!dream large diamond encrusted golden crown (schiaparelli diamonds crystals swarovski and jewelry iridescent) on head of a Medieval King, Castle, Throneroom, artstation, masterpiece, extremely detailed, very wide angle, long shot —n 4 --seed 2940085326

!dream large diamond encrusted golden crown (schiaparelli diamonds crystals swarovski and jewelry iridescent) on head of a Medieval King, throne room, artstation, masterpiece, extremely detailed, very wide angle, long shot —n 6 --seed 2267557132 --v 0.25

!dream tall anime girl, black straight hair, straight hair bangs, blue eyes, pale skin, timid pose, goth outfit, black bowties, platform shoes, full body pose, ink + pen, Tomohito Oda style, extremely detailed, high contrast --n 4 --seed --steps 25

!dream large diamond encrusted golden crown (schiaparelli diamonds crystals swarovski and jewelry iridescent) on head of a Demon, Miyazaki style, dark souls, artstation, concept art, masterpiece, extremely detailed, very wide angle, long shot —n 4 --seed 2940085326

try --sampler k_euler_a for low number of steps. can set it as your default with !dreamset sampler k_euler_a

The demons are all women. Shocker.

@s so like, how do you uniquely identify each variation since they have the same seed right? as in to recreate it

and i notice if you use generate multiple images you only get one seed. i assume that's for the first image?

midjourney in april vibes 👀

!dream tall anime girl, black straight hair with bangs, blue eyes, pale skin, timid pose, goth outfit, black bowtie on hair, platform shoes, full body pose, pink background, concept art, ink + pen, Tomohito Oda style, extremely detailed, high contrast --n 2 --steps 25 --seed 123456789

1) idk lol. im not sure you can, but there could be a variation number or something. i'll look into it
2) yeah. i thought the seed numbers of the following images were deterministic but im not sure anymore. either way i'll make it show the seed for each image. just need to figure out where...

my guy

!dream hyper realistic depiction of the end of humans, robot takeover, extreme detail, fantasy, digital painting, unreal engine, epic scene, raytracing, cinematic lighting —n 4 --seed 3440606267

!dream tall anime girl, black straight hair with bangs, blue eyes, pale skin, Gainax Pose, black goth outfit, black bowtie on hair, platform shoes, full body pose, pink background color, concept art, digital art, Tomohito Oda style, extremely detailed, high contrast --n 2 --steps 25 --seed 123456789

!dream in BRAZIL --seed 24049218

!dream a car taking a dog for a walk ultra realistic 4k very detailed --seed 3741329328

I've seen some samples for "Asuka Souryuu Langley from Neon Genesis Evangelion", with a few variants like "illustration of", "pixiv skeb.jp", "manga of", "artstation" etc. They generally come out looking like Western illustrations or vaguely 3D CGI-like, with red eyes, no hair clips or plugsuits or school uniforms or NGE-related imagery, instead, emphasizing very long red hair in Star Trek-esque uniforms and soccer shirts. The 'manga' prompts, strikingly, sample photographs of manga volumes with a red-haired girl on the cover.

My best guess is that OA filtered out almost all of the anime in their training dataset (they seem to be extremely aggressive with the filtering, as I guess they have enough data from Internet scraping to saturate their compute budget so they would "rather be safe than sorry" when it comes to PR, no matter how biased their anti-bias measures make the model), and so what we're seeing there is all of the Western fanart of Asuka, which is not all that much so it picks up the hair but not all the other stuff; the soccer shirts are because for some reason she's been associated with the German soccer team so every World Cup Germany is in, there's a whole bunch of fanart with her in athletic gear.

Considering how very limited the training data must be, the DALL-E 2 anime results are arguably actually very good! Better than the ruDALL-E samples, definitely. Global coherence is excellent, sharp lines, basically all works, just uncertain and clearly out of its comfort zone. It is doing anime almost entirely by transfer/priors. You can easily imagine how good it would be if it was not so hamstrung by censoring, and in general, that scaling it up would fix many of the current issues.

My conclusion: between this and Make-A-Scene and compvis, it is clear that anime image generation, and any other genre of illustration, is now a solved problem in much the same way that StyleGAN solved face generation.

EDIT: so far the only explanation I've pried out of an OAer is, to paraphrase, "DALL-E 2 doesn't do good anime because it wasn't trained on much anime, but CLIP knows about anime because it was trained on the Internet" - which completely ducked my point that this should be an impossible failure mode if they used any kind of Internet scrape in a normal fashion, because anime is super-abundant online and DALL-E 2 clearly can handle all sorts of absurdly niche topics for which there could be only handfuls of images available. (EDITEDIT: and this is especially obviously true when you look at models like Stability which were trained on Internet scrapes in a normal uncensored way and exactly as expected, do way better anime...) So, it's increasingly obvious that they either didn't use Internet data at all, or they filtered the heck out of it, and don't want to admit to either or explain how it sabotages DALL-E 2 capabilities. But it does at least explain why DALL-E 2 can generate samples like the Ranma 1/2 '80s style girl+car where the overall look is accurate and the textures/details extremely low quality; that's what you'd get from a very confused large diffusion model guided by a semi-confused CLIP.

Long post about why Anime art is still looking like a 10 year old's drawing compared to the other types of art created

!dream hyper realistic depiction of Jesus returning on judgement day, extreme detail, fantasy, digital painting, unreal engine, epic scene, raytracing, cinematic lighting —n 4 --seed 3440606267

!dream hyper realistic depiction of AI apocalypse, robot takeover, extreme detail, fantasy, digital painting, unreal engine, epic scene, raytracing, cinematic lighting —n 4 --seed 306571049

!dream Mona Lisa in California, modern day, 2022, hyper realistic, cinematic lighting, style of wes Anderson, 4K, extremely detailed, masterpiece, —n 4 --seed 2721927140

!dream Illustration + Asuka Souryuu Langley from Neon Genesis Evangelion, artstation, sitting on a chair, peace sign, red clothes, smiling, eyes closed, digital art, flat colors, flat shading, high details --n 4 --seed 1234567890

I want my credits refund just for the 2nd pic

!dream hyper realistic depiction of the end of time, extreme detail, fantasy, digital painting, unreal engine, epic scene, raytracing, cinematic lighting —n 4 --seed 336018624

Holy s*** the third one

!dream Illustration + Asuka Souryuu Langley from Neon Genesis Evangelion, long red hair, sitting on a chair, holding a peace sign, red clothes, smiling, digital art, flat colors, flat shading, artstation, high details --n 4 --seed 1234567890