Jump to content
Existing user? Sign In

Sign In



Sign Up

Playing with AI generated fetish art


Recommended Posts

On 5/4/2023 at 3:09 PM, omuai said:

Wow, you really are diaper AI pioneers in here :cat_hearts:

After some unsuccessful attempts last summer, I became interested in AI generated pictures again this year. I've had some good results using Stable Diffusion and the anime DiaperAI model available on civitai.com (a lot of thanks to whoever made this or the models that were mixed for it!) and that encouraged me to dive into creating a diaper LoRA. After more than 40 attempts I'm still mostly trying to figure stuff out, but some of the results are nice in my opinion.

Figured it would be fun to try to put a diaper on @ihcah's old favorite image. Here's my attempt. I couldn't get it any brighter, but at least the diaper somewhat fits the lighting conditions inside. I used my own diaper LoRA + AOM3a1b (by accident) and lots of img2img/inpainting. Then I added the tapes and smoothed out some weirdness and did another few rounds of inpainting.

Added an anime picture with a similar setting and a somewhat more omorashi beach picture just because. For pee, the LoRA "peeingSelfV2" was used.

fin.jpg

anime.jpg

beach.jpg

Awesome!

Any chance I could get your Lora and some of the prompts you used? Those are phenomenal! (Gonna try to train my own as well, but 1) I am a noob, and 2) 8 gb of VRAM-)
I didn't know PeeingSelfV2 really was a LoRA -- I was told to add it into model checkpoints, is that not correct?

When you do inpainting, do you just mark the parts that seem off and regenerate?

Sorry if it's a lot of questions, I would really appreciate it if you would give some advice. ^^

Link to comment
  • Replies 189
  • Created
  • Last Reply

Top Posters In This Topic

Top Posters In This Topic

Popular Posts

@Azazbel Good questions and findings (and your results are very nice)—I’m happy to have another person trying this and compare findings.    As you probably saw in one of my previous posts,

Padded (and wet padded) Taylor Swift lingerie shoot : ) I used a public LoRA for her face, and used a woman modeling lingerie as the base image for ControlNet to get the pose, then my custom pads

I worked for a while on a new one today when I got inspired thinking of an old favorite image.I'll show my process a little on this one, for fun. Here's where I started - it's a great shot from A

Posted Images

And I know that’s not true to Taylor’s figure 👀

but of course my figure-base was a lingerie model image (taking her overall shape via controlnet), then I added wider hips and other bits to perfect her, then the face as Swift, and then the pad Lora & tags. 
 

what a world we live in, where I can ask the computer for a fantasy of Taylor Swift with hips and a pad, and have it generated for me so easily 

Link to comment
22 hours ago, Jimbob19080345 said:

The faces look horrifying in the beginning stages

Well, the free online stuff still sucks a lot of arse when it comes to faces, though they're getting better. The trick is to run stuff locally and inpaint until you get something that doesn't give you nightmares and/or make you piss yourself - even though I'm sure many of us wouldn't mind the latter.

Link to comment
  • 2 weeks later...
On 5/18/2023 at 9:13 AM, HentaiKami said:

Awesome!

Any chance I could get your Lora and some of the prompts you used? Those are phenomenal! (Gonna try to train my own as well, but 1) I am a noob, and 2) 8 gb of VRAM-)
I didn't know PeeingSelfV2 really was a LoRA -- I was told to add it into model checkpoints, is that not correct?

When you do inpainting, do you just mark the parts that seem off and regenerate?

Sorry if it's a lot of questions, I would really appreciate it if you would give some advice. ^^

I uploaded a similar LoRA to civitai this evening (ID 83342, omuai diaper lora). I've trained at least 40 diaper LoRAs with different settings, but the first ones were really bad. I've also experimented a lot with various settings (learning rates, dimension, batching, etc). The one I uploaded to civitai is pretty close to the one I used for the pictures I uploaded here, sadly I can't upload the same one because the .safetensor contains too much meta data that I'm uncomfortable to share.

Doing LoRA is possible with 8GB of VRAM, I didn't have more when I made my first good LoRAs.

When I use inpainting, I try to mark the area as closely as possible, yes. There are some ControlNet addons that can make inpainting better (i.e. closer to what you might want, especially regarding anatomy if it is off), but sometimes the more random approach of not activating it is better to achieve better results after more time/generations.

The PeeingSelfV2 LoRA I have is just usable by itself. I have never had to merge anything into checkpoints.

On 5/16/2023 at 3:14 AM, ihcah said:

@omuai nice to see others getting into this hobby : ) 

I saw the civit models too, but wasn’t too interested since I’m more into the niche of “incontinence panties / pads real women wear” rather than ABDL-like products.

When you trained your own LoRA, it sounds like it went well. Did you use regularization images? In my last revisions I did, seems to help. 

