mailto: blog -at- heyrick -dot- eu

Unreal humans

I've been playing around with the Generated Photos Human Generator.

Let's just say right from the start - it's bloody marvellous that somebody can rock up and make themselves a fake girlfriend/boyfriend/son/daughter, as appropriate). There's no waiting list, special invitation, sign in, or anything like that. It's completely open. The only restriction is that generating can take a while (expect 3-4 minutes) because, you know, freebie people are low in the queue priority.

It's also extremely impressive in what it creates. As JGH said a couple of blog entries ago, "firmly uncanny valley stuff". I completely agree. These creations look like actual people, only something is just a little bit 'off'. It's so close, you can see why biological actors are getting a bit freaked out about what AI actors could do.

From the point of view of what the technology is now capable of, it's really rather impressive.

I'm going to gloss over weirdness like too many (or few) fingers, because AI seems to like to do things like that from time to time. As I said previously, while the Human Generator is capable of making mistakes (especially with freaky clothing or mismatched eyes), it's generally quite a bit better than my experiences with Dall-E2 that often just throws together stuff at random and thus creates... something that looks like what you asked for if you view it from the other side of the room and squint.

Here, look: A photo of a little girl with dark hair and a light brown dress skateboarding on the moon.

Dall-E2's creations
Dall-E2's creations.

I only have two pictures. It just gave up and offered two blank squares for the rest. And, yeah, in the right-hand picture it took the "on the moon" painfully literally.
But note the block hands, it reminds me of Tomb Raider 2 (that I used to play on Windows 95). Note that the shoes aren't much better.
And the faces. Oh my god.

Aaaargh!
Aaaargh!
(a face only a mother could love)

On the other hand, Generated Photos doesn't attempt to be an "if you can imagine it, we'll make a picture of it". Instead, it concentrates on one thing and does that really rather well.
Take that little girl from a few days ago (at the end). Make her clothing "Earth tones", switch the backdrop to a waterfall, and...

An unreal female child
A little girl that isn't real...
(don't look too closely at her fingers!)
Link to imagery
Yup, standing in the water. I nearly lost my tea over that when the image popped up. ☺

Now, since I'm a Brit and we're a bunch of pessimistic sods (we'd never vote for a political party that said "Make Britain Great Again", for example), let's look at some of the problems.

The first problem is the pricing. I totally get that your twenty dollars a month (or $200/year), will get you twice the resolution, no watermark, commercial use rights, and no need to provide attribution. That's not a big expenditure for a company that might want to create licence-free "people" for advertising campaigns, but what about people like me that aren't interested in commercial use, just looking for a way to not have to wait for several minutes per generation?
I think it might be of interest to people to offer, say, the ability to skip the queue and to obtain a better resolution image for a couple of dollars a month (maybe $20 for a year). In return, there's a more subtle watermark and a strictly non-commercial licence. So it'll offer a better level of generations without the commercial options that aren't necessary.

 

The second problem is probably fixable. WTF are all the options? I gave my girl "Bangs" as a hairstyle basically because I had no idea what it meant. I'm guessing it means a fringe. I mean, I could Google but looking at the output...
Likewise, there are many styles available. Let's say, for her top it can be one of these:

  • T-shirt, Blouse, Shirt, Tank top, Dress, Mini dress, Maxi dress, Midi dress, Shirt dress, Sweater, Hoodie, Jumpsuit.
For her bottom...
  • Jeans, Pants, Wide-leg pants, Cargo pants, Flared pants, Skirt, Mini skirt, Midi skirt, Maxi skirt, Pencil skirt, Pleated skirt, Shorts, High-waisted shorts, Leggings, Trousers.
For outerwear...
  • Pea coat, Trench coat, Jacket, Suit jacket, Denim jacket, Bomber jacket, Leather jacket, Parka, Cardigan.
There are underwear options (not on the children), and given that these disable the top/bottom I'm guessing it's for, shall we say, more adult oriented images.
Footwear...
  • Sneakers, Boots, Sandals, Heels, Flats, Shoes, Loafers, Oxfords, Ballet flats, Platform shoes.
Accessories...
  • Hat, Belt, Scarf, Necklace, Earrings, Glasses, Sunglasses, Handbag, Headband.

Now, I'm not female so I don't know what all those dress/skirt options mean. Plus, wouldn't choosing a dress mean that nothing is required for the bottom? Well, maybe "Tights" should be an option? Uh... I think that's "pantyhose" in Leftpondian.
It would be nice if there was a gallery of examples. Messy bun hair? Okay, looks sort of like that. Pea coat? Oh, one of those. Midi skirt? Nope, it doesn't play music...

For styles, you can choose from a frankly bewildering array of options. Ready?

