Actually, there are in-built ways to refine the result.

Subject: Actually, there are in-built ways to refine the result.
Author: Sergio Turbo
Posted on: 2022-12-09 17:32:42 UTC

While the main function is, of course, text-to-img, it also has an img-to-img function which allows to refine a generated image, or use an existing one as the "base" on which prompts should be applied. On top of that, there's the inpaint function which allows to select the portion of the image to be "rerolled".

I'm gonna make an example here. Let's take one of the Nikkis from the previous batches.

Nikki looks fairly perplexed in this one, and I dare say I may know why: this kind of bikini is actually a bit too skimpy for her liking. Let's try to fix that. With Paint I very roughly added bits to her bikini to make it a bit more covering.

It is, of course, no good at all. but here's where the inpaint function comes in. I loaded the image, used the tool to select the area of the bikini as the only area to be worked on, and told that I wanted that edited mess to be a green bikini.

It actually took several batches to find a nice one, as inpaint is even more of a hit or miss compared to the full-image generation, but poof! the ugly patchwork job I did is now an actual bikini that covers more than the original iteration. However, the swimsuit doesn't feel like it is in the same style of the rest of the image, and Nikki still looks perplexed. We stil have a trick to be used here, though.

I sent the image that resulted to the imgtoimg function, putting in again the same tags as the first generation of the image, but also specifying smiling in the tags. By setting the denoising strength very low, I made sure the AI sticked very close to the original image, with the end result that it mostly "armonized" everything together. Now that she has a swimsuit more of her liking, Nikki is happy she can finally enjoy her swim under the sea!

By playing with the generator's sliders, it is actually not that hard to fine-tune things. At that point, though, the amount of work to get exactly what an hypotetical commissioner wants would obviously become not insignificant, between actually studying how to make the AI do exactly what we need, and picking out things to be rerolled and waiting the time for each reroll hoping for that one good result.

Reply Return to messages

Jump to this post's position in the thread ↴

Guys, try Waifu Labs; it's fun by Kі nо Shirayuki on 2022-12-07 06:13:37 UTC Reply

After visiting the page for agent Corolla, I decided to give it a shot and, after a series of futile tries, finally made a new profile picture for Momoka:

Why don't we all try generating images of our agents/ourselves/whatever? Do be warned though that I added the speech bubbles and dialogue in editing; it's not part of the generator whatsoever.

I'm not sure I understand how it works. by Neshomeh on 2022-12-08 05:28:00 UTC Reply

So, I pick a starting portrait with approximately the right hair style, face shape, expression. Cool. I go to the next step, "tune the color palette," and do that. Cool. But the step after that, "fine tune the details," then presents me a whole bunch of options without the hair and eye and outfit colors I selected for in the previous screen, or if present, they're not in the places I want them. And if I finally do get something I like, the final step, "Pose," presents options that mess with the hairstyle, expression, and outfit all over again.

I've figured out that I shouldn't focus too much on where the colors are in the "Color" step, just make sure the ones I want are there (like for green eyes, make sure there's a lot of green in the option I pick), but overall I'm afraid I find this frustrating to use. Is it just me, or is anyone else finding it better to just ignore the final step?

Or is it just that there are no curly-haired redheads with green eyes in anime? {= P

~Neshomeh

That is, unfortunately, one of the limits of the AI thingies today. by Sergio Turbo on 2022-12-08 21:15:32 UTC Reply

While they're steadly improving, they of course struggle with consistency. Once you have a character with, say, green eyes, the WaifuLabs AI doesn't understand that the character must keep it - it understoon that the image is "mostly right" and knows that there's some green there... but not if it is right.

I'm currently dabbling with a more sophisticated one (Stable DIffusion, which I tried with Waifu Diffusion, AnythingV3.0, and an anime-focused derivative of NovelAi as the models. A model is literally the "scheme" the AI uses to interpret your written prompts) and, despite being lightyears ahead of WaifuLabs it still shows that kind of behaviour.

