Welcome To The Hentai Diffusion GitHub
Hentai Diffusion has been made to focus not only on hentai, but better hands and better obscure poses.
Dispite the name there is no adult content on this page.
HD 17 now keeps a more consistent style.
We've improved animation abilities. Now subsequent images will be more coherent.
We've been working on improving style consistency. In the upcoming release of HD-17 characters will look more cohesive. This can be easily changed by using any embedding or hypernetwork. HD-17 also allows you to change the style of images a bit more easily, though this may be placebo.
Effort has been put into getting HD working better with img2img using real photographs as references.
Recently I've released a few images into the wild on different reddit accounts and pages. While most people seem indifferent or are unable to notice any differences, a small group of users have been incredibly hostile. No posts were ever removed and it seems the more vocally negative users are an extreme minority, and most users are happy to get more art. The reasons for hostility range from inferiority or jealousy to something more of a perceived justice of "You didn't do that". Despite the vocality of some users I do feel it comes from a much deeper place than just not liking the produced art.
Seeing as posts of HD's art have reached the top of a few pages and the majority of users enjoy the content I feel this experiment was a success, but I do warn against doing this outside of collecting data. Since no harm was done, nothing was stolen, and overall it was a net gain to the communities as they gained some new art of their favorite characters, I am happy with leaving the art up. If you plan to post art generated by this A.I I recommend labeling it as such.
Below is one of the most successful examples.
Below is a graph to help guide you on which CFG scale works the best for the style you need.
Euler ancestral provides the best results and doesn't need to be changed.
This model was built with the universal negative prompt in mind. The text file can be found on Huggingface as well - copy and paste it or save the text file and enter it as the negative prompt. I suggest you save as a style in the webui.
Windows Defender is the only anti-virus that has chosen to flag some checkpoint files as false positives. I know it's shocking, the king of false positives gave a false positive, but to the few people concerned, no it's not a virus and I recommend using MalwareByes or anything other than Windows Defender.
Current Version: HD-17
Negative Prompt: Download
Old Version (Archived): HD-16
Copy and paste the negative prompt text into Neg Prompt area then save as a style.
smmrr.mp4
You'll need a way to run the ckpt file. We recommend using AUTOMATIC1111's UI
Once you have the stable_diffusion_webui zip file, unzip it into the directory you want to use for Hentai Diffusion.
Open the folder containing the files you just extracted and go to this folder.
stable-diffusion-webui-master\models\Stable-diffusion
Place the .ckpt file you downloaded into the folder.
Return to the main directory.
stable-diffusion-webui-master\
On windows run the webui-user.sh file. For linux use webui.sh.
Let it run and download all the needed requirements. It may seem frozen at times but it is doing fine, just don't close the terminal.
Open your browser to the local address shown in the terminal. It should be something like 127.0.0.1.
Make sure to select Hentai Diffusion (or whatever the correct ckpt file name is) in the dropdown menu here.
Copy the negative prompt you downloaded and paste it into the negative prompt section.
Make sure the prompt field above is empty.
Click the save icon and name it to whatever you want. Something like "Master Negative" is a good choice.
Whenever you start up the webui, select style one, and then the name of the master negative. You then click on the clipboard icon and it will add it to the field.
This is only needed once per session as long as the browser isn't closed.
Add "1girl, anime," without the quotes to the beginning of your prompt.
- Merged model with WF using the Inverse Sigmoid interpolation method with a 0.25 merger value.
- More Characters. Current supported characters include most characters with over 5,000 results on R34 and Gelbooru. Suggestions welcome on the Issues tab on this page.