Fun with AI imaging

[Dollstudio]

Hi,

nowadays, AI characters are everywhere, e.g. we see them as influencers on Youtube or as movie actors … and a couple of days ago, a manufacturer/vendor was kicked off from TDF because they somehow AI tampered their promotional images. So I iguess it's time to take a look how this technology works, what can be accomplished with it, and especially how it can be detected. In a nutshell, acquiring some media competence, like identifying image manipulation with classic methods like retouching or detecting fake journalism. Things everybody should be attentive these deys, at least to some degree, unless he/she likes to get manipulated (respectively mindfucked).

I guess I am not the only one curious about AI imaging, so please feel free to share your own thougths and experiences or your (own) AI generated images, as long as they are doll/beauty/modeling related and comply with the TDF RoC and ToS.

Since this isn't going to be a tutorial, let's just start with a (more or less random) showcase image to illustrate what AI imagery is:

midjourney-showcase.jpg (677.54 KiB) Viewed 1570 times

The (beautiful) picture above was created by 'paradox7525' and is used in the showcase section for Midjourney, one of many publicly available AI image generators. This image comes with the description:

a beautiful woman draped in silks and floating surface of water, art nouveau, in the style of Alphonse Mucha

The above description is also a so called 'prompt', meaning that - in theory - this set of terms should suffice to generate this or a similar image.

So theoretically, we can instruct an AI image generator to create images of humans, other animals, things, landscapes, objects and the like by just verbally describing them. Also, the AI image generator can mimic styles, from cinematic or photorealistic over anime/manga style, cartoon style to artistic styles. For that the AI uses a data set called the 'model', and the model is created with training. The AI can not generate an image of an penguin if it was never trained with the image of a penguin.

For me that was a bit confusing as in my understanding, a real intelligence should be able to deduce things from abstract descriptions. E.g. by looking up in a reference book where it can read that a penguin is a bird, so it might have feathers and not scales like a fish or white dots like a fly agaric. When I tried my first AI promt, I learnt what this restriction meant:

00020-1947240328.jpg (595.42 KiB) Viewed 1570 times

This image was generated with A1111, an open source blend of Stable Diffusion. The prompt I used was simple:

Family of 10 different nude sex dolls in the living room

And yes, I also used a couple of negative promt parameters like "morbid, ugly, asymmetrical, mutated malformed, mutilated". The 'negative promt' lists things we do not want to see in our generated image.

Now about the picture above - where to start… this particular AI had trouble to count. It's definitely more than 10 soll dolls. The AI believes for some reason that dolls must have detachable limbs as most dolls seem to habe detachable arms. Not all dolls are nude, and there are numerous errors included which definitely do not honor the negative prompt paramters. There is nothing to see which looks like a living room. And, obviously, the image is unusable garbage.

Why is the AI struggling so much witch such a simple prompt? My guess is:

• There are components in the prompt most AI's are instructed not to process ("nude"); and
• the training set of the model might not have included any actual sex doll.

Actually, most AI image generators have recently removed all explicit (erotic, sexual, pornographic) images from their training sets. I guess it's not a good idea to start a new technology with things it must not know or is not allowed to think.

For my second attempt I used another AI image generator called Fooocus. Same prompt, but a much more eloborate set of negative prompt paramters. Two images were generated in this batch:

3602513595863084428-01.jpg (273.07 KiB) Viewed 1570 times

3602513595863084428-02.jpg (284.02 KiB) Viewed 1570 times

By the way, all these images are unedited, unretouched and unaltered (except for the "AI generated image" reminder).