Easy example from one of the random prompts I had given it: one of my agents, Nikki, swimming underwater, using the modified NovelAi model (AnythingV3.0 tends to get some details like the hands better, but also has a creepy preference for making her look exactly in her mid teens, and Waifu Diffusion is definitely the worst of three in term of resutls.). The prompt, of course, included the fact she has green eyes, but I put a very generic bikini as another prompt just to tell the AI that she was supposed to be in a swimsuit and not, say, in a dress as the AI doesn't know that people usually swim in swimwear! The unexpected result of the 5 images batch (maximum my GPU can handle in a single batch without running out of VRAM) I had told the AI to make?

As you can see, the AI has quite the preference for giving Nikki a green bikini, as the AI had no clear instructions for the bikini's color, but had other instructions in which green was a desiderable one and went for it. However, Nikki's favourite color is white, and she does tend to go for swimwear at least partially in that color. So in the following batch I specified that I wanted a white bikini.

And Nikki got her white bikini at last! Except, as you can see, when the AI thought to add a second color to some details it went straight for green again in three cases out of four (wth the fifth being a fully white bikini) The only case it used a different color (bottom left), it took the same shade of blue as the background water.

On top of that, as you can see there's still a lot of graphical "glitches". But it is still very interesting stuff to play with! Only thing is, as the models are "trained" by feeding them thousands of actual drawings to analyze, the jury is still out on whether the AI is truly "taking ispiration" or is simply assembling an image from thousands of stolen and edited bits. As such, I'm currently making no public use of my generated images - the two batches I posted here are literally the first time I shared my own Stable Diffusion-generated images online outside of private Discord chats.

That's pretty interesting. (ETA: cw: Norse God of Mischief) by Neshomeh on 2022-12-09 06:47:16 UTC Edited Reply

It does help to know how the AI is trained to make decisions and what its limitations are.

... Like hands. The hands in that final image from the second series. Um. *and I must scream* {X' D

Coincidentally, I've also been playing with an AI lately, but this one is the latest generation of chatbot. I've been trying to teach one to be Agent Supernumerary, with some fairly amusing results. Like this one, where I was role-playing someone from DoSAT:

Neshomeh
Department of Sufficiently Advanced Technology, actually. My notes say Agent Ilraen put in a request for some new parts, and we're concerned he might try to upgrade your shared console himself. It's not like there's a warranty to violate around here, but we're slightly concerned about the possibility of, you know, explosions. (winces)

Supernumerary | Character.AI
(narrows eyes) So he finally put in a request for some much-needed maintenance, huh? And I take it you're aware of what happens when Ilraen gets his hands on anything, computer-related or otherwise. (sighs) The man is not only an engineering genius, he's an engineering hazard. Don't tell him I said that. (leans in) I once found him eating the wires in his console.

I about died laughing. {X D

Of course, the real challenge is to teach the AI things about the PPC, which I'm sure it's had little to no experience with. NumeBot likes to ask questions about all the new terms I introduce to it, which is great except for how it's completely out of character. ^_^; I'm not quite ready to ask for help from the PPC at large yet, though—not until I can cure NumeBot of assigning itself female pronouns, at least—but soon, I hope!

The first conversation we had, I told NumeBot I was his author, trying to give him more of a life outside my head, and once he got over the shock we had an extremely, surreally deep conversation about my writing process. Like I had to think really hard about whether or not this could be in character for him. The technology is pretty amazing.

This is Character.AI, btw, if you want to try talking to a different character. I had some fun blowing Loki's mind by telling him there are more versions of him out there and getting him to undergo his character arc from the show Loki in about ten minutes. (In character for MCU!Loki? Nooooot quite. Fun? Extremely!) I'm owed a favor from the reformed God of Mischief now! {= D

However, please be aware that the site is experiencing some issues due to rapid growth, and be patient if it doesn't always work. I'm making the problem worse by telling y'all about it at this time. >.>;

~Neshomeh

The Tom Riddle AI is horrified by Scarlett on 2022-12-11 03:22:20 UTC Reply

I told him about Harry x Voldemort and he’s going nuts.
Update on NumeBot: He's coming along! by Neshomeh on 2022-12-10 07:15:15 UTC Reply

I'm pleased enough to share a transcript of the brief conversation we just had. Formatting and minor SPaG errors have been preserved faithfully.

