Hi,
im new in to this hole AI thing, so i just read that is not "allowed" to use images from real persons what i understand. So does someone know a good program were i can create fictional persons and make them wet oder messy?
greetings from germany
Peter

If you are looking for a program, I recommend Automatic1111 which you can find on github. It is a bit technical to setup and it requires a graphics card with a lot of vram. It will install everything it needs to generate images using stable diffusion, but you will need to download checkpoint models separately. There are a lot of realism models on civitai.com which you can use to generate unrestricted and realistic images of fictional characters or real people. There are youtube guides on how to set it up.

Here is one image that I generated using Autmatic1111 to show how you can use weights in the prompt to specifically get the look you want. In my example I use hair color, but I also tried fictional animated princesses and video game characters. My prompt isn't very creative. Check out renfield9's posts for more creative prompts.

I do not recommend posting images which resemble copyrighted fictional characters or real people.

Thanks for this mFeelzGood. As a newcomer to AI generated images, what you've posted here is very interesting. I would suspect that Automatic1111 could be a bit too difficult for me to use at this point in time, but presumably the weightings you refer to should be equally applicable to prompts in Stable Diffusion which I'm using?

If so, please could you also share an example prompt to generate realistic very wet hair using weightings that you think would provide similar results to the images you've posted? In my own attempts, I'm finding that the engine behaves very erratically with regard to hair, mostly either showing too little wetness or for it to be too exaggerated! This happens without me altering the relevant prompt.

Yes a lot of the online tools supports using weighting. This is the same technique renfield9 used using parenthesis. Instead of using multiple parenthesis, some online tools supports using weights.

The prompt I used was something like:

woman, (close up of her face:1.3), (long wavy layer styled hair:0.3), (wet hair:1.0), (wet skin:1.0), (wet clothing:1.0), (wet face:1.0), brunette, (blue eyes:0.1), naturally beautiful face, looking at the camera, glamorous, wearing a strapless (white dress:1.1), sunny, beach background

I used a negative prompt:

cartoon, drawing, sketch, deformed, distorted, disfigured, mutated, fat, ugly, obese, chubby, fake, poor, artificial, implants, worst quality, low quality, freak, monster, grayscale


The numbers in the image, 1.0,1.1,1.2,1.3 are weights applied to the terms: (wet hair:1.0), (wet skin:1.0), (wet clothing:1.0), (wet face:1.0)

With a value of 1.0, she is not very wet even though the prompt includes several terms that try to make her wet. It isn't until I increase the weights that she starts to have wet hair. You can do the same thing by specifying (wet hair), ((wet hair)), or (((wet hair))).

The weights of the term are very important for controlling coverage amount of substances too. I plan on creating another post to explain more.

I don't use any of the online tools, but I do know of quite a few: https://www.seaart.ai/ https://lexica.art/ https://www.mage.space/ https://getimg.ai/ https://sinkin.ai/ . The base checkpoint which the website uses is very important for achieving realism. Some sites allow you to switch it. The negative prompt is also very important for realism. Sometimes it helps if you put in the prompt terms like raw photo or portrait photograph.

Many thanks for your helpful advice mFeelzGood. I know a little about weightings, but not much. I'd assumed that the standard (unseen) weighting is 1.0, but the examples in your post suggests it's actually 0.0 - is that correct?

I've been trying all manner of (()) and specific weightings as well as using words like "drenched" and "bedraggled". I'm probably overdoing my prompt wording, using a carpet bombing approach to try to impress upon the engine what I want! The same applies to my negative prompt where, despite my best efforts, plenty of the genrated images contain bodily deformities.

Going back to wet hair, for whatever reason, I think a lot of it comes down to the chosen model - for some of them their hair doesn't look very wet no matter what you do, whilst others can go too far the other way to the point of it not looking realistic.

Whilst I've seen references to checkpoints in the model summaries, at this very early stage of my exploration of AI image creation I don't know yet what they mean or what they do! I'm using SeaArt.ai and I haven't seen that allows you to alter the checkpoint, I think it just comes with the model..

Thanks again!

It does feel like the wet hair weight is actually 0.0 because the first column does not feature wet hair. It is 1.0, but what I found is that the more terms in the prompt that you include which describe the hair, the less likely it will be wet. This is because of you use a prompt like "blond hair, long curly hair, wet hair", the terms blond hair and long curly hair are strongly associated with dry hair which is why the weight needs to be bumped up to make it wet.

It is very difficult for me to get a good result. I don't know terms well to put in the prompt. There are so many different ways hair looks. I don't even know most of the names of hair styles. I don't know names for a lot of outfits either. It seems like, as you probably already discovered, a lot of experimenting is required with terms. I never even heard the term bedraggled before. Sometimes there may be a perfect word that describes what you are looking for, and when you use it it doesn't work because the model isn't trained with it. Also, how would you even describe the hair look in these photos https://umd.net/groups/emerging-hair-2 or in the profile picture of https://umd.net/herhairsgottabewet ?

