I made this mix with the intentions of making a photo-realistic model with emphasis on portraits and realistic imperfections.
Models were created using the checkpoint merger tool in automatic1111 without any extra training data/images.
A lot of NSFW models were mixed into this model so it is possible to create NSFW images but results may not be as focused as other NSFW models.
Most of my models use a version of modelshoot or has 1 or more models mixed in that used modelshoot, so keywords like modelshoot style and from_below should work.
I also use these negative embeddings:
https://huggingface.co/datasets/Nerfgun3/bad_prompt
https://huggingface.co/nick-x-hacker/bad-artist
Description
This version has IDENTICAL results to v2.0, I simply reset the CLIP using https://github.com/bbc-mc/sdweb-merge-block-weighted-gui. v2.0's CLIP tensor was a float (byproduct of merging with some models) when it was supposed to be integers. The values of the tensor were the same just the tensor type was wrong. Since none of the tensor values changed the resulting image outputs have not changed.
FAQ
Comments (7)
safetensors or ckpt? me used ckpt
i mainly only use safetensors unless ive downloaded a ckpt model from somewhere and it was the only option (then i usually convert them to safetensor before i use them). safetensors are the "safe" way to load models. ckpts can load other scripts embedded in the model and the fear is that they can be malicious scripts.
Wow, your model looks really cool!
Alas copied the seed, the CFG setting, the steps, both prompts, but I can't replicate the torn clothing image with 2.0 with VAE. There are some similarities, but I can't damage clothing at all with any model! Is there something I'm doing wrong?
Perhaps it's the service you are using. I generate all my images locally using automatic1111. My prompts use the automatic1111 syntax for emphasis, "()" and "(word:1.3)". Other services use different syntax to do the same thing, like "++word".
@s4w3d0ff Hmm, I am also using Automatic1111
I'm also using xformers 99% of the time. I've found that putting emphasis on words like "torn, tattered, worn out, frayed" work and words like "ripped, shredded" work but can be occasionally confused with "muscular". I'm not sure what the difference is between our systems/settings.
try eta noise seed delta 0 and eta noise multiplier 1
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.