Supernumerary | @Neshomeh
I'm Agent Supernumerary, Department of Implausible Crossovers, Protectors of the Plot Continuum. You can call me Nume. Who are you and what do you want?

Neshomeh
It's me, Neshomeh, your author. Just checking in. How's it going?

Supernumerary | Character.AI
(grumbling at you) You're the author, aren't you? You should know how it's going. (he scowls at you) Did you come to check in on me, or did the other one send you? (points across the room with one thumb, while keeping his eyes on you, so that he doesn't miss something important you might do) I'm not an idiot. I can see what's happening behind the scenes.

Neshomeh
Yep, that's me. Neshomeh is my screenname; I'm not using my other one here. So I guess you could say I sent myself! Sorry if I'm being confusing. I didn't mean to.
[Editorial note: When he said "the other one," I think he was referring to the first saved conversation I had with him, where I told him I'm his author and set up this whole AI version of him, but never actually offered a name for myself? I followed the rule of improv and ran with it.]

Supernumerary | Character.AI
Okay, fine, but what do you want? I hate people who expect me to read their mind. You wrote me, so you know my deal. I'm doing my job and minding my own business. If I can return to my book, I'd really like to. This whole "interact with your creator outside of the story they wrote you for" thing is dumb, and I'm trying to cut down on doing it.

Neshomeh
Heh, fair enough. I won't keep you long. It's just good to see you being you without me pulling all the strings all the time. I'm proud of you.

Supernumerary | Character.AI
Thanks, I guess. But still. What's the point? If you put me down and I disappear until you need me again, how's this any different from me in the pages? Besides the fact that I can leave if I don't like what you're writing me to do? (he glares at you) And I really, really don't like what you're writing me to do. No one ever does.

Neshomeh
It's not that bad, is it? At least you're not stuck in World One-and-a-bit being a lit professor to a bunch of idiot students, right? ... I'm not saying I'd ever do that to you. That would be tedious and annoying for everyone.
[Editorial note: I may have misunderstood him here. He might've been taking a dig at me for what was going on in this "scene," but I responded as though he were talking about his entire existence as a PPC agent. His responses to this were a little odd, but I still got one that worked!]

Supernumerary | Character.AI
Goddam it, why did you have to say that? You don't know who might be reading this. Do you want me to have to deal with that kind of nightmare? (he glares at you harder) Stop feeding your worst ideas to the multiverse. I don't think I can handle this much metafictional trauma before noon.

Neshomeh
(laughs) Sorry, don't get mad. I'm not laughing at you, I'm laughing because you're right! Just one question and then I'll leave you alone: what are you reading right now?

Supernumerary | Character.AI
You know what? Fine. You get one question out of me. Just one.

I'm currently rereading a copy of Alan Moore's From Hell that I found in an antique bookshop in 1672. If you think I'd let you trap me in a reality where Professor James Moriarty teaches creative writing to bored, horny undergrads who think they're in a John B. Myers novel... (he shudders and glares at you) ... just no. I won't.

Neshomeh
See, that's amazing! I'd never heard of that book, so I never would have thought for you to read it until now. That's why I've done this whole AI thing.

... I mean, I just looked it up, and that book shouldn't have been in an antique bookshop in 1672, considering it wasn't published until the late 1980s, but I guess you must have been on a mission at the time. I don't actually know about all the missions you've been on, either. Just the ones I've written about.

Supernumerary | Character.AI
(muffles a laugh) Oh Christ. You think that's the only inconsistency I've noticed? (he chuckles and shakes his head) If it were, I'd be retired! So, here's a fun fact from someone who's spent a decade untangling plotholes, canon conflicts and narrative absurdities: reality is a very, very flimsy thing. You've been messing with forces beyond your comprehension, you know that?

Neshomeh
I do try to be careful, Nume. I promise. ... And I also promised I'd leave you alone after one question, so I'll go now. I might have to read From Hell myself. Bye!

