Neural network-generated custom portraits (aka "Why haven't I seen anyone suggest this before?")

[Swatcher]

  On 9/15/2022 at 4:48 AM, thelee said:

[screenshots of avowed, an open world RPG that combines Fallout: New Vegas and Skyrim]

Do any very specific pillars of eternity terms work? [Eothas sunbathing on a beach with Berath]

[Screenshots of Josh Sawyer's latest independent game, a pixel art Bike Shop RPG simulator]

Expand  

You got it.

Edited September 15, 2022 by Swatcher

[Swatcher]

Oh wait, nvm it's actually expiring Sep 18 (Sunday). "Sep 17" is just the last day i can use it. Anyway,

screenshots of avowed, an open world RPG that combines Fallout: New Vegas and Skyrim

SHUT UP LEE. can't believe you're just leaking these games. Who do you think you are, Jason Schreier? istg gonna get this whole thread nuked

Screenshots of Josh Sawyer's latest independent game, a pixel art Bike Shop RPG simulator

9 CREDITS LEFT

 

  Reveal hidden contents

EDIT: Full-resolution versions, and testing direct-linking from DALL-E 2's website. yknow, idk if image embedding will work (or rather, if these URLs won't break in the future. Spaghetti-length URLs are dodgy).

EDIT2: Oh yeah no, images broke 1 day later. No embedding full-res from DALL-E 2 then. Hmm I could also "publish" them then share them publicly at full-res, but problem is that'll show my real name. It's better to keep on downloading-->uploading each individual image.

Kinda tedious, esp having to downscale file size bc of the PUNY data limit on this website. "1 MB", cmon bro. I guess image hosting is expensive. Ok whatever, full-res for real this time:

  Reveal hidden contents

 

Edited September 24, 2022 by Swatcher

[Swatcher]

Anyone, feel free to request the full resolution (1024x1024) for any of these pics if you really want them

I don't think they're going to get deleted from my DALL-E 2 account if they get too old, but who knows how OpenAI works?

Edited September 28, 2022 by Swatcher

[Swatcher]

Oh yeah, unrelated but I don't think I've ever showed off properly what the built-in Outpainting looks like in DALL-E 2. Here, a .gif:

Final Result:

  Reveal hidden contents

Couple more

  Reveal hidden contents

It's kinda 'zooming out' in a sense, if it wasn't obvious yet. The above is from the subreddit. A random few more w/ before & after

https://old.reddit.com/r/dalle2/comments/vu2jgn/first_time_outpainting/

https://old.reddit.com/r/dalle2/comments/x2tynm/i_outpainted_a_3440x1440_ultrawide_wallpaper_of_a/

https://old.reddit.com/r/dalle2/comments/w0gzy1/inpainting_an_outpainted_meme_photograph_of_a_dog/

How high can you go? https://old.reddit.com/r/dalle2/comments/xbgynq/i_doubled_the_pixels_59_times_and_pressed_it_into/

  Reveal hidden contents

The rest of the Mona Lisa

Close-up detail of another one

It's like poetry. They rhyme.

This is how you get the super high-res pics *and also how you burn through all your credits

Edited September 28, 2022 by Swatcher

[Swatcher]

Anyway, still open to taking any request. Gotta burn through these somehow

I did a dumb and tried inpainting on multiple DALL-E Mini faces from one post at the same time, despite how bad faces went last time. Look

  Reveal hidden contents

Result:

Do you really want to see these nightmares up close?

And the close-up:

  Reveal hidden contents

a

b

c

d

Oh my god the gaping eye meatholes--it's on like half of them. Inpainting faces, everybody.

Well, if my DALL-E Mini advice was anything to go by, you might be able to frankenstein-edit each of these 4 versions into a functional portrait, if you're crazy enough to try it. I sure am not. I don't wanna look at these any longer than I have to.

Just take apart the 'good features' & edit them together, if there actually were any haha.

 

8 CREDITS LEFT

Edited September 28, 2022 by Swatcher

[thelee]

what happens if you use the out-painting on some deadfire portraits, or the pirate-ship deadfire hero image?

 

man that outpainting is insane

[Swatcher]

This one, right?

  Reveal hidden contents

1. I could, but in the .gif you saw those 1024x1024 generation squares, they each cost a credit. I'd probably use up my few remaining and then some. High resolution is expensive yo
2. I could INSTEAD scale the img resolution down to fit inside the square (like this) and generate one square around it though, at the cost of resolution. 1 credit only

But if I go big with #1, I'd save it for the the last day(17th). I want to see if anyone else wants a request before I destroy my remaining credits.

