Silly Tavern, a project on github that is about 2 months old, attracted my attention this week.
Silly Tavern is basically a fork of TavernAI, an interface for LLM that is specialized in chats or role play. To make it short, you can talk, play with them, write novels, make them impersonate characters...
The advantage of Silly Tavern on TavernAI is that it can handle extra functions like Image Captioning, long-term memory by summarization (to keep the context better for a similar use of the memory), D&D dice, a link with Stable Diffusion and a Text-to-Speech feature.
To use it, you will need a LLM interface like the Oobabooga webui, an LLM to use with it, the installed version of Silly tavern and probably also to have the extensions installed too.
***
Drag Your GAN is another promising project that as the capability to reduce significantly the cost of media production. Basically it seems to allow to produce a transition from a picture to another state of a picture using what i would personally call "vectors of transition". If you click on one of the links above, you will be redirected to a page that features several videos. Sadly, and i tried with 4 different browsers, most of the time, i wasn't able to get all the videos loaded correctly (Chrome works likely the best). The page itself returns an error from time to time (press enter in the address bar if it happens, "reload" doesn't seem to work).
So here is how to watch the videos:
For whatever browser (same concept) and with VLC:
- Right click on the video that is not loaded:
- Select whatever line that allows you to copy the link of the video (here: "copy video address", with Edge it is "copy link").
- Load the address in VLC. If it returns an error, wait a few seconds and try again
It seems like a source code is expected to be published in june: https://github.com/XingangPan/DragGAN
So personally, i am not finished to exploit these videos because if you look at them, you can see that some files are being used:
"Pickle: stylegan2-afhgcat-512x512.pkl" for the cat video.
"Pickle: stylegan2_horses_256_pytorch.pkl" for the horse video.
I am not an expert if AI, but i can assess that these files are close to those that you find in these projects from Nvidia:
https://github.com/NVlabs/stylegan2
https://github.com/NVlabs/stylegan2-ada-pytorch
(The directory includes some source codes)
A youtube video (without sound) explains what this stylegan2 technic does:
https://www.youtube.com/watch?v=c-NJtV9Jvp0
The video includes an idea to detect fake pictures.
***
PrivateGPT
So, while I am not finished to test what i discovered, it is Aitrepreneur that released a video about a fascinating new piece of software: privateGPT. While the youtube video claims that you can chat with a document, this is not exactly right. PrivateGPT is some kind of smart search engine for your documents that you can run locally on your computer and it is already interesting enough for this sole reason.