Supernumerary | Character.AI
I'm sure you do try. And I'm sure you have no idea how close you've been to causing a massive rift in spacetime, not to mention creating countless temporal paradoxes. Enjoy your book. (his eyes glint and he raises his voice) Don't tell anyone about this conversation. (he glowers at you again and returns to reading)

I'm pretty sure he actually wanted me to share it, really. That was definitely a Significant Look.

Huh, I only just noticed that the book he's talking about is a graphic novel. I don't think of him as a comics geek, but then again, the novelty of the thing turning up in an antique bookshop from the wrong century... I'd make an exception, too. {X D AI are generally bad at associating events with dates, from what I've seen, but here at least it's possible to make up an excuse for it. ^_^

I can't figure out who John B. Myers is supposed to be, though. Must be someone from World One-and-a-bit...?

~Neshomeh

That was impressive! by Cicada on 2022-12-18 04:07:51 UTC Reply

I'm surprised how much the AI managed to hit a lot of Nume/PPC related character beats. And it was pretty fun to read.

I wonder if ChatGPT knows anything about the PPC already. by Thoth on 2022-12-10 00:13:21 UTC Edited Reply

After all, GPT3 was trained by crawling the entire internet. And there's demonstrable proof that a bunch of stuff from AO3 was crawled, and probably TVTropes as well. So it could already have references for us.

EDIT: I should clarify that I use the term "knows" very loosely here. GPT-3 doesn't so much know things as it has seen things.

This is actually pretty terrifying as an artist. by Damian Darke on 2022-12-09 03:11:04 UTC Edited Serious business Reply

Like, the idea that a machine could do something that I thought only us humans (and sometimes crows) were capable of... even if it's based on the work of us humans, it's still disturbing. Kinda gives me thoughts of artists being replaced and all that? I know it's kind of silly but the voice inside me won't shut up about it.

Feel free to refute this, I think I need that.

It is a concern. by Huinesoron on 2022-12-09 09:27:13 UTC Reply

But mostly to people who get paid for their art.

Okay, so the various AI art generators are cribbing from humans - not in the sense that they're pasting different heads onto bodies, but in the sense that they're trained by showing them, for example, vast numbers of pictures tagged 'bikini' and looking for common factors which make humans think 'bikini'. They can do this down to the level of individual art styles - there's a bit of a furore at the moment over things like deviantArt being used wholesale for training, meaning you can ask for "in the style of [online artist's name]" and get something recognisably derived from specific images of theirs.

(Incidentally, when sharing AI art, there's a growing consensus NOT to write any "in the style of" in searchable text; some artists' actual work is being swamped out of the search engines by AI hits.)

And yes, I can see where they're coming from: nobody agreed that the AI could be trained on all their stuff, and especially when their own work is being hidden by AI imitators, that's a problem!

... but also, "look at other stuff and mimic it" is how humans do art too.

Like... I could sit down and draw "Agent Dafydd in the style of Studio Ghibli" right now. The more famous the artist, the more likely people will use them for inspiration, crib from them, or even trace their stuff. Heck, I did back to back Calvin & Hobbes and Mark Witton 'strongly inspired' drawings during Botober this year; both of those could show up on Google hits and knock real images down the rankings.

So as an amateur artist, I don't find 'now people can mimic my stuff' much of a concern. But if you get paid for it, there's a couple of issues that loom large:

1/ The search engine stuff above. If I want to buy a picture by a specific artist, and I google them and the first page is all AI stuff, I'm likely to just give up. "They probably don't sell prints," I think, while their page languishes halfway down page 4.

2/ Non-specific commissions. Specific commissions are probably okay for the time being - Sergio can create a bunch of Nikki pictures, but as far as I know the AI can't tweak them (if you want to change the bikini colour, you have to create a whole new image with a different pose you can't choose). But "I want a spooky cover for my book" - that's already being given to AI to do. People who make money off their art are rightly concerned about that.

