Home Creators Posts Import Register Favorites Logout
haven't archived this post yet. have a subscription? use the importer!

Content

Lots of AI stuff going on these days and I know many of you are interested so here's a few things in case they help!


STABLE DIFFUSION

There has been a new release of HassanBlend (1.5.2), to my knowledge the best AI model for StableDiffusion around so far. It can be downloaded from https://huggingface.co/hassanblend/HassanBlend1.5.1.2/tree/main, either HassanBlend1.5.1.2-pruned.ckpt or HassanBlend1.5.1.2.ckpt will do. They should be the same, pruned just means that it was cleaned up of some redundant stuff (to my knowledge). The dude training the AI model posts more about it and has other stuff on his patreon: https://www.patreon.com/sd_hassan if you want to check it out.


STABLE DIFFUSION UI

If you've been using this UI (https://github.com/cmdr2/stable-diffusion-ui) you've probably noticed that it gets lots of automatic updates and people add new features to it often. It very recently got renamed also to Easy Diffusion. If you encounter any issues with generating images after updates, you might have to hit ctrl+f5 to fully refresh the browser page cache. I had that happening to me a few times now.


ADULT STORIES AI

I looked a bit into what's going on with free open-source local text AI generators and if there's anything easy & good to play with, particularily to integrate it in VAM. The good news is that there is something very cool available, the bad news it's too slow still to be practical for VAM purposes.

Kobold AI (https://github.com/KoboldAI/KoboldAI-Client , https://www.reddit.com/r/KoboldAI/) is what I tried and it's pretty amazing. It's simple to install, and it works the same as stable diffusion UI, it's a web browser UI.


KOBOLD AI SPEEDRUN TUTORIAL

For people that want to try it, here's a speed run:

1. Install https://github.com/KoboldAI/KoboldAI-Client#installing-koboldai-offline-bundle-on-windows-7-or-higher-using-the-koboldai-offline-installer-easiest
2.  in the UI you can click AI and there you'll find the available AI models. They get downloaded the first time you pick them

I tried NSFW > Erebus 8gb which seemed pretty good. The bigger the model, the better they are but they might be too slow or return errors depending on the GPU. Some of those are meant for special setups, won't work on consumer PCs.

3. When picking a model you'll also have to select how something called layers. Think of them as AI brain points and it's sort of like the game mechanic where you set attribute points on characters. In this image it asked me to set 32 layers and choose between GPU & Disk. The more you put on the GPU, the faster the responses will be. But you might have to put some point on the disk as well if you don't have a great GPU card and you get GPU errors. In the image I had 32 layers and set 26 on GPU and 6 on the Disk cache for example, and hit Load.

3. If you want the AI to be in chatbot mode, you can go to Settings and check Chat Mode and the AI will try to answer to you in dialogue
4. To begin you might want to give the AI some info by clicking on Memory and setting some Author Notes (genre, setting description, etc)

5. Type an introduction to describe the setting, the situation, the characters, and to get the AI going

6. The AI will then keep generating text. It's a bit slow depending on your settings. Turning down in settings "Amount to Generate" and "Max Tokens" helps with speed at the cost of intelligence.

I had chat mode enabled so the AI assumed the role of Julia and tries to continue the story as replies. With chat mode disabled, the AI tries to write prose.

I had it set to the minimum Amount to Generate so it generated this, a partial response

And I hit Submit again to let the AI to continue its thought and then I started typing my reply:

The AI will keep going and continue the story as you help it along. Depending on the settings it's going to be narrow-sided and goofy sometimes but it can be used to write erotic literature or just for fun:


It's not as cool as ChatGPT, this one is based on an slightly older GPT AI, and the AI models are considerably smaller & less smart. But since the online AI text generators are going on a crusade against erotic content, this seems to be the best solution for stuff like that. There's other AI models included but I haven't tried them yet. For more general purposes I'd recommend:

CHATGPT

It's everywhere in the news these days and you probably know about it already. If you don't, you should make an account right now (it's free) and try it out, it will blow your mind. It works with other languages too. Pretend it's the smartest thing ever and ask it anything. Family-friendly stuff though as it will trow fits at anything remotely controversial.

Here's ChatGPT continuing the story from the local AI above. It's in a completely different league.

Or, another example:


I also looked at integrating ChatGPT with VAM and I think it's technically possible and pretty straight-forward. The downsides are that it's not a free service and it won't work with anything remotely erotic for VAM. IBut the benefits are that it's insanely smart. I don't know if I'll pursue it myself yet but it's in the cards.

Files

Previews only

Comments

Anonymous

I've been tinkering with TavernAI frontend with the new Pygmalion model. Made due to Character.AI which was fucking amazing for chats and worked really well, yet neutered by corporate and lobotomized to be SFW... Pygmalion though is designed ground up for chatting with it's model and NSFW, so might want to check it out! You run it remotely on Colab or Kaggle. Locally though 6B parameters requiring 16GB of VRAM... I really think VAM would be fucking awesome to have chatbot capabilities with something like how TavernAI does it. People could possibly import personas with portraits of their characters or from the Pygmalion Discord which has lots already then add them to your VAM characters. I think it utilizes the saved meta in the image to store the W++ information for each character and their persona. Then sending it to the Pygmalion backend. You can check out the GitHub to see how it works. I'm no programmer but that would be fucking cool, probably boost creators to making more famous characters and grabbing a chunk of that base to flock to VAM and your Patreon. Least forgetting Replika users who want the VR experience! Fascinating thing is sometimes I notice text for NovelAI Krake on TavernAI outputs current personality/behavior tags from the conversation being had. It would be cool to have Alive detect those tags and adjust behavior while chatting for immersion. I wonder if Pygmalion output does the same? Don't really know, but the AI is capable for such on NovelAI. Here's links to the stuff. https://huggingface.co/PygmalionAI/pygmalion-6b https://rentry.org/pygmalion-ai https://discord.gg/pygmalionai https://github.com/TavernAI/TavernAI

SPQRAeternum

Thank you! yup, I actually started by looking into pygmalion and that's how I ended up installing KoboldAI for the UI. I think I had some trouble with it, likely because of my gpu, and stuck with a simpler AI model for my tests. For the local AI stuff, it's not as simple as just looking at the github code to understand how it works. Not for me anyway, I don't know python and have no idea how long it would take for me to figure things out. There's also the problem of being too messy, myself I don't like all these extra software and lots of pieces in constant change that you have to glue together, and I think many people feel the same and wouldn't bother with it. And there's also the problem of running it. There's no pygmalion backend, that's just a big file, an AI model that you have to run somewhere and taking all th GPU there. Most people can't run that just by itself, not to mention adding vam on top which is heavy on the resources. To host it myself on the cloud like the discords bots do seems too much hassle for me. I'd get people DDoSing my server to troll me, people complaining about privacy, people complaining about speed or not working for them and be mad at my service as if i'm some startup etc etc. But it will definitely get there soon where it's something clean to implement and I'm looking into it. Right now I think it's just way too messy to be something cool for everyone