on my local machine running the hlky stable diffusion build

!dream hyper-realistic, 8k, uncanny valley, Peter Griffin as a real human, 3d render, Family Guy, Seth Macfarlane, fat guy, Rhode island, middle age, glasses, white shirt, green pants, overweight, new england

hey so it doesnt work like midjourney with weights for individual parts of the prompt. it works more like dall e 2. a basic way of increasing weight of words is repeating the word.

also in general i think a pretty sure way to make sure faces are in the image w any model is mentioning a specific facial expression/action like smiling etc.

it’s really not tryna render her face for you

!dream drake slapping T.I. In a movie theater

the hlky & lstein forks supports weighting wasn’t sure if you had that installed was doing a test

and yeah good idea i’ll try it like that 🤔

!dream beautiful face Jennifer Lawrence as a jedi knight, yellow lightsabers, face, highly detailed, artstation, full figure, sharp focus, hyper realistic, art by ralph mcquarrie and greg rutkowski and alphonse mucha and j. c. leyendecker and katsuhiro otomo

i think maybe cause i have so many concept artists for clothing lol

!dream darth vader wearing traditional samurai armor, ukiyo-e style painting, textured sepia background, colorful

oh got you. yeah uh i think it uses a similar tokenizer and encoder at least bc its clip, but im not 100% sure how comparable they are. the limit won't be much different though, or at least not any longer.

what was your local set up like? i had it running on apple silicon at around 10 minutes per image but that was the 1-3 model checkpoint just before the public release, and literally just cpu only

gonna try out the public model with mps at some point but idk if its super worth experimenting with

!dream Nas and Jay-Z rap battle

oh interesting. its the same model so they're probably doing processing of the prompt before sampling idk. i was thinking of some ways to do it after i saw your post but idk what the demand would be like

but ill take a look at their code, thx

!dream Jay-Z's head on a camel, colorized, 2001, realistic, 8k, desert, Shawn Carter, Jay-Z, Jigga, Hov, humps, sand

!dream oil painting, purple, red, and blue tones, heavy texture, bloodborne, gothic victorian setting, eldritch horror, HP lovecraft

!dream oil painting, purple, red, and blue tones, heavy texture, bloodborne, gothic victorian setting, eldritch horror monster, HP lovecraft, werewolf, Cthulhu

!dream oil painting, yellow, red, orange, and brown tones, heavy texture, bloodborne, gothic victorian setting, eldritch horror monster, HP lovecraft, werewolf, Cthulhu, high-contrast lighting

!dream a painting of an Angel falling from Heaven into darkness in the style of Caravaggio

yeah it seemed comparable if i keep it at 75 tokens on the GPT3 tokenizer i never get the error

I’m running a 5800x and RTX 3080

I can render one sample at 768p in 18 seconds or at 512p in 6 seconds

yeah i’ve heard it doesn’t use the GPU on the M1, i didn’t even bother installing it on my Macbook lol.. I heard stability is working on Mac and AMD support tho

I made this with Dall-E 2 today, took me like 40 prompts..also used a lot the Outpainting mode.
Had to use Photoshop to add and combine some stuff (like the ghost, the swirly effect and some borders of the final product)

The prompt was:
Man riding a high-speed motorbike through a gloomy dark green tunnel at night, 4k quality, highly detailed, low angle photography.

Started with this:

End result: