...I've been on an involuntary complete hiatus from SD for a few months after my old gpu and psu broke down.

Now I'm finally getting back to it with a brand new humble little 3060 with 12gb vram. I know it's not much, but compared to what I had before it's a completely different world for me now. Went from 2-3 sec/it to 4-5 it/sec with SD1.5

I have collected and/or created 180 images for the dataset so far and I feel it is time to try training again...

About 55-60 of those images are from various artists (including a few well made manips). The rest I've either made with some hentai -ish checkpoints (like the first pic) or with some semi realistic cartoon models like the second to fifth samples. While in the five posted pics all characters lay on their back, the dataset includes many different poses, different ethnicities, hairstyles, clothes, age and so on. Also I've tried to have a moderately limited variety of the used weapons (although I now suspect that I should have gone for even fewer different types of bladed weapons)

I've been out of the loop when it comes to the newer lora training methods and new SD-developments in general. So any infos or tips that might be usefull would be greatly appreciated.

For now I have mainly these questions:

Is there a better program for local training than kohya ss gui? (I am so far shying away from collab and the likes, since I don't like the idea of uploading my dataset to a google server. But maybe I'm just misunderstanding that aspect of that tool so far?)

one of the greatest problems I've encountered with the results of my first training runs half a year ago were the fact that despite extensive tagging the loras would tend to replicate backgrounds, hairstyles, faces and even the general artstyle of the majority of the dataset long before the desired concept was even vaguely present. I know this is in part due to the very small size and quality of the dataset back then, but especially the style-bleed seemed to become greater the more images the dataset contained. So my question; is there a way to "force" the training to focus on the relevant part of the image other than just cropping? Because I've tried that before and then all outputs where cropped, too, and it wasn't even useable for img2img. Could training with a negative prompt help? If so, how? Is there a way to diminish the training of an artstyle other than a greater variety in styles? Would any of you more successful lora - trainers be so kind and caption one or more of the images I've posted the way you would for your datasets?

I'm sorry to bother anyone with this little wall of text, but it's not like I could expect help for this kind of lora on reddit or civit, I think.

Also if anyone here has any similar images and would like to contribute please feel free to post them here.

Constructive critique for the pics is also welcome, of course :)
I've never published any images before, so please be gentle xD ^^

Ooops! Can't have a sword related thread without a Katana now, can we?

Oh and @mods: I hope these artstyles are acceptable?

>I hope these artstyles are acceptable
yes
anything that clearly can't be confused for a real picture at the first glance should be good (for the record, this only applies on /ai/ and not on /3dcg/, and is for legal reasons)

Ah, damn it, I grabbed that last image from the wrong folder. Avert your eyes, don't look at that hand! >.<

>>1544

Thank you for that clarification! :)

>>1539
>180 images
i think that is more than enough to start training.

A lot fo epople here made loRAs with just 20 images and i love how they work.

I once made a LoRA with 1000 images and honestly it worked pretty meh. So please, start training. i would LOVE to see your pussy stab lora.


BTW you can train a model on civitai. You can use a burner account if you are afraid. I neveer had any porblem. I can pass you the credits necessary to do a training.

PS. I do not have pussy stabbing images, but i have generic stabbing images
https://files.catbox.moe/nibfth.zip

>>1548
or if yo uprefer you can just share the dataset and I will train it.

PS. I love the models you produced so far

>>1548 >>1549

>A lot fo epople here made loRAs with just 20 images and i love how they work

Really? I've been living under the impression I'd need at least a few hundred pics to get workable results. (some people even told me I'd need at least 400) >.<

>PS. I do not have pussy stabbing images, but i have generic stabbing images

Thanks, these are great, too! :)

>BTW you can train a model on civitai.

Yeah I've not yet checked their training tools out really. I uninformedly assumed it would be basicly the same deal as with colab, to be honest.

>or if yo uprefer you can just share the dataset and I will train it.

that would be awesome! :D

Do you have a Pixiv or another way for me to give you a link in private? (or are you the guy who started the reentry, by chance? If so we may have already had contact on pixiv.)

