Thread by @fabianstelzer

DALL-E 2 vs Midjourney vs StableDiffusion mega thread: photography, illustration, painters, abstract

these image synths are like instruments - it's amazing we'll get so many of them, each with a unique "sound" 🤯

rules: same prompt, 1:1 aspect ratio, no living artists

"Beatles lego set, catalogue photograph"

funny how this immediately also pulls 60s color schemes

"photograph of a surprised cute crazy robot screaming into a microphone, 1990s TV, VHS texture, thrilling reality TV, film still, realistic render"

DALL-E's really great for facial expressions 🤪

(for @battleprompts :) )

"a human anatomical heart made of flowers, pastel, matte, masterpiece"

MJ wipes the floor with the others when it comes to these types of prompts, aiming for textural details

"Behind the scenes of shooting the moon landing, Hollywood studio, 1969, backstage photograph, astronaut actors, lighting" 😬😅

"pixel art of a beautiful vaporwave sunset with palm shadows, DOS game, retro pixel game, 1990s, screenshot from a 90s retro pixel art dos game"

for pixel art, @KaliYuga_ai has an amazing model, too, that even generates sprites...

for otherworldly devices like this I'd almost always use MJ - it's incredibly creative in putting together the various pieces, eras and materials - perfect for my AI movie @SALT_VERSE

"a spooky 1970s floor plan of a haunted house, worn paper, scary atmosphere, pain and regret"

interesting to see how they try to squeeze in spookiness symbols for this one

"vaporwave underground swimming pool, digital painting, procreate, cgstation, 8k blender, hyperrealistic render, 3d photoshop, award-winning digital art"

MJ water 😍

"Pixar movie scene of a dark skull wizard fighting against Kermit the frog as a gladiator, incredible render, Presto"

DALL-E's usually my go to for scenes involving 2 or more clear "actors" - will be cool to render battle scenes for my prompt fighting game @battleprompts

"portrait of a man who looks exactly like super mario,
photography, portrait photograph"

all of these can do amazing portraits, with DALL-E and SD being better at photos, while MJ does more refined facial textures in a painting context

MJ does "historical" / worn photos really well though

"low poly game asset, Cthulhu monster, 2000 video game, isometric view"

this will be one of the absolute killer instant use cases: generating game assets on the fly

just add a 2d -> 3d model...will be crazy fun for v2 @battleprompts monsters

"a 1990s logo design, cactus online store, dotcom bubble style"

"1990s clip art of a laughing crazy fax machine, windows 3.1, MS-DOS, early computer clip art"

"photograph of a cat with white fur and pink stripes, incredibly soft fur, photorealistic"

#stablediffusion can do incredible photos, too, but you need to be careful to not "overload" the scene

"mathematical art, 1924, litography, abstract generative art"

the moment you put "art" into a prompt, Midjourney just goes nuts

"an incredible bouqet of flowers, highly detailed, black background, wonderful art, astonishing detail, trending on artstation, octane render"

good example of how DALL-E's imperfections look very digital, unlike MJ's - SD otoh is ultra clean here

"Hubble Telescope photograph of an incredible nebula, deep space photography, astonishing photo, wormholes and nebulas"

MJ goes drama
DALL-E goes realism
SD goes wild

some illustrations

when it comes to copying specific styles, SD is absolutely 🤯🤌

personally, I mostly prompt without specific artist references, and esp avoid prompting with living artists

but yes, it can get incredibly close if you do

emulating classic painters works really well with each of these, but SD has an edge here, and DALL-E won't let you do a Botticelli painting of Trump

just for fun, some math

"plot f(x) = 2x"

MJ = X
DALL-E ?
SD ????

how about words?

"the word PROMPT, magnificent typography"

curiously, SD doesn't do words at all

going super abstract can yield interesting results - just describe some sort of vague feeling / concept. also works well if you add "collage"

other notable differences:

- DALL-E has inpainting, which let's you edit part of an image, super powerful

- Midjourney has an incredibly large and active community of almost 1M people

- StableDiffusion let's iterate on a single "seed", staying very close to an output you got

also, staying in the instrument metaphor: you want to play'n'prompt DALL-E / Midjourney / StableDiffusion individually to their own strengths, so it isn't 100% "fair" to use the same prompt

nevertheless maybe helpful as an initial overview of what these can / will output

doing images is fun & incredible, but the real 🤯 starts when you consider what this enables composably

example 1: a community-narrated 70s sci-fi film that only uses image AIs for its visuals - I produce these on my laptop:

https://twitter.com/SALT_VERSE/status/1539588118961168384

example 2: @battleprompts, an experimental twitter game where you summon monsters through prompts, then let GPT3 narrate their fight

https://twitter.com/fabianstelzer/status/1559517876943523846

if you wanna help build this, DM me

and because why not, let's give that first guy a voice and aline. so here's the man himself (=Midjourney), "face-acted" by me (using Avatarify), then added a Synthesia voice - all done in a minute 🤯 - SOUND ON!

these actors are soon coming to @SALT_VERSE 👀

thought it'd be fun to give that first guy an AI voice, too:

https://twitter.com/fabianstelzer/status/1561102406292996096

why the "no living artist" rule?

1- to me, it feels wrong'n'cheap, even tho results can incredibly satisfying

2- i'd immediately do it + monetize if there was a way to compensate artists for their signature style "seed"

@spawning_ has sth in the works here - do follow!