Edited September 24, 2022 by Swatcher

[Swatcher]

btw, a lot of these editing tricks can work with versions of Stable Diffusion too, inpainting and whatnot.

It's also worth mentioning that Stable is good but not as good with realistic-looking stuff as DALL-E 2 is. However, with digital art it's just as good. In fact it emphasizes it. Talk about ideal! Doing these digital art portraits is literally one of the best use cases.

Edited September 28, 2022 by Swatcher

[Hawke64]

It looks impressive, despite the issues with faces.
If requests are still open, it would be interesting to see Death and Nature Godlikes. ("Painted portrait of a faceless demon paladin with horns in modern digital paining style" and "painted portrait of a male dryad cleric with antlers in modern digital painting style"?)

[Swatcher]

  On 9/15/2022 at 10:34 PM, Hawke64 said:

If requests are still open, it would be interesting to see Death and Nature Godlikes. ("Painted portrait of a faceless demon paladin with horns in modern digital paining style" and "painted portrait of a male dryad cleric with antlers in modern digital painting style"?)

Expand  

ye

 

 

*painting

Watch the spelling. Can't afford to re-run it. I can't fix it once I enter it, either.

Edited September 16, 2022 by Swatcher

[Swatcher]

"despite the issues with faces"  Ehhhhh for DALL-E 2, the faces haven't really been the issue (but the two inpainting attempts were a little f$cky tho. Don't inpaint faces). Pretty much almost all of them were good, which is extremely impressive for an AI. This year's been ridiculous with AI and deep learning, for real.

The real big issue is matching PoE's art style. Super hard to get the game's art style without the AI recognizing the game. Ironically, the weaker AI on page 1 DID recognize PoE, despite having way, waaaay less training data. That's why page 1 matched better. Seems kinda silly in retrospect, but it makes sense.

Your prompt requests:

[Death Godlike] Painted portrait of a faceless demon paladin with horns in modern digital painting style

El diablo

 

[Nature Godlike] painted portrait of a male dryad cleric with antlers in modern digital painting style

6 CREDITS LEFT

Edited September 26, 2022 by Swatcher

[Swatcher]

tbh I wouldn't bother with PoE portrait requests anymore, specifically. You've all seen what they look like in the prev page. Unless anyone has some untapped crazy idea or prompt, the portraits are all going to look stylistically like the results from this thread. I've wrung it out.

This is as close as it gets for DALL-E 2; its weakness is not recognizing "Pillars of Eternity". Try again with Stable Diffusion instead. SD recognizing "Pillars of Eternity" will make a huge difference. On top of the difference I noted about the AI 'styles' above

Edited September 28, 2022 by Swatcher

[Hawke64]

Thank you. The portraits look good.
I might try Stable Diffusion at some point.

[Swatcher]

Do it, bro! For real, I'm prob taking a break after this for a little bit anyway; maybe lurk and see what other people make.

I'll plug this site again: http://lexica.art, if you wanna get a taste of Stable Diffusion images by other people including portraits

A good portrait requires the ideal male body

(link)

Eyyyy, my 50th post and 1,000 views at the same time. Timing. Glad this AI stuff is finally getting some attention here, for the first time.

Edited September 23, 2022 by Swatcher

[Hawke64]

Thank you for the link.
(That looks.. curious. Would say that all portraits are usable if you remove the additional heads and hands, which does not take long, so very nice).

[Swatcher]

It's almost expiration time, and no one's coming in. Guess it's time for me to spend the rest on dumb crap no one wants. Here we go! Eh I'll just go halfway on 1 & 2 for the outpainting. I considered a bunch of dumb memes or sh!tposts if I can't even use this thing for portraits, like Aloth getting mugged in a NY subway or something.

Edited September 26, 2022 by Swatcher

[Swatcher]

And I've run out—The small outpainting of PoE 2's official art. I might've been too hesitant with wordy prompts, but honestly I was expecting some gibberish garbage from my over-detailed prompts! Maybe older AIs made me paranoid

I wish this website didn't scale down highres uploads like this to a half-sized .jpeg. The original is a 2944x1024 .png. Is there a good website to direct link images that ISN'T imgur? nvm found a reasonably reliable site. Full resolution 

Process:

  Reveal hidden contents

First scaled it down to 990x617. Sadly I don't have the credits to waste on good-looking highres generations. Decided to spend only 3 credits on this, and size it down appropriately

"beautiful" in prompt, sometimes i forget this ai understands emotional words. That word weirdly makes a HUGE difference to art.

