm fascinated to know what is possible in this area of censorship. It's one of the biggest issues in AI right now. Though I'm more interested in it on the LLM side than images. I don't want AI being politically driven.

Click on filters, and "SD 3". The category is empty, waiting to be filled. With cum.
LOL!
 
Is it even possible to prevent nudity in retrained models/LoRAs by sufficiently mucking up the base model? I think all of Stability's efforts are focused on making sure the default products they offer aren't naughty and that they have plausible deniability in the name of "muh safety". I don't think anything they do is actually going to stop or slow down the coomers, unless they stop releasing models. I could be wrong.
They could poison the model via false tagging on tags only coomers would use. So, for example it ingests a photograph of a baby duck and then "dog penis" and "cum" get inserted into the tags before it trains on it. And so on for every single image to the point where it can't figure out what the coomer tags are supposed to mean anymore. You wouldn't be able to do a LoRA since the base weights are so fucked up.

I doubt they will do that though.
 
They could poison the model via false tagging on tags only coomers would use. So, for example it ingests a photograph of a baby duck and then "dog penis" and "cum" get inserted into the tags before it trains on it. And so on for every single image to the point where it can't figure out what the coomer tags are supposed to mean anymore. You wouldn't be able to do a LoRA since the base weights are so fucked up.

I doubt they will do that though.
You could still get around that with trigger words. For example, if you were training a custom Lora with Barney and Barney was blacklisted, you would just pick out a diverse sample of images of Barney and train the trigger world mixlgf. Then replace the weights of Barney with mixlgf in a checkpoint. You would probably want to do this a token at a time though.

Edit:This was actually the last and crucial step which prevented me from creating my own Lora. It was only when I eliminated all common element keywords from tagging that my Lora became complete and accurate. Looked great before I deleted it and my stable diffusion install.
 
Last edited:
  • Informative
Reactions: std::string
You could still get around that with trigger words. For example, if you were training a custom Lora with Barney and Barney was blacklisted, you would just pick out a diverse sample of images of Barney and train the trigger world mixlgf. Then replace the weights of Barney with mixlgf in a checkpoint. You would probably want to do this a token at a time though.
You'd have to do this for every poisoned token.

I mean I guess you could compare old model vs new model using cosine similarity to see where the poisoned tokens are and retrain everything but that sounds like too much trouble for the average user.
 
  • Like
Reactions: macrodegenerate
Are there any good gallery extensions for Automatic1111? I'm moving back to it from SDNext thanks to several extensions not working well on it, but I really miss the gallery tab from SDNext.
 
Back