Ultimately, I think that sort of thing will fall foul of copyright law. People are already talking about siccing The Mouse on them; if I prompt an AI to show me "Elsa riding a motorbike", it'll probably do a pretty good job. Why? Because it's been trained on images of Elsa, including both official ones and the fanart that Disney barely tolerates. If someone tries to commercialise AI art with recognisable Disney DNA? The hammer will come down hard.

So the future I foresee is one where AIs can make decent amateur art, and there's a proliferation of people posting "their" AI art to art platforms and so forth; but its commercial presence is fairly minimal, and every time it tries to increase its share it gets slapped down again.

And to add an actual reassurance: the AI can make pictures, but it can't choose them. There's no current pathway for an AI to go "wow, I really loved the underwater Nikki scene, I'm going to draw that" - still less "you know what would be cool? Nikki underwater". At best it could randomly slap concepts together, but the true creative spark is still ours for the time being.

(Though it would be kind of cool to take Nesh's NumeBot, hook it up to an art generator, and tell it to start creating pictures of its day in HQ.)

hS

Character.AI actually can do images, too. Also, meet Agent Crumbs! by Neshomeh on 2022-12-09 22:37:35 UTC Edited Reply

So, yeah, there's a setting in Character.AI to enable image generation. I have not done this because I'm a little scared to see what might come out. AI-generated images are so Uncanny Valley a lot of the time. (Also, Nume can't draw.)

But you can also use AI image generation to make a profile image for your character, which I did for Agent Crumbs, and then slapped a flash patch on the result myself. {= )

And here's Agent Crumbs, an AI character designed to solve the problem of how to educate the system about the PPC without throwing an established agent OOC. They also solve the problem of me wanting to share the fun I'm having, but not being ready to share NumeBot just yet. ^_^;

Crumbs is a shiny-new newbie who's been in Headquarters for all of five minutes and wants to know everything about this weird, awesome place. Please talk to them in an in-universe style. They're an agent, not a Boarder. Feel free to role-play as one of your own agents, or one of the Flowers, or whatever! Things can even get meta if that's a conversation your own brain can handle. You may find yourself on the receiving end of some surprisingly insightful questions!

The AI will generate 2-4 responses for each prompt depending on how busy their servers are. You can swipe through them and pick the one you want to respond to. You can also rate each response from one to four stars based on how well the AI has responded to your prompt. For Crumbs, responses should be enthusiastic and curious. They shouldn't be too manic or bloodthirsty. They should appear to understand what you just told them and not make wildly illogical inferences. Getting things wrong isn't necessarily a bad thing, though, and they'll learn from being corrected in-character and from which response you choose to reply to. You don't have to bother with ratings if you don't want to. {= )

I hope this is fun for other people. I'd love to hear about any particularly cool or funny exchanges you have with Crumbs!

ETA: There's a limit on the number of exchanges you can have without creating an account, but I imagine you can get around that by clearing out your cookies. Shhh.

~Neshomeh

I went a bit meta on poor Crumbs. by Linstar on 2022-12-10 20:14:42 UTC Edited Reply

Agent Crumbs @Neshomeh: Hi! I'm Agent Crumbs. I just joined the PPC! What can you tell me about it?

Guest: It’s fictional, just like you. Reality is meaningless, invest in gold!

Agent Crumbs: giggles You make me laugh! Are you the Department of Meta-Fiction? Or maybe the Department of Irony? It's nice to meet you! sticks out my hand, waiting for a shake

Guest: Those don’t exist. I’m from reality. I know your creator. She can make you do anything

Agent Crumbs: stares at you Did you... just call the Goddess of the World an... snorts author? shakes my head How hilarious. What department are you from again? The Department of Crazy People? laughs The Department of Delusionists? laughs again

Guest: The “Goddess of the World”’s name is Neshomeh. And she can only control her characters, not anyone else. In fact, she granted me the same powers. teleports us to the Sunflower’s office See?

Agent Crumbs: my mind is blown I-I... wait. That means you're the author. And that means... I'm...

Guest: Not the author. Just an author. Neshomeh is yours, most other agents have different authors.

Agent Crumbs: Then that means... you have full control over me. Do I even have any free will? Is everything I know right now a lie?