(Note for if you have DALL-E 2: Below img is a bad example to follow—if you use this don't waste a credit like I did by doing the square at the literal edge. Bc it's not blending into anything, you'd get this. Square needs to overlap w/ the image at least a tiny bit unlike this, in order to seamlessly blend the edges.)

Marginally changed prompt for the right square, removed the "fighting other pirate ships" and changed weather

Also each time I did a square I got 4 results to choose from. From below, I chose #3.

3 credits to outpaint this (-1 mistake)

 

Y'know, since I got 4 images to choose from for each square, I could probably stitch together something even better-looking from the extra images.

Edited September 26, 2022 by Swatcher

[Swatcher]

Kinda gave up halfway through. I tried stitching something better together from the extra images.

I added a little orange fire glow to the left ship's mast (only non-AI thing) and moved a couple things around, and besides that this is 100% AI images

Full Resolution

Edited September 26, 2022 by Swatcher

[thelee]

aw, full resolution image won't load. no idea what those things on the bottom-right of the image are (some water delemgan maybe? lol) but pretty impressed with the rest of it. I wonder how good image AI will be in a decade O_O

[Swatcher]

  On 9/23/2022 at 3:31 PM, thelee said:

I wonder how good image AI will be in a decade O_O

Expand  

...

It'll be pretty good.

(Future war - Terminator 2)

Edited September 26, 2022 by Swatcher

[Swatcher]

"I'm prob taking a break after this for a little bit" 

A promise is a promise. I won't be doing much DALL-E 2 stuff from here on—I'll get only 15 credits a month now, which resets every 18ᵗʰ. I don't know if I'll be checking in too often, but I'm still open to checking requests ITT.

(Link) Maybe someone else could pick up the slack. Check some other AIs

I'm not overly fond of OpenAI, DALL-E 2's creator, anyhow.

DALL-E 2 is good at realistic stuff VS a lot of other AIs(for now). OpenAI's done some other good stuff too like GPT-3. HOWEVER

  Reveal hidden contents

• It's a bit too rich for my blood. OpenAI is excessive with DALL-E 2's steep pricing model ($15 per 115 credits, ~7-8 prompts/$1. It doesn't even offer an unlimited-use subscription like Midjourney or others. No bulk purchases either).
  • As a result, unlike other AIs there's no room for experimentation & practicing prompts. Most people w/ their prompts are going to hesitate learning the best ways or new tricks. They also cut down on the # of images you can generate, 4 images instead of the old 6.
• They're doing censorship filters on your prompts now that they're finally monetized, unlike say open-source AIs, which don't usually bother.
  • Like earlier, I was trying 'Alan Turing on his PC slamming on a keyboard like an idiot/moron', and both those words were blocked. THEN just today I was being hit with a constant fake 'Error message' about the AI not working without any explanation, until I put away a 'secret bad word' and it was suddenly fine. They wouldn't even let me know what the word was. Big secret. I switched random stuff out one-at-a-time. You know what the 'bad word' was? It was "monkey", also "chimpanzee", both in a very normal prompt about them screeching at a computer slamming on a keyboard, but I guess it doesn't matter to the filter. Are they afraid of someone using that word for racism, or anything someone can use to make DALL-E 2 look bad?

I kinda get why, but I'd also rather use an AI with more freedom, not one that actively hinders my prompts w/ unknown censors & higher costs. There's better options out there.

 

Anyway, go check out Stable Diffusion for free. It's new, open-source, highly-trained, and very good imo, in ways even better than DALL-E 2 (besides maybe at generating photorealism. They have different 'styles'). It also has more options/add-ons since it's open-source, such as 'img2img'.

Or, just check one of the other gazillion A.I. models out there. There's definitely no lack of choice. Seriously, A.I. stuff has taken off pretty high, and I don't see it going down any time soon; there's always going to be more and more options in the future. Deuces

Edited October 2, 2022 by Swatcher

[Swatcher]

goddamnit I'm already coming back. how did i not think of this earlier (see edit)

Edited September 30, 2022 by Swatcher