  • Casual, Formal, Bohemian, Vintage, Streetwear, Hipster, Athleisure, Preppy, Gothic, Punk, Minimalist, Tomboy, Retro, Glamourous, Edgy, Classic, Chic, Romantic, Sporty, Artsy, Ethnic, Business, Casual Chic, High Fashion, Rock, Country, Rugged, Avant-garde, Elegant, Boho-Chic.
A what now?

Most of my generations are either Casual or Classic. I did try making one set to "Business" expecting a pleated skirt and blouse combo like an office worker, sensible shoes, etc. Instead I got a bared midriff and weird clothes. Unfortunately I didn't bookmark the page, and the image ID is... not the UID of the page, so I can't go back and find it. Oh well.
Another place where some examples would be helpful.

 

The third problem, perhaps more for commercial users, is the lack of pose options. There's a reasonable range of basic poses (and a few weird ones), but there are none holding or pointing; the sorts of things that an advertising wonk might want to photoshop in something, like this woman is happily using their latest bag-free tornado-suction vacuum cleaner, or "Oh, extreme levels of glee because my roast chicken looks like a roast chicken and not a cremated toddler - and you too can make perfect roast chicken with our brand new NukeEm3000 oven!".

Also, from the point of view of, say, game assets, there is no way to get images of the character from the sides or from behind.
Which might be hindered by...

 

...the fourth problem, and something of an issue for AI, is that it is wildly inconsistent. I can't say "Okay, I like this human, let's tweak her outfit a little".
On the left, a girl at a waterfall. Remember the girl in the forest from last time? I just removed the glasses and switched the scenery and voilà, voilà, voilà... And on the right? Exactly the same thing, but simply saying to change her location to be "countryside" instead of "waterfall" (where she's walking on water). Sadly she's not ankle-deep in mud and being slobbered over by a cow, so not the sort of countryside around here. ☺ In fact... what?. I was expecting at least a field, not whatever this is.
But what you will note is the differences not only in her outfit but also her face. She looks older in this one. You can see there's a definite similarity, but they're not the same.
(and is it me, or is she taller in the picture on the right?)

Waterfall to Countryside
Waterfall to Countryside
Waterfall link and Countryside link

Okay, let's take the so-called "Countryside" and gen a "Mountain Range" with the exact same settings. Then, finally, a "Shop" from that one (which looks like a living room to me).

To Mountain Range to Shop
To Mountain Range to Shop
Mountain Range link and Shop link

While there's a definite similarity here, it's not the same, which means that if you wanted to try different options or even to picture her from the side, the results wouldn't necessarily match up. I'm not sure if this is exactly fixable given that the AI can sometimes hallucinate things. For example, I have a picture of the girl with glasses and a necklace, which shouldn't be possible. Also one has a wonky two-and-a-half handled handbag that wasn't asked for.
Certainly, she seems happy walking on a road in the mountains, but looks absolutely knackered "in the shop living room". So let's tweak the descriptions a little to see if we can't improve things.
Let's go back to the mountain and gen another with the tags "Smiling, happy" and then gen one of the shop ones with the tags "sad, tired".

Mountain Range and Shop with emotions
Mountain Range and Shop with emotions
Mountain Range link and Shop link

I didn't ask for the handbag in the "Shop". The AI decided that unhappy girl wanted one. Maybe her debit card just said "oh hell no"?
The mountain one looks much better. She's definitely an outdoorsy girl.

Note also that by now it has given up on the two-tone sweater-over-sweater or whatever that was.

 

The final problem is that there's a definite propensity for skin. The pleated skirt is short (I was expecting sort of near the knees), and you'll note varying degrees of visible midriff which goes from subtle (as shown here) to ridiculous. And it sometimes also outputs skirts/dresses that are twice as long at the back than the front. It's as if the AI was trained to "show some skin whenever possible", which is just disturbing when it's a child. You'll notice the little girl above in the water is... yeah, pull your jumper down sweatheart...

I think they ought to have an "I like porn" tickbox. If you check that, then your humans look like they'd hang around insalubrious places offering favours for money. Otherwise, your humans will look like their idea of a nice evening is a good book, a big hot chocolate, and a favourite blankie.
Sometimes when creating non-adults, the AI will throw an error about detecting nudity. It does some nice victim blaming in asking you to revise the options. Firstly, this is likely the AI's doing and not yours because, here's the thing, why is there even the ability to apply nudity to underage not-people? The AI should be smart enough not to throw an error on detecting it, but to simply not do it at all.