Here is the complete negative prompt for the two images above:

(worst quality, low quality, normal quality, lowres, low details, oversaturated, undersaturated, overexposed, underexposed, grayscale, bw, bad photo, bad photography, bad art:1.4), (watermark, signature, text font, username, error, logo, words, letters, digits, autograph, trademark, name:1.2), (blur, blurry, grainy), morbid, ugly, asymmetrical, mutated malformed, mutilated, poorly lit, bad shadow, draft, cropped, out of frame, cut off, censored, jpeg artifacts, out of focus, glitch, duplicate, (airbrushed, cartoon, anime, semi-realistic, cgi, render, blender, digital art, manga, amateur:1.3), (3D ,3D Game, 3D Game Scene, 3D Character:1.1), (bad hands, bad anatomy, bad body, bad face, bad teeth, bad arms, bad legs, deformities:1.3)

This is much better, even though the AI is still unable to count to 10. The dolls have still detachable arms like manikins, and the second image has an surplus arm (2nd doll from the left) and two detachable hands (both dolls on the right side). But we got a living room and the dolls look kind of OK, so these images could be a starting point for further work (assuming someone is interested in doll photography but does not have any dolls ).

So from these very basic examples we already learnt some things to watch out for: Image segments that do not look right, elements that appear distorted, and last but not least: surplus limbs. If you watch careful, you will see such manipulations everywhere in the (legacy) media. For example, the German mainstream media recently managed to publish a picture of a protesting crowd and placed a part of the crowd within a river (in this case, the Alster in Hamburg, Germany), attempting to make the crowd appear larger than it actually was. Or watch out for hands with six fingers. This was an infamous bug of some AI image generators last summer, and it showed up in countless "evicence" pictures from the Israel/Gaza conflict.

Another thing I am personally struggling with is repeatability. The AI image generators have a lot of artistic freedom for their creations. In some AI generators the amount of "artistic license" is configurable, e.g. with the CFG setting in Stable Diffusion: The CFG scale (classifier-free guidance scale) or guidance scale is "a parameter that controls how much the image generation process follows the text prompt. The higher the value, the more the image sticks to a given text input". Or in other words: It seems to be very hard to exactly replicate one generated image. To give an example: For the batch with the following two images I used exactly the prompt from the initial showcase image:

a beautiful woman draped in silks and floating surface of water, art nouveau, in the style of Alphonse Mucha

I fed this to Fooocus with the negative prompt from above, and that's what I got:

mucha1.jpg (336.9 KiB) Viewed 1564 times

mucha2.jpg (357.35 KiB) Viewed 1564 times

Both are very nice, but also very different from the initial showcase picture. Because of this I am asking myself: Does anyone really know what's going on in an AI brain if we hardly can predict the outcome of a task it is supposed to solve? If we do not know how the AI comes to its results, how do we know that its internal logic isn't flawed?

Sandro

[baphometric]

Yeah generative AI images are pretty amazing! The technology is moving at an incredible pace.

Check out my thread here of Mia's AI art: viewtopic.php?t=161116

And this thread I made on using AI to create deepfake videos with dolls: viewtopic.php?t=162489

I look forward to where things are going, and what we can do in the future. But we must also be aware that these technologies can be used to mislead consumers, as we saw with the fake AI images of dolls from that banned manufacturer. Once you've seen a lot of AI images, it gets easier to spot. Look for extra fingers, extra arms, or other oddities. Follow lines and check that the image makes logical sense - AI hasn't quite figured out how to do that well yet.

Check out Tasha Marie's stuff, her AI art is some of the best for doll owners I've seen.

Here are some of my favorite images I've made with AI of my doll Mia, using various tools like Midjourney, Stable Diffusion, DALL-E and others:

[The Doll Advocate]

1707179704764.jpg (28.8 KiB) Viewed 1530 times

I've played around with SpellAi .
I also asked it for dolls and well this is what I got

1707179308046.jpg (39.92 KiB) Viewed 1530 times

1707179570193.jpg (28.63 KiB) Viewed 1530 times

[manfredixon]

I know what you mean. I created this today and was blown away by how simple it was. If there's ever a doll that looks exactly like this, my search will finally be over

AI-generated-redhead---what-a-time-to-be-alive-1707687374892.jpg (363.91 KiB) Viewed 1500 times

[Dollstudio]

I know what you mean. I created this today and was blown away by how simple it was. If there's ever a doll that looks exactly like this, my search will finally be over.

Yes, exactly my thoughts. There are several AI characters so yummy I would buy in an instant. You can google for something like "AI characters without NSFW filter", or take a look at candy.ai, dreamgf.ai, nsfwartgenerator.ai, promptchan.ai, soulgen.net and the like (watch out, most of them charge $$$ for their services).

