MelodicMisfit's Profile

Just someone who likes ML and generative AI (botmakie too). Check out my CivitAI if you want, sometimes I post things.

Thanks for checking out my profile! I'm not sure how often I'll make more bots, but when I do, I'll post them here.

Card Artwork

• All of the artwork in my cards is made by me running Stable Diffusion Forge WebUI locally, with a lot... lot... lot of inpainting and manual touch-ups. The models I most use are Pony Diffusion V6 XL and AutismMix SDXL (also based on Pony Diffusion) as I believe these are the best anime models with an unparalleled comprehension of prompts and artist tags. Also wow, it's kinda insane how the best anime model since the NovelAI leak is a furry model.

Models I Use

• I run all of my cards locally on an a 3090 with SillyTavern as a frontend and llama.cpp/oobabooga as a backend.
• I recommend an iQ2 quant of lvlz_70B for 24GB VRAM. Although it is trained at 4k context, you can squeeze 8k context in with some RoPE scaling while running at a comfy ~19 token/s. This model lacks the 'GPT-isms' that a lot of others seem to have, and is really good at writing and staying in character - even at its low bpw.
• ~~Other good models for RP in order of decreasing VRAM requirement are:~~ ~~1. Nous Capybara 34B~~ ~~2. Noromaid-20B~~ ~~3. Noromaid-13B~~ ~~4. Kunoichi-7B~~
  • Just use Llama 3 8b or an l3 finetune lol