>PS. I love the models you produced so far

Thank you!

P.S. if anyone here can read chinese could you please tell me if the plaque roughly says what it's supposed to say or if I got trolled by google translate, again? xD

>>1565
>the guy who started the reentry
Yeah, I'm that guy. so just hit me on pixiv.

>>1565
aweome keep em cumming:) great work!

>>1566

Just sent you a pn :)

>>1567

>aweome keep em cumming:) great work!

Thanks! Allright, one more for now :)

>>1818
Tried it with the girl walking but it didn't seem to like that. So standing it is.

>>1824

Ooops, didn't realise you had already switched over to here...

These are amazing! Especially >>1827. That's exactly what I was always hoping for! :)

Hmmm, walking eh? I didn't even think about that pose/concept when creating the pics for the training. :o

Also I made almost no fully frontal posed pics like >>1828 because due to the perspective the blade then would always look just like a weird thin line. I could (/should) have done some pics like >>1825, but one of my main goals was to always get the swordblade to penetrate straight in and sitting upright. But it's really nice to see that the lora can be so adaptive! :D

I bet all that could be further improved with more images(?)

These are absolutely gorgeous! Thank you for sharing, my friend.

I may not yet be abled to use that awesome lora (>>1800), but I can't let this wonderfull day pass without contributing at least a little bit myself... I think there'll be some more tomorrow. Been working on some stuff...

mixing with piss drinking didn't give the best results. These two are the best out of the bunch.

>>1844, >>1845, >>1846, >>1847

These are lovely, many thanks! :D

Any idea what may cause the piss drinking lora to struggle with the pussy stabbing one?

Also thanks to Entropy and little luck I can finally use ponySD! So here's my first nice result: :D

...just amazing! Knives seem to work best for now. Only have experimented for maybe half an hour with pony running and then 10 minutes to narrow down on that prompt and the right setting for Adetailer. No inpainting, just upscaled and used Adetailer for that last one. (I'll try some inpainting later though)

For comparison this next one here I've just finished prior to getting pony to run on my system and it took about 5 or 6 hours with my workflow and hardware.

The next one also took quite some time to finish, but I believe the two or three evenings spend were well worth it!

Since I really really like this one I'll upload all three versions; before, short- or broken sword and all-the-way-through. Hope someone will enjoy! :)

...aaaand no. 3.

Because I can run Pony now I got a little distracted from giving some minor finishing touches to that little project I hinted at yesterday, but I'll upload that in a few hours if all goes well. :)

>>1854
Well judging as how swords would be coming out of the penis instead of urine, it would be that it thinks that that is where the stabbing should happen.

>>1859

Ouch! >.<

...do you think it could possibly help to retrain including more pics wherein penisses are present, that are explicitly NOT pissing swordblades?

Finally finished this:

When SD gave me this pic I immediatly knew I'd have to make something nicer than just a single pic with her...

So I changed her appearance a bit to my liking and after finishing the usual inserted-sword-pic for the training data I started to make a whole little story thingy with her.

Is this a comic strip? I really don't know what the proper term for this would be.

In case someone is interested, I've packaged the original sized images and uploaded them here:

https://files.catbox.moe/q4ecui.rar

(170mb)

Password is

Aloha

...I know the character likeness isn't perfect and I'm not too happy about the 4th frame, but at least it conveys what I wanted it to, I think.

Again, I just hope somebody else can get some enjoyment out of it. ^^

(P.S. I had to severly lower not only the size but also the quality to uploaad this pic here. So please consider this only as a preview for what inside the .rar)

>>1862

Wow, you made this with AI? It's phenomenal work. I can't even get any sort of consistency from one pic to another.

absolutely stunning work, thanks for these! can you make the girls masturbating with knives & smiling as they gut themselves?