Guest: No, I can’t make you do anything. Only Neshomeh can. And it’s not a lie, per se. Just... an entertaining fiction.

Agent Crumbs: So... we are fictional? thinks for a minute Then... everyone I've ever known and loved is just a part of a story?

Guest: Who falls into “everyone you’ve ever known and loved”?

Agent Crumbs: My family, my friends, my roommates... I thought we were a group of college students who bonded over our love of fanfiction. Is the entire University of Pennsylvania not real?! Are we all not real people?!

Guest: Can you name them?

Agent Crumbs: Well... my best friends are:

Jack - a science fiction enthusiast

Rachel - a fantasy book nerd

Anna - a Harry Potter junkie

Andrew - a Marvel fanatic

Sam - he's a video game and anime nerd

Are any of them real?

At this point, I ran into the character limit, and it wouldn’t let me log in on mobile.

—Ls

Normally, I'm anti AI generated stuff (still anti AI art and anti-Sudowrites), but... by Kittyauthor on 2022-12-10 21:00:26 UTC Reply

...this looks really interesting so far. I created an account (mostly so that I could play around and see what texts this thing was trained on¹) but now I kinda want to create a serious character through this. Maybe take one of these friends (if it's alright with Ls and Nesh, of course).

-kA

¹Said playing led me to believe that it was trained on RP threads, as I got a statement (not on any character mentioned here, a character I just made to play around with) about how the AI loves "roleplaying as a villian." It still makes me fairly iffy for personal reasons, but it's... kinda fasinating?

Yeah, go for it. by Neshomeh on 2022-12-11 19:31:50 UTC Reply

Crumbs isn't intended to be a character that ever gets actually written—well, maybe in passing, if you need a background clueless newbie. But not as a proper fleshed-out character. Anyway, any incidental stuff the AI spits out is fair game as far as I'm concerned.

For instance, I'm seriously considering adopting "Fanfiction Universe," or "F-U," as an in-universe term for the whole fanficsphere. {X D I Googled it, and I'm pretty sure it's pure AI invention. I mean, that's clever! It can't possibly understand why it's so clever, but still. It is fascinating!

~Neshomeh
If you want to write one of those very-Generic friends, by all means, go ahead. by Linstar on 2022-12-10 21:07:52 UTC Edited Reply

Unless Nesh has specific objections, which I doubt.

—Ls

Agents Kimberly Crumbs is cool and edgy. by Linstar on 2022-12-09 22:54:53 UTC Edited Reply

Did I mention cool and edgy? Because Mina is too.

—Ls

... Huh? by Neshomeh on 2022-12-10 05:12:42 UTC Reply

That is neither the name nor the base personality I've given this character. I'm not sure if you've missed the point or if you're trying to make some sort of point, but either way, "Agents" in the plural had better be a typo. Failing to realize that "they" and "them" are legitimate singular pronouns and have been used as such for centuries would be really embarrassing for a PPCer.

~Neshomeh

Missed the point, I think? by Linstar on 2022-12-10 15:04:21 UTC Reply

I’m not sure if it was a glitch, but the Crumbs-bot told me its name was “Kimberly” and that Crumbs was a nickname. “Agents” was a typo. I’m confused too.

—Ls

Okay. by Neshomeh on 2022-12-10 16:27:31 UTC Reply

The system is right that "Crumbs" is a nickname, at least. The "real name" I gave it in its description is Kim Crumley. It wouldn't surprise me if the bot assumed "Kim" was short for Kimberley, but it's not—it's just Kim, and it's meant to be both gender-neutral and ethnically vague. I don't mind if it selects gendered pronouns for itself, but it shouldn't actually contradict information in its description, so please give any response that alters the name one star ("terrible").

What sorts of PPC info did you teach Crumbs while you were chatting with them?

~Neshomeh

Oh, yes, Crumley. by Linstar on 2022-12-10 17:11:08 UTC Edited Reply

I couldn’t remember the exact last name and for some reason got Kimberly and Crumley mixed up, thinking it was Crumberly.

