Conversation

Notices

1. katkat@shitpost.poridge.club's status on Monday, 01-May-2023 06:39:47 JST katkat

   • ✡️ ((fuggy)) :swastika:

   @fuggy@skippers-bin.com you're the ai art guy right? I haven't played with it in a long while what's the best setup now?

   • ✡️ ((fuggy)) :swastika: (fuggy@skippers-bin.com)'s status on Monday, 01-May-2023 06:39:46 JST ✡️ ((fuggy)) :swastika:

     in reply to

     @katkat@shitpost.poridge.club I am mostly playing with chat bots these days, but I can show you what models are good these days

   • ✡️ ((fuggy)) :swastika: (fuggy@skippers-bin.com)'s status on Monday, 01-May-2023 15:08:40 JST ✡️ ((fuggy)) :swastika:

     in reply to

     @katkat@shitpost.poridge.club Yes, but many advancements have come out
     
     1. Lora - small files that store fine tuning information that can be easily used with any model and are high quality and easy to make
     
     2. Control net - let's you pose characters etc.
     
     3. Explosion of custom models for anything you would want on https://civitai.com/
     
     4. More optimazations allowing for lower and lower VRAM, most recently a PyTorch 2.0 fork https://github.com/vladmandic/automatic of https://github.com/AUTOMATIC1111/stable-diffusion-webui
     
     5. New upscale methods that sometimes yield better results
     
     6. Early text to video models are released, they aren't great so far but it's a start
     
     7. DeepFlyod is released a text to image model that can generate text well, however with a restrictive licence and harsh hardware requirements
     
     8. MidJourney v4 is released with great results
     
     Anyway, as for chat bots:
     
     1. ChatGPT is released a chat bot
     
     2. Facebooks LLaMA LLM leaked which is currently the best model you can run locally(realistically)
     
     3. Fine tunes of LLaMA to act be closer to ChatGPT released such as Alpaca, Vicuna, GPT4ALL, WizardLM, Stable Vicuna
     
     4. Pygmalion released which is a alternative for https://c.ai has been released, this is due to character AI censoring NSFW content
     
     5. A version of Pygmalion AI based on LLaMA was released which is generally better than old GPT-J based version due to better model architecture https://huggingface.co/PygmalionAI/pygmalion-7b
     
     6. From the same people that made Pygmalion AI, Metharme was released. It seems to be a ChatGPT style model fine tune, but with roleplaying data mixed in? I guess you can have anime girl as your assistant now? Local models aren't good enough to be trusted with facts yet though, so it's rather useless. But in theory it's a ChatGPT style model that you can rp with and ask questions and get answers to best of it's ability
     
     7. Open Assistant starts collecting data for chat bot based on LLaMA trained on human data
     
     8. Some Open Assistant models released to public on website, and other models not based on LLaMA released publicly for download. Due to legal issues, LLaMA ones are only on the website for now at least(likely to change)
     
     9. GPT-4 is released and can see images, but that functionality is not available yet
     
     10. ChatGPT 4 is released but requires a subscription but is much more intelligent
     
     11. Bing releases a GPT-4 powered chat bot that can search the web
     
     12. Google does the same, but it's not great compared to Bing surprisingly

   • katkat@shitpost.poridge.club's status on Monday, 01-May-2023 15:08:41 JST katkat

     in reply to

     • ✡️ ((fuggy)) :swastika:

     @fuggy@skippers-bin.com hm, sure. Are we still using SD?
     
     you can tell me about chat bots too tho

   • ✡️ ((fuggy)) :swastika: (fuggy@skippers-bin.com)'s status on Monday, 01-May-2023 15:10:21 JST ✡️ ((fuggy)) :swastika:

     in reply to

     @katkat@shitpost.poridge.club likely am missing a lot as it's quite overwhelming

   • ✡️ ((fuggy)) :swastika: (fuggy@skippers-bin.com)'s status on Tuesday, 02-May-2023 01:43:35 JST ✡️ ((fuggy)) :swastika:

     in reply to

     @katkat@shitpost.poridge.club idk lol

   • katkat@shitpost.poridge.club's status on Tuesday, 02-May-2023 01:43:36 JST katkat

     in reply to

     • ✡️ ((fuggy)) :swastika:

     @fuggy@skippers-bin.com do the latest optimizations and LLaMA work on navi2/AMD? I remember pt2.0 releasing as nvidia/cpu only.