>>1884
Thank you! :)

With some SD1.5 checkpoints you can achieve at least some degree of consistency for char likeness (at least concerning facial features) by just giving them a random name. Just put something like: "(Irene Clara:1.1)" somewhere within the descriptions of your desired subject.
Sadly as far as I can tell this "feature" that worked very well with older base models has become less reliable with newer models and I don't know if it still works with sdxl/pony at all. My guess is that the closer the model is to the SD1.5 base model the better it works. ...but if it does work with the model you're using it can even help to control your subjects ethnicity and skincolor to some extent (Although not very reliable when using clip skip 2 or higher). A handy tool I'm using is this random name generator: https://www.behindthename.com/random/
Sadly this little SD - gimmick doesn't help that much with the hairstyles so that's been the first thing I had to manipulate and inpaint for almost every pic from that comic thingy.

Paint.net is the 2nd tool I'm using to make all these pics. Its a nice little free alternative to photoshop, etc. Most importantly it can do layers. For example I just sketch some altered details or simply copy-paste some elements (like the orchid in her hair) and then let inpaint have a go at it to nicely blend it in with the rest of the image. I know a1111 also has an inpaint sketch function but for some reason my system almost freezes when I try to use it so I've just stuck to what has worked best for me so far.

Overall I could not have done any of these pics without using some photoshop-esque techniques and inpaint.

For example no matter what I prompted, SD refused to just give me a decent rearview of the girl walking out of her hut with just some jungle background outside. But after a few hundred failed gens I at least had enough to composite something out of them. Took the girl I liked most, removed the shitty background (just with what is available in the "extra" tab in a1111), took two halves of hut interior I liked, a doorway and then grabbed 3 or 4 snippets of jungle for the background. This also made it easy to have the colors of the background washed out to get that sort of blinding effect - also something I couldn't get the AI to do just by prompting in this case. Once the composite was done I just had to use inpainting along the seams of the snippets again.

For the sword stuff it is basicly the same deal. Got some matching pics of the weapon of choice in roughly the correct angles for the subjects poses and camera angles, inserted that and then let inpaint do its magic. In this case with controlnet, either canny (for more finely detailed swords) or softedge (when the weapon is less detailed and/or closer to the viewer and thus bigger).

>>1886

Sure, happy and smiling I can easily do now thanks to the lora. But holding the knife herself may be a bit more tricky. I think it is possible though because I had some such gens randomly appear in my testing since yesterday. But I haven't figured out how to prompt for it yet.

When you say "gut themselves" do you literally mean with guts coming out of the pussy? I'd have to experiment with that.

Maybe I'll make something like that with my older non-pony methods I just described in that last post. Probably a good idea to have some more happy consensual and/or self inflicted examples for the training data for a potential retraining.

For now I hope this happy little lady at least partially scratches that itch for you. :)

(That Lora is so good at making knives, I still can't believe it!)

>>1886

Oh, just got one where she at least kind of has her hand on it. Altough her aim is a bit off.

love these, can a girl be cutting herself open, pussy to sternum? the smiling happy knife holster girls are awesome!!!

Awesome stuff. A long shot but can someone also make a butthole rapier/arrow/etc stabbings lora as well?

>>1997

Patience please! I will eventually make something like that. ...Been busy irl and also there's just so much new stuff popping up that I needed to test out the last few days, that I just haven't had the time yet.

So until I get around to it I hope this one will suffice. I know she hasn't gutted herself yet, but she at least has better aim and she's definitely enjoying it!... :)

...and even though this one here is not happy about it and she isn't sliced open, there are at least guts present... somehow... even though I didn't order any... >.<

I was just trying to combine the quad amputee Lora and the pussystabbing one. But instead at first I got all kinds of AI weirdness, mostly hybrids between deformed remnants of knifeblades and weird tumor looking fleshballs but also this one fine looking pic.

...could almost be the aftermath for...

...this one. If it weren't for the different background and stump length.

And now to showcase what you can do with the anytest controlnet model: This...