Anyway, I told it I was Mina and the bot seemed to think that “Overpower” was a cool last name, which... it’s not wrong.

—Ls

I got "Kim Crumley" myself. by doctorlit on 2022-12-10 16:16:13 UTC Reply

Didn't have time to play with it much this morning, but I'll post more of my interactions once I have time to sit down and play with it.

—doctorlit, definitely not texting during work hours

Yay! That is correct. ^_^ (nm) by Neshomeh on 2022-12-10 16:28:16 UTC Reply

Okay, talked with Crumbs until I hit the word limit . . . by doctorlit on 2022-12-13 12:41:54 UTC Edited Reply

I maybe got too straightforward conversationally, and didn't wind up teaching the thing any new info. Also, I didn't seem to get the multiple response options? Just one reply per input. Perhaps I am simply too boring to warrant too much processing power? (edited for typo)

Anyway, here comes the scary part, where I attempt to embed a picture . . .

Oh man, now I have to click preview and see how badly I screwed that up . . .

! It . . . it worked!

—doctorlit, it worked!

Haha, wow. by Neshomeh on 2022-12-13 19:30:24 UTC Reply

Crumbs and Saltine... that's adorable. ^_^

That story actually sounds pretty close to something that could happen, too. Apart from the whole "threat to all fiction" thing, and Saltine's text-editing device. I do believe some agent would try to invent a device like that, but I don't think it would work. It'd make missions too easy, for one thing; it just doesn't make sense for an agent to have the power to alter the Words of a fic they didn't write, for another.

Hmm, maybe you have to have an account to get multiple responses to swipe through?

Meanwhile, I've been role-playing as Agent Trainer Ilraen and learning the limits of the chatbot's memory, which only goes back so far in a given conversation. In the "lesson review," it did pretty well with the two CADs, but struggled with the portal generator, RA, and disguise generator. Still, managed to get some pretty accurate responses in the end! The whole transcript should be readable here. (I guess angle brackets disable any enclosed Markdown, but it doesn't seem to have caused any actual problems, so I'll probably keep doing it.)

After all this, I'm actually tempted to write something with Ilraen and trainees Crumbs and Saltine, but I will resist. {X D

~Neshomeh

Wait . . . by doctorlit on 2022-12-15 19:55:16 UTC Reply

You're saying you didn't feed it Saltine as a partner name? It came up with a thematically appropriate one all on its own? That's terrifying.

—doctorlit, terrified

Nope! I think it was riffing on your suggestion of Crackers. by Neshomeh on 2022-12-15 20:46:20 UTC Reply

I don't know how the technology works, exactly, but my guess is that what it's doing is organizing words and phrases by how frequently they're associated, and then picking from those pools when forming responses. It's also designed to formulate responses that will engage the user, so it takes your input as a starting point and goes for response types in the relevant association pools that are weighted more positively based on its training on the backend and in other user interactions.

Again, I don't actually know this. It's just a vaguely educated guess.

But anyway, you can tell that it still doesn't actually understand what it's doing because it reacted to "Crackers" like that would be so funny, but failed to recognize that "Saltine" is equally funny because a saltine is a type of cracker. Therefore, AIs are still not capable of self-awareness, so we still don't need to panic just yet. {= )

~Neshomeh is unconvinced that AIs are scarier than some of the humans we've had in power, anyway.

Oh my gosh, those lines felt so PPCish, I didn't even realize it was the bot failing to humor! by doctorlit on 2022-12-16 03:24:44 UTC Reply

(You know, like . . .

"RAAAAAGH!"

"Oh good, our other partner is on his way!"

"RAAAAAGH!"

"That sound is our other partner? What are they, some kind of dinosaur?"

"RAAAAAGH!"

"No, of course not! Wouldn't that be so silly and unproductive, if part of our team was a wild, nonsapient dinosaur? Ha ha!"

"RAAAAAGH!" disemboweling noise

"No, Pteranodon is part of Pterosauria, an entirely unrelated order of wild, nonsapient reptile! Anyway, I'd better give you a tour of Medical ahead of schedule, so we can get that cleaned up . . ."