 

All in all, it is impressive - and kudos to them for not only making people that pretty much look like people, but also for making it accessible to everybody. It's certainly an interesting thing to play with, though I think it's going to take plenty of creations to try to figure out what the options and possibilities are. When I said "Smiling, happy" the original prompt was "Happy, perky" but this just gave the same neutral expression. Asking for tights (or the peculiar American word) pretty much wasn't recognised. I wonder if there's a list of keywords that can be understood? Particularly as I think we're also going to be hitting a language barrier here. I mean, what would happen if I asked for a brown jumper?

For those who don't know the difference, on this side of the ocean it's another word for sweater, while over that side of the ocean it's what we'd call a pinafore dress (as in little girls) while I think what they'd call a pinafore dress has no sides or back, just straps. I learned this at an early age when an American auntie was wondering why a boy was wearing a dress... you know, once she'd worked out that "jumpah" was "jumperrr". ☺ Americans do weird things with the letter 'R', while us with RP/Estuary/Home Counties accents just don't say them at all if they follow vowels.
I tried it with one of the child pictures and it doesn't seem to be recognised.

Here's a final picture to enjoy. I took the little girl that we began with and set her body to "average" rather than "thin" as thin didn't look so good on a little one. Then I just hit "Generate another" a bunch of times.

Another unreal female child
Likes getting her feet wet, doesn't she?
(again, don't look too closely at her fingers!)
Link to imagery

 

 

Your comments:

Please note that while I check this page every so often, I am not able to control what users write; therefore I disclaim all liability for unpleasant and/or infringing and/or defamatory material. Undesired content will be removed as soon as it is noticed. By leaving a comment, you agree not to post material that is illegal or in bad taste, and you should be aware that the time and your IP address are both recorded, should it be necessary to find out who you are. Oh, and don't bother trying to inline HTML. I'm not that stupid! ☺ ADDING COMMENTS DOES NOT WORK IF READING TRANSLATED VERSIONS.
 
You can now follow comment additions with the comment RSS feed. This is distinct from the b.log RSS feed, so you can subscribe to one or both as you wish.

Gavin Wraith, 22nd November 2023, 23:07
! asked ChatGPT what three letters does "Washington" end in. It replied "Washington" ends in the three letters "g," "t," and "o." So I told it to try again. This time it said: My apologies for the mistake. "Washington" ends in the three letters "o," "n," and "g.". Some room for improvement there, I think. 
 
 
 
 
Rick, 22nd November 2023, 23:43
I fared slightly better: https://chat.openai.com/share/01ced242-1f7b-417a-a78c-294025f525e0
Anon, 23rd November 2023, 00:42
The AI-generated girl on the top left (the skateboarding one) looks like the deformed character from The Goonies. 
 
And the kid at the bottom of the post? 
 
Let's put it this way. If someone were to re-make the film Child's Play, with a female 'Chucky' character, that's what I'd expect her to look like. 
 
Honestly it's like a cross between Child's Play and Orphan. That AI image looks like a kid with a body count. A high body count. 
 
Thanks for that. Nightmares for the next week!
Gavin Wraith, 23rd November 2023, 10:34
My original question to ChatGPT was a bit harder - how many athematic Greek verbs have a root ending in \nu, \upsilon? This is not a random question. It concerns the relative influence of Anatolian languages on Greek. Its feeble answer indicated that it had no access to a digitized Greek lexicon. Perhaps it soon will.
David Pilling, 23rd November 2023, 15:28
Impressive stuff. Go back 30 years there was not much on the 'net, not enough to train AI. We've now gone through the era of content creation - cheap digital cameras, everyone typing. There's a lot of stuff. 
 
What comes next, the era of AI generated content, where the majority of the stuff on the 'net becomes regurgitative AI. I would not call fake on most of the girl images above. They'll be on Google, and copied away from their text soon. 
 
Something like Chat GPT has digested around 50% of everything ever written. A lot more than the typical person. Just doing the computation costs a significant amount of money. But there's not a lot of stuff to digest for some languages. 
 
I read this week that AI can decide based on a voice sample if you have type 2 diabetes - oddities in the voice model it has show up the condition. 
 
Rick, 23rd November 2023, 15:38
Gl^WAnon ;) This shows the difference in our attitudes here. I think the child at the end is "cute, but not quite right". As for a body count? Well, I watched Child's Play (and one of the many sequels) and I thought the premise was kind of dumb (but not in a funny way like The Toxic Avenger), and the humans were SO annoying that in the end I was rooting for the doll... 
 
So I don't see a girl with a body count. I see a computer generated image that is really quite impressive (but not quite there yet). 
Rick, 23rd November 2023, 15:43
Hey, maybe GenPhotos should offer a bunch of free humans to anybody who wants to write about them online? 
That'll get them a shedload of publicity plus an absolute tonne of images screwing up Google's/Facebook's biometrics (the ones they don't admit to). 
Tweak the AI so it returns more consistent results, they could do an "adopt a human" campaign for people with overactive imaginations and too much free time to see what sort of backstories people can come up with. 
 