...turned into this in less than an hour! (10 minutes to get the weights right and have an initially good gen, another 10 minutes for hires + adetailer on a good seed and the rest of the time was needed to correct some details. and it took only so long because my computer is shit. I'm getting aroung 4-5 sec/it with pony compared to 5-6 it/sec with sd1.5.)

I really have to thank you anon from >>2087! Anytest is amazing! A complete gamechanger!

>>2000

almost forgot; >>2117 was made with a combination of the pussystabbing Lora and an ordinary dildo masturbation Lora from civit. There are at least 2 of those for pony, only tested with one so far. Of course it's rather hit and miss but since the dildo Lora I used had the triggerwords "vaginal" and "anal" I'd assume that it may be possible to get some anal stabbings done with it. I also very briefly tested combining the arrow in body -Lora (>>1604) and the pussystabbing lora even before you asked here. I just wanted the girls to have arrows in their chest/belly as well as a sword or knife in the pussy but instead it gave me weird metallic arrow thingies in their cunts. So again I bet with some prompt weight experimenting and then inpainting I'm sure you should be abled to come up with something like what you're after.

I'll be working on an improved dataset for the Lora. If you have enough pics relevant to your request and if you can upload them somewhere I'll see if they can be incorporated in the training. :)

(<=== an older one from the Lora trainingdata done with SD1.5. With a sword that size in there I'm pretty certain her pussy and asshole have become one. Also bonus whipmarks on her ass, so I hope you like her. ^^)
_____


In general for requests: if you can come up with a crude sketch like >>2120 it would make it much easier and less time consuming for me and you'll get something closer to what you desire. I'll be happy to give things a try if I've got the time and if I like the subject/setting.

This one initially started out as just a quick manip of someone else's sketch drawing in order to get more images for the Lora training. But I had unintentionally set the denoise for the anytest-pass so high that it gave me nice looking backgrounds. This one to me had too much potential to just throw away, so I made this.

SOmeone released a similar modle on civitai

I have not yet tried it

https://civitai.com/models/545455?modelVersionId=606601

>>2433

Nice find. But sadly it doesn't appear to work very well. I've genned a little under 100 pics, using just the example prompt but only two pics so far came out with the blade inserted. Everything else is just the hilt/grip inserted with the blade on the outside (which is something SD always could do).

But maybe I'm doing something wrong. Did you have more success? :)

>>2399
The blonde loli top right looks amazing! Would love to see more scenes with her - anal sex, rape and/or guro <3

>>2486

Yeah, she's cute! :3 ...I think I'll do something multipanel-y with her. (well, 95% her. Didn't use a Lora for her so I can't guarantee to exactly get her 100% like in >>2399 again.) I've already got a few other single image- works in progress, as well as one really long- and one shorter multipanel story also currently in the works, so it may take a while until I get to her eventually. So please be patient. :)

But in the meantime: (<===) the girl from >>1862 had a little sister... (Older work made with SD1.5, now slightly overhauled with pony) ...not the same kind of cute, but also cute, I hope... :)

...And now for something completely different...

I was just making some simple txt2img gens for this character because I need a few NON-guro pics of her in a few different outfits for one of the multipanel stories I just mentioned. This was the 3rd outfit, 10th (and last) pic with that prompt. Nowhere in the prompt or neg was the word GURO (in fact "text" was in the neg) and I didn't use any controlnet, outfit- or guro- Loras or other shennanigans that could have intentionally resulted in this.

What is happening here?! :o

...This must be a sign from the AI-gods! Rejoice, for they smile upon our works! xD

...in case someone doesn't believe me and wants to replicate it, here is the raw non- adetailer, non- highres/multidiffusion/upscale image with prompt etc. in the meta.

(she'll eventually get her pussy stabbed, too, (amongst other things) so you can also consider this a sneak peak...)

>>2519
love this one. Great hopes to see what happens

>>2517

What is pony exactly? I couldn't really figure it out except that it seems to be some format or something for newer models that are higher res. Nobody every explains shit in ai world.