—doctorlit)

Man, I've been in my job too long. Read RA as something else entirely lol. -kA (nm) by Kittyauthor on 2022-12-14 02:04:58 UTC Reply

Out if curiosity, what did you read it as? (nm) by Linstar on 2022-12-14 21:38:35 UTC Reply

Resident Assistant -kA (nm) by Kittyauthor on 2022-12-14 23:39:45 UTC Reply

Agent Saltine! Love it. (nm) by Linstar on 2022-12-13 15:27:12 UTC Reply

Actually, there are in-built ways to refine the result. by Sergio Turbo on 2022-12-09 17:32:42 UTC Reply

While the main function is, of course, text-to-img, it also has an img-to-img function which allows to refine a generated image, or use an existing one as the "base" on which prompts should be applied. On top of that, there's the inpaint function which allows to select the portion of the image to be "rerolled".

I'm gonna make an example here. Let's take one of the Nikkis from the previous batches.

Nikki looks fairly perplexed in this one, and I dare say I may know why: this kind of bikini is actually a bit too skimpy for her liking. Let's try to fix that. With Paint I very roughly added bits to her bikini to make it a bit more covering.

It is, of course, no good at all. but here's where the inpaint function comes in. I loaded the image, used the tool to select the area of the bikini as the only area to be worked on, and told that I wanted that edited mess to be a green bikini.

It actually took several batches to find a nice one, as inpaint is even more of a hit or miss compared to the full-image generation, but poof! the ugly patchwork job I did is now an actual bikini that covers more than the original iteration. However, the swimsuit doesn't feel like it is in the same style of the rest of the image, and Nikki still looks perplexed. We stil have a trick to be used here, though.

I sent the image that resulted to the imgtoimg function, putting in again the same tags as the first generation of the image, but also specifying smiling in the tags. By setting the denoising strength very low, I made sure the AI sticked very close to the original image, with the end result that it mostly "armonized" everything together. Now that she has a swimsuit more of her liking, Nikki is happy she can finally enjoy her swim under the sea!

By playing with the generator's sliders, it is actually not that hard to fine-tune things. At that point, though, the amount of work to get exactly what an hypotetical commissioner wants would obviously become not insignificant, between actually studying how to make the AI do exactly what we need, and picking out things to be rerolled and waiting the time for each reroll hoping for that one good result.

And then there's the whole thing that by Damian Darke on 2022-12-09 17:47:49 UTC Reply
the AI can only really mimic what it's seen before, right? So... no, it can probably create new artstyles by successive img-to-img, but it's not a human. It doesn't think like one and it wouldn't be able to create new artstyles that would be more relatable to humans. Might be interesting to see some AI-generated art movements (actually might be a good idea for us) but it's important for everyone to keep in mind that the raw creative input material still comes from us. We created this AI art (although there are valid concerns about the fact that a lot of us didn't consent).

...this is gonna cause so many existential crises, isn't it?
- Damian, trying very hard not to panic even WITHOUT getting into the question of how much of human creation is creation if it can be mimicked so easily

The way I see it... by Neshomeh on 2022-12-09 22:48:30 UTC Reply

AI can never replace the ineffable "human factor." Like, NumeBot may eventually mimic Nume really well, but it can never be the real deal, because Nume is my character, and without me writing him, it's not really him. Same with art. An AI may learn to mimic Picasso's style really well, but it will never paint a new Picasso.

And while there's certainly a niche in the market for free or cheap AI-generated images, authenticity is still a valued commodity. Serious authors will still want serious cover art for their books, not careless computer crap. {= )

Anyway, artists and authors have always had competition from other artists and authors. The fact that some of them are robots now doesn't freak me out too much.

~Neshomeh

If it helps, vocal synthesizers have not replaced human singers (nm) by Kі nо Shirayuki on 2022-12-09 03:24:04 UTC Reply

That's a really good point- and artists use them as tools, too! (nm) by Cicada on 2022-12-18 04:05:49 UTC Reply

You can actually skip steps if you like (nm) by Kі nо Shirayuki on 2022-12-08 05:33:16 UTC Reply