Stable Diffusion 2.1 (base) is also supported by default in Doohickey (and diffusers in general!)
set the model name to "stabilityai/stable-diffusion-2-1-base" and go ham https://t.co/iKHHzpPG0s
the prompts extended by my RAT (https://t.co/PvkSlrDL48) do translate okay to the v768 model ! automatic prompt engineering !!!
(also using @minimaxir 's <wrong> token in the negative with "in the style of <wrong>, amateur, poorly drawn, ugly, flat, boring")
colab demo below
I'm fine-tuning a prompt-extending GPT with RLHF for SD2.0 (by myself, I don't have a userbase to get ratings from)
Custom personal prompt engineer!
(a ton of the code was generated by ChatGPT too! here's some intermediate unconditionally generated prompts)
[BREAKING NEWS] New diffusion model coming to Doohickey/HuggingFace soon™ (probably sooner than "soon™" you've heard elsewhere) https://t.co/glZ9tFAY4s
working on a model that does higher resolution better-ish, it's not perfect and not done training
(from lexica vs from this model, both 1024x1024)
w/ text inversion for dalle paint style that doesn't work 100% but does add flavor
removing the last layer of CLIP for waifu diffusion like in NAI, I'm thinking of fine-tuning wd on the second to last layer for like just a few/10s thousand images
(removed vs not, "miku, masterpiece, best quality, miku")
so i dont have to prompt engineer for profile pictures, an anime style icon generator using stable diffusion
"1girl red hair, twintails, neutral expression, angry, short eyebrows"
https://t.co/FPppwQfeCt
i didn't realize the settings i was using were weird without perlin noise, without / with
raw ViT-H-14 guided Stable Diffusion at [640, 512] (just pushed this update properly!) https://t.co/d0q3nMptm2