I'd gen a slightly older human, and submit her to be the leader of a made up political party with a desire to become Prime Minister and "Make Britain Great Again" ;) - tell you what, she'd have more personality than the incumbents. 
David Pilling, 23rd November 2023, 22:08
Hey Rick I know what you're up to: 
 
"Meet the first Spanish AI model earning up to €10,000 per month" 
 
"They created Aitana, an exuberant 25-year-old pink-haired woman from Barcelona whose physical appearance is close to perfection. The virtual model can earn up to € 10,000 a month, according to her creator, but the average is around € 3,000." 
 
Rick, 24th November 2023, 07:31
An AI model makes between three and ten thousand a month? 
 
You know what they say about fools and their money, right? 
Anon, 24th November 2023, 09:59
Rick - I don't think I'm the only person posting as 'Anon'. But seeing as you've figured out who at least one of the 'Anon' posters are... 
 
I drove across the Huntworth Viaduct a bit back. Doughnut Island appears to still be there. 
 
How long ago did you lot blow that crater in the marshes? ;-)
David Pilling, 24th November 2023, 12:34
Think of the ethics, you own an AI person, you make them go out to work for you. Or are you them - you did the creative work, no. But you're not a pink haired exuberant 25 year old... 
 
Mostly investment is about the "bigger fool theory", so don't knock fools. 
Rick, 24th November 2023, 15:31
Yes, maybe you should pick a different name than "Anon"? May I suggest a tree-dwelling mammal? 
 
Doughnut Island must be coming on to thirty years now? Maybe a couple more? 
What's a greater surprise is that they haven't turned the place into luxury apartments or some dumb crap like that. 
 
For everybody else - I learned a lesson that day. The lesson being that if you have a big arse firework, the little stick attached goes up with the rocket, it doesn't detach. 
So having buried it nice and securely into the ground so the thing wouldn't fall over, I lit the touchpaper, saw the fire underneath, had the realisation that this was *not* where I should be standing and... I legged it while the firework went SpaceX on me. 
 
So, boys and girls, if the nanny state hasn't banned fireworks as "tools liable to be of use to a terrorist" (or some other nonsense), just remember... the stick goes up! 
Rick, 24th November 2023, 15:34
As for the "Doughnut Island", it was a big firework, that began life as...more than one. The rest ought to be obvious. 😉
A tree-dwelling mammal, 25th November 2023, 02:34
So what gave it away? A reverse lookup on the static IP address? Or the strange love of 90s country music? 
 
I doubt they'd attempt to build anything on the area where Doughnut Island is. It's all marshland and flood plain (which does flood regularly). All of the Somerset Levels drain into the Huntworth marshes, so building anything on it is a no-no. 
 
(Not that it would stop the present government or their cronies from approving planning for 'luxury apartments' of course.) 
 
And yes, I recall you telling me the entire story of Doughnut Island one night on the phone (before you moved to France and we could have 2-3 hour phone calls without having to remortgage the house to settle up the phone bill). As I remember, I'm not sure which one of us was laughing about it the most. 
 
Anyway, like I said, when you drive across the Huntworth Viaduct on the M5 and look across the marshes, you can see something that looks exactly like you described. An island with a hole in the middle, that appears to have been created with explosives. Funny, that. ;-)
A tree-dwelling mammal, 25th November 2023, 14:01
And for the record, I just took another look at that AI image of the last girl. I stand by what I said before, Bride of Chucky meets Orphan. 
 
She looks like she'd murder you in your sleep!
jgh, 26th November 2023, 08:21
Those skirts trigger my "poor mite must be *freezing*!" response. 

Add a comment (v0.11) [help?] . . . try the comment feed!
Your name
Your email (optional)
Validation Are you real? Please type 59088 backwards.
Your comment
French flagSpanish flagJapanese flag
Calendar
«   November 2023   »
MonTueWedThuFriSatSun
  123
678
14151617
202124
272930   

(Felicity? Marte? Find out!)

Last 5 entries

List all b.log entries

Return to the site index

Geekery

Search

Search Rick's b.log!

PS: Don't try to be clever.
It's a simple substring match.

Etc...

Last read at 19:41 on 2024/04/16.

QR code


Valid HTML 4.01 Transitional
Valid CSS
Valid RSS 2.0

 

© 2023 Rick Murray
This web page is licenced for your personal, private, non-commercial use only. No automated processing by advertising systems is permitted.
RIPA notice: No consent is given for interception of page transmission.

 

Have you noticed the watermarks on pictures?
Next entry - 2023/11/23
Return to top of page