>>2636
Pony-diffusion-v6-xl (often abbreviated "Pony" is a Stable Diffusion base model that has become one of the, if not the most used consumer level model since it was released in january. The majority of current base models and Loras have been trained/finetuned on/for Pony. AFAIK the creator initially just wanted to create an SD-model for some "my little pony" enthusiasts and furry enjoyers but the model was also trained on "normal" data (anime/hentai/cartoon etc.) and it was so superior to "normal" StableDiffusionXL (and it's derivatives) -the back then most developed version of stable diffusion-, that it quickly became the basis for a huge amount (if not the vast majority) of community driven developments. (https://civitai.com/models/257749/pony-diffusion-v6-xl)

The SD1.4/SD1.5 -models that first came out in 2022 and all the stuff that derived from it was trained on a resolution of 512x512 pixels. Then came SD2.0 which was basicly dead on arrival, mostly due to a censorship related mistake when it was trained by StabilityAI (The developer company behind baseline Stable Diffusion). Though it was supposed to be better than SD1.5 because it had been trained on a resolution of 768x768 pixels on a larger dataset, its output was shittier, because a nsfw filter slider was set to a wrong value during training, thus fucking up the quality of everything anatomy- related. Then StabilityAI made SDXL with a base resolution of 1024x1024 but since we live in the age of enshittification its quality again was in many ways lower than expected. Again greatly in part thanks to censorship. Then a guy named PurpleSmartAi trained pony on the SDXL basis and now that is what most people use. Just a few weeks ago the greatly anticipated SD3.0 was released but as of now it seems very likely to be a case of "dead on arrival" again, so for the foreseeable future it looks like pony will be around a little while longer.

>>1997 Hope you'll enjoy:

1st I had to remake >>2117 with MultiDiffusion...

...then I tried to fullfil your request. Is this at least a little bit kind of what you've had in mind? I also tried having some entrails hanging out, but it didn't look good at all imho.

>>2639
>>2638
Neat high effort pieces

Multidiffusion looks cleaner especially on nipples and pussy, 0.33 denoise?
Remember to tick mixture of diffusers if you aren't.
Good job though, very nice

>>2637

Oh, I should have mentioned, that models/Loras/Controlnet-models and some other things are made with/for their respective base models and therefore not compatible to one another. So for example Loras trained on SD1.5 cannot be used in SD2.0, SDXL or Pony and vice versa etc. That's why Pony even has it's own filter category on civitai.

>>2640

Thank you, much appreciated! :)

>0.33 denoise?

yup, 0.33 iirc. Did you take a guess or do you have THAT good of an eye for that? :o I think I ticked 'mixed', but now that you mention it I'm not a hundred percent sure anymore. Is there something in the pic that leads you to believe I may not have had it ticked? :)

>looks cleaner especially on nipples
yes it does, but also I'm not even sure if I even used breast- or nipple- Adetailer, let alone inpainting on the nipples in the original pic, to be honest. Still, even I can see that the overall quality is improved, despite me usually beeing rather blind to this sort of thing. (So again many thanks for showing me the light! :D )

On a related note, by doing this pic I came to notice, that apparently Adetailer is working far worse in Forge in some cases for me. At least concerning hands. I tried for hours but no matter what I tried, the output never came even close to the results from a1111. (neither with the original settings used to create >>2117, nor with changed values, settings, prompts. I know I can relatively easily fix them with inpaint, so its just a convenience thing. But if you happen to know a way to fix it I'd love to hear it. :)

>>2650

>I think I ticked 'mixed', but now that you mention it I'm not a hundred percent sure anymore. Is there something in the pic that leads you to believe I may not have had it ticked?

nevermind, i went back and checked the meta on the initial multi diffused gen. And yes, 'mixed' was ticked. (Phew!)

>>2650
>yup, 0.33 iirc. Did you take a guess or do you have THAT good of an eye for that?
It was the guide's default, but I also a guess given the almost-perfect but still imperfect quality on certain parts when zoomed in.
0.33 is a good