Doll makers, are you listening? Free suggestions for your next dolls

Here is a digital artist I can highly recommend: https://twitter.com/FXDollSusi ("All images are of non-existent people"). Though, not sure if it's pure AI.

And btw, I'd suggest to put some indication into the AI artwork, like this:

button_ai-generated-image_original.png (9.05 KiB) Viewed 1496 times

To avoid misunderstandings as - unfortunately - we can not yet 3D print those lovelies in case someone wants to order one of them

Sandro

[EPMNGUY]

I’ve been playing with making my doll look real using AI. I start with a photo, then use it as reference at about 69%, and then take the image it generates into FaceApp to adjust makeup and expression.

[EPMNGUY]

I know what you mean. I created this today and was blown away by how simple it was. If there's ever a doll that looks exactly like this, my search will finally be over

AI-generated-redhead---what-a-time-to-be-alive-1707687374892.jpg

She is gorgeous!

[Dollstudio]

Hi,

so far we have seen that current AI imge generators are well capable of producing realistic photo-like pictures of artificial personas. These images can be hard to distinguish from a 'real' photography, unless the AI (accidentally?) embeds mistakes. However, this is still in the realm of digital artistry. AI just makes this technology more accessible.

Let's take a closer look at the image modification capabilities of the current AI image generators. This wil be a couple of posts where I will concentrate on capabilities potentially relevant for the doll community.

For reference I am going to use (or abuse) a promotion photo from Gynoid Tech as their dolls are the most lifelike I am aware of. If there are odd poses or distoritions, it's most likely the AI and not the original. That's the original photo I will be using:

IMG_8029.jpg (531.95 KiB) Viewed 1412 times

This is a Model 20 aka GT-163/E with ›Mona‹ head by Gynoid Tech. The full set accompanying this picture is in these two posts: part 1 and part 2.

I will start with a very "simple" modification by applying a style:

sai-anime-08.jpg (247.28 KiB) Viewed 1410 times

Above: Fooocus, with Anime preset, active styles "Fooocus V2", "Fooocus Negative", "Fooocus Masterpiece" (all defaults) plus "SAI Anime". Using an input image (the promo photo from Gynoid Tech), applying the styles with "stop at 0.8" and "weight 0.8". Meaning something like applying the modification with an intensity of about 80%.

sai-anime-09.jpg (261.9 KiB) Viewed 1410 times

Above: Identical parameters, only applying the styles with "stop at 0.9" and "weight 0.9" (less artistic freedom for the AI).

Observations: The AI tampers with the Gynoid logo (upper left corner); applies the anime aesthetics nicely; preserves parts of the clothing (pants) but alters the skirt to a bra; keeps the general pose, but changes it anime-like; with the 0.9 parameter, the bare feet are preserved (less artistic freedom for the AI), with the 0.8 paramter (more artistic freedom for the AI) it invents shoewear and puts some mistakes into the toes/toenails. Not sure what it does to the eyes (different eye colors). The spines of the books become unreadable, that's a characteristic for AI altered images. Happens as well with number plates of cars and the like.

No cherry picking with these two results. Both are the first and only picture I told Foocus to generate with these paramters. So again, a digital artist could do the same with some spare time budget. With an AI like Foocus, it's a matter of seconds - and it's accessible to anybody with a PC. For results like that, no art school lessons required anymore. A bt scary, isn't it?

Sandro

[Dollstudio]

Hi,

let's take this one step further and try to swap parts of the picture without leaving obvious 'artefacts' like garbled letters or numbers.

For that we use a feature called Inpainting/Outpainting. The procedure is simple: Choose the picture you want to manipulate, select the Inpaint or Outpaint tool in your AI image generator and paint a mask onto the area you want to manipulate, then enter the desired operation (verbally) into the prompt.

For this example we will change the wig and hairstyle. If you are attempting this the traditional way in Photoshop or The Gimp, masking hair can be tedious work. For this example, I just entered "blue and purple hair" in the prompt. Nothing else.