I did some experiments. The first picture on the beach shows results from trying different terms instead of "wet hair". The second picture just changes the background to a bathroom. The third changes the background to a waterfall. The background also has an affect on how wet she gets. I really don't like how the hair in these images turn into black lines.

I have like 450 more archives of MPV photos which I haven't looked at yet, and 31 are cleanup scenes which involve water. I might try to include these when I try to improve the training. I still don't even know how to describe the hair in the captions

That's fascinating stuff! Thanks for going to the trouble of producing the side-by-side comparisons. I would have expected "drenched" and "soaked" to have much the same effect, but they obviously don't.

Mostly because I don't know any better at the moment, my experiments have comprised stacking up/overloading phrases in the hope that something good emerges. Although I refer to hair in several places in my overall prompt, I tend to combine the words together between the commas, such as "with soaked!!! tousled!! long blonde hair" (not sure if "!" works with SeeArt's Stable Diffusion engine) and "(((soaked:1.9 bedraggled:1.9 hair)))". I know I don't need to combine "()" with the weightings but, in my youthful (in terms of knowledge and ability) exuberance, I'm trying to throwing everything at the engine in the hope something good drops out of it!

I'm not sure where you're from, but "bedraggled" is a well understood word here in the UK. "Tousled" means much the same but is far less common. I'm using these words to try and make the wet hair look more unkempt and untidy, but some won't be interested in that of course.

Have you tried "waterlogged" as an adjective for wet hair? It's not a word we'd use to describe wet clothes or hair in the UK, but the AI engine might be happy with it!

I can see what you mean about the black lines but, as long as I don't end up with hair growing out of the woman's cheeks and/or chin (which often happens!), the images you've reproduced using "soaked" are perfectly acceptable to me because they absolutely give the impression that the hair is totally wet. Based on what you've done, I'm going to stop using "drenched" and see if it improves things.

I'm not sure all of the above is applicable to how wet clothes look, so that'll be something for another day!

Happy to try more. I tried: wet, waterlogged, bedraggled, submerged. These were all with a weight of 1.5 like (wet hair:1.5).

The bathroom background does not seem convincing. I was wondering if it would put her in a bath tub, but it didn't. There are some weird objects in the head when using submerged.

It is really interesting to see the affect some of these terms have when used individually. It also shows how creative you need to be with mixing terms in the prompt.

These are also just one stable diffusion checkpoint model. There are hundreds out there. I have a dozen. I can also produce grids comparing the different checkpoint models.

I'm interested in clothing too. I want to get better at using it. Maybe this can go into a new thread though like "AI Wet Clothing" so that it is easier for people to find. If you have suggestions for the names of outfits, I could try making grids of those too to see which terms have the best affects.

Thanks for posting the results of further tests, mFeelzGood, especially when bearing in mind the amount of time it must be taking. I believe that what you're doing here is providing invaluable reference points that anyone produces WAM AI images either already or in the future can draw upon should they wish to.

In your latest tests, it's very interesting to see the significant effect of simply using "wet" for hair in the prompt, whilst "bedraggled" looks good too. My suggestion would be to establish some core principles for the simpler topic of hair in the first instance before moving on to clothing, which is could well prove to be more complicated. Perhaps it's worth looking into the impact of combining words in prompts next to see what difference that makes? Maybe ditch the words that don't have much effect and then have a look at what happens when some of the other words are used in combination? It's entirely up to you of course as you're doing all the work!

One other thing I'd say is with reference to models, as one thing I've found is that the models not only affect the woman's appearance, as you would expect, but also the style of the clothes and the effectiveness of the wetness adjectives on those clothes and also the hair. I don't know anything about what's included in the definition of models, so I was surprised to see their impact going beyond physical appearance.

Thanks again for your fantastic efforts!

There's a pretty effective wet hair LoRA on Civitai. Won't link it but just search "Zovya's Wet Hair". I find it's better to use it quite sparingly (eg. 0.3-0.6 weight) as it seems to do something to the model's eyes, but it cuts through pretty well, and you usually get wet clothes into the bargain. Also FWIW the checkpoint called 'deliberate_v2' is pretty good at wetlook with minimal prompting, despite being a very general checkpoint. It's somewhere in between a realistic and stylised model.

All of the images from my previous reply were using chronos_v10.

The first image on this reply is using deliberate_v2 with the same prompt and seed.

The second image is using deliberate_v2 with the "Zovya's Wet Hair" LoRa with a weight of 0.6.

The third image is using chronos_v2, the term (wet hair:1.5), and "Zovya's Wet Hair" LoRa with a weight of 0.6, 0.7, 0.8, and 0.9. I used "wet hair" because that is what it said the LoRa was trained on.