I’m also curious if you used ControlNet. For my LoRA, I don’t feel that the results are great as far as overall composition if I only use prompts or image-to-image. But if I take a photo of a woman in panties, convert to a “canny” sketch and just draw a single line or modification to suggest the diaper, the model will produce lovely results. The images I posted today were based on using ControlNet with just a normal picture of a couple of Hooters girls in their shorts; I mostly used canny in ControlNet, with a slight hand edit, then my LoRA understood and created the realistic diapers. 

I’m sure there are better workflows. And the tools keep getting better. 

Yeah, we need more people that play around with this, there's just not enough diaper images! :)

I actually noticed that you were more into incontinence panties/pads after I had already sent off my post, but I didn't dare edit it because I wondered if the post would get approved later because of it. Sorry for the confusion :)

Regularization images have never worked well for me. In the beginning I had one instance where I thought it improved things, but my first LoRAs were generally pretty bad so it might have just been better by random chance. I've read a lot about creating LoRAs, and the common wisdom seems to be that regularization is not worth the effort or actually bad, and considering that LoRA actually add some layers to the matrix, I wonder if that isn't to be expected because to get the same images, those layers would have to revert to go inactive again, maybe?

I like ControlNet, there's many great features to play around with. It totally depends on what I am going for, though. I think it is great for getting the background right or to do certain poses, but it's difficult to get the bone length right with openpose and anime characters. I'd say I sometimes use ControlNet canny or something else to get the background/surroundings right, then I might img2img, and then I alternate inpainting with SD-webui and with my drawing tablet and Krita because it is so much easier to guide the AI to fix certain areas if you just try to draw what you want and refine it with AI.

Edited by omuai (see edit history)
Link to comment
4 hours ago, omuai said:

The one I uploaded to civitai is pretty close to the one I used for the pictures I uploaded here, sadly I can't upload the same one because the .safetensor contains too much meta data that I'm uncomfortable to share.

Interesting, I’ve never shared models so I haven’t looked into the contents of the safetensors, are you saying that it preserves some info beyond the weights? 

Link to comment
On 6/3/2023 at 10:52 PM, ihcah said:

Interesting, I’ve never shared models so I haven’t looked into the contents of the safetensors, are you saying that it preserves some info beyond the weights? 

As far as I know, various scripts that float around the net to train LoRAs have an option to save some data about the training into the metadata portion of the .safetensors. It's enabled by default for some of them. Removing some info changes the hash of the file, so it is not trivial to do so without a good explanation of what portion of the file is being hashed. It obviously can't be a hash over all data, as that would include the hash inside the metadata itself, which is generally not doable. Maybe it's all but the hash itself but I didn't bother to find out after a few failed attempts to find the information somewhere.

When I checked the info, it had a creation time and training finishing time, and that lets you guess the hardware that was used. Image tags are counted, so if you ask deepbooru and a few mostly unique tags fall out, people may be able to make good guesses about which images have been used. For me, the timestamp + potential hardware info were the main reason to not want to share metadata.

Link to comment
52 minutes ago, Peena said:

Yeah if I had anything better than my laptop and ANY knowledge of the process I’d make one of her. But I don’t 😭

I feel ya' there. Not only is this AI stuff complicated as heck to get decent at- it's also super expensive. Well, one COULD rent a better GPU (cloud computing), but eh, 'dunno how safe that is/ anonymous it is, so I won't mess with it for now. 😅

Link to comment
On 4/17/2023 at 10:47 PM, ihcah said:

This one is from a paid service online, sort of exploiting their demo a bit by rotating my VPN / browser id so I can keep submitting text to it --for that reason I don't want to print their name in this thread, since they're a newer company that might google themselves regularly, ha (and this use is definitely against their TOS). I'll DM the name.

There are some offline models but so far I haven't found any quite as good... but if anyone finds one, let us know. Sooner or later the open source models will beat the paid services for voice just as they have for everything else.

Can I get a DM also?

Link to comment
On 5/19/2023 at 4:19 PM, ihcah said:

Padded (and wet padded) Taylor Swift lingerie shoot : )

I used a public LoRA for her face, and used a woman modeling lingerie as the base image for ControlNet to get the pose, then my custom pads LoRA. I'm getting better at figuring out how to use this LoRA effectively, fortunately.

tsf1.jpg(2).thumb.png.0bcbd5196e7ab0291f7bfeb55e4a6853.png

tsff3.thumb.png.1fab623184b61be8b90447ba4182ceb8.pngtsff1.thumb.png.00ef401cabe0d13af54d45b3b64424cc.pngtsf1.jpg(1).thumb.png.e97282c118ca4915b5b437d54ba57294.pngtsf1.jpg.thumb.png.ed53ad3643f7c89ad9ccf174658a70d2.png

These are so hot!!!!

 

Link to comment
  • 1 month later...

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.
Note: Your post will require moderator approval before it will be visible.

Guest
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.

  • Recently Browsing   0 members

    • No registered users viewing this page.

×
×
  • Create New...