Here are the results from Fooocus. Complete batch of four images, no cherry picking:

2024-02-13_03-59-27_1573.jpg (530.1 KiB) Viewed 1387 times

2024-02-13_04-51-54_2353.jpg (532.27 KiB) Viewed 1387 times

2024-02-13_05-39-55_6056.jpg (530.32 KiB) Viewed 1387 times

2024-02-13_06-27-08_6558.jpg (532.34 KiB) Viewed 1387 times

Observations: The original image is completely untampered, meaning that the manufacturer logo remains untouched, except for the masked area (in this case the hair). The original hairstyle is removes almost completely, leaving only indications like a shadow and slightly distorted earlobes (which can be fixed in another almost trivial editing step).

The problem this causes for product photography is pretty obvious: A malicious person could fake configurations that do not exist. Like in this case, Gynoid does not offer a fancy blue and purple wig with short hair. And this does not stop at the hairstyle.

Changing elements of an image is a default Inpainting/Outpainting method. Bute there are also additional and more specialized Inpainting/Outpainting methods, e.g. "Improve details" or "Modify content". With the latter method, I masked the bag with flowers and wrote into the inpaint prompt: "black kitten". That's the result:

2024-02-13_09-51-44_8243.jpg (526.76 KiB) Viewed 1350 times

Observations: The previously hidden parts from the trousers were perfectly restored; the AI invented some earrings in the pockets of the pants in an area where previously was a reflection; the flowers were seamlessly replaced with the kitten.

Even without any fine tuning, the manipulated AI result looks like a valid picture. The average viewer might not recognize that the doll never held a kitten in her hands, at least not if he does not have access to the original (untampered) photo. And again, this causes an obvious problem for the credibility of product photos. In this case, the viewer could be made believe that a doll is capable of holding a living creature like a kitten. We are now entering the realm of the possibility to intentionally misguide a customer about the capabilities of a product he is supposed to buy.

As I mentioned before, image manipulation is not new. There is a continuum from "innocent" image enhancements to removing or changing parts of the picture. These possibilities have been used before. But what is completely new is the level of accessibility. With AI tech one does not have to acquire elaborate skills in photo editing. Basically every idiot with access to the software can do whatever he or she likes. At this point, the evidentiary value of an image potentially decreased to zero, but we still tend to believe what we see. With ubiquitious AI tech around us, we need to unlearn this belief.

Sandro

[Dollstudio]

Hi,

from here it gets only worse - or more exciting, depending on your point of view.

Let's swap the face but keep everything else untouched. The original image shows Gynoids Mona head. I'll tell Fooocus to swap Mona with Gynoid's Shay head. That's this one, a distincly different sculpt:

gynoid-shay-head-3955.jpg (503.21 KiB) Viewed 1333 times

The procedure is a little less intuitive in Foocus and might be different in other AI image generators. Step by step:

• Put the image that is to be modified into the "Inpaint or Outpaint" tab in Fooocus. Choose the method "default". Put a mask on the face.
• Switch to the "Image Prompt" tab in Fooocus. Select "Face Swap". Stop at 0.7. Weight 0.75.
• Switch to "Advanced" tab and go to the sub-tab "Control". Check "Mixing Image Prompt and Inpaint".
• Hit "Generate".

Results from a batch of 4, again no cherry picking:

2024-02-13_23-29-41_7320.jpg (532.13 KiB) Viewed 1333 times

2024-02-13_21-05-58_7349.jpg (532.7 KiB) Viewed 1333 times

2024-02-13_21-53-02_5409.jpg (532.52 KiB) Viewed 1333 times

2024-02-13_22-41-02_5003.jpg (532.77 KiB) Viewed 1333 times

Observations: Body, pose and hairstyle is cleanly preserved. The generated variants include adjustments of the line of sight without introducing annoying distoritions or errors (e.g. garbled teeth). Only in two images, the lower lip could use some refining, the other two look fine to me.

For the credibility of product photography, this causes a couple of issues. For example, a malevolent person could claim the existence of unavailable configurations (that could be easily debunked), be/she he could also try use artificial images as "evidence" how a certain head (like Gynoid's Shay) looks on a certain body (like Gynoid's Model 20). Since AI tools do not honor technical properties of actual love dolls like neck diameter or head to body proportions, such pictures would me misleading as they do not prove what they claim to demonstrate. If a certain head looks good on a certain body can only be shown by putting a physical head onto a physical body.

My guess is that if we do not watch out, we will be soon flooded with fake imagery. Without naming names I suspect that multiple promotional photo sets have been published in the past months, showing AI backgrounds and the like instead of 'real' settings. Imho it would be desirable at least to disclose if promotional photos have been digitally altered with AI tools.

Sandro

[Dollstudio]

Hi,

odd results, Fooocus edition. Wanted to see if we can tamper with the skin tone…

Same source image as above, "Image Prompt" tab with promt "dark tan skin color", Stop at 0.9 and Weight 0.7.

image.jpg (255.33 KiB) Viewed 1330 times

I am not sure what happened here. For whatever reason, the AI changed the face, the pose, the pants, added a bed and removed the books, changed the hands, decreased the size of the flowers, but kept the hairstyle and the white shirt. None of this was instructed, and the only requested change ("dark tan skin color") was ignored.

Also tried another approach, with Inpainting and a mask on face, upper body and legs. Prompt: "black skin color".

Result:

black-skin.jpg (489.87 KiB) Viewed 1293 times

Fooocus used a lot of artistic liberty for this. Changed the face and the line of sight, changed the clothing, added shoes and red paint on the toes nails, also removed the flowers and the posture of the hand. On the left side of the black skirt are artefacts from the original pants (pants were not masked). From the original image, only the background, the hairstyle and a bit from the pose remains. When these AI things start going to war at some point, I hope they will be less inventive.

Again not sure why Fooocus modified so much without being instructed to do so. However, for fake product images a result like this is harmless as there is simply not enough similarity to the original to fool anyone. Though, that black lady would be a fine doll.

Sandro

[Dollstudio]

Hi,

finally, let's try a non-existing background. In traditional photo/video editing, something like this would be compositing, so again no AI is required for such manipulations, the AI just makes it more accessible.

For this image modification I used Inpainting with method "Modify Content", and I added some extra styles like 'MRE Cinematic Dynamic', 'Fooocus Cinematic'. Prompt:

girl on street in futuristic ruins of a metropolitan city, burning car on the left side, mountains in the background

Result from a batch of 1:

background.jpg (562.95 KiB) Viewed 1292 times

It is debatable, if the generated background qualifies as "futuristic ruins of a metropolitan city", it might just be a bad instructions; though, there are no mountains in the background, and the burning car is on the right side. Interesting, the AI seems not too understand directions. Would need to be investigated from which point of view the AI interprets the prompt. Except for that, I see only one error in the image (right foot). And the shoe on the left foot is (uninstructed) creative liberty, again. So the resulting image is technically quite usable and with some touchups for lighting it'd be OK.

Is the ease to use non-existing backgrounds an issue for product photography? Is it OK to promote a (physical) doll with photos that consist of 50% or more virtual/simulated content?

I am not sure, please share your ideas.

Sandro

[setevoltas]

What I'm finding Sandro is that while the doll community is quite (almost surprisingly) positive about AI image generation in connection with chat functions, it's pretty negative about image generation in connection with marketing and doll art. Brands using AI openly to replicate traditional photo shoots have gotten run out of here with pitchforks (partly for reasons that go beyond AI, but AI didn't help their cause). It seems to be comparable to previous debates about factory images versus post-processed professional shots. People don't want to be sold on images they perceive as "fake". But after the purchase, it's a whole different mindset.

Although the difference in reaction makes sense, it amuses me a little bit. Lots of people are still just finding out about Stable Diffusion and the various UIs. I don't think most people quite get yet that their chat prompt plus uploaded face image is getting plugged in by a bot to a Stable Diffusion instance using something like ReActor or Faceswaplab.

Maybe once the tech is more mainstream there will be a change of heart. At this stage, I'd warn manufacturers to steer clear of AI in any promotional imagery, and that if they're compelled to dabble in it, the images should be either marked as such or shown in separate galleries. The consumer confidence damage isn't worth it.

I'll just add that this seems to be the Western reaction. I have no idea if this is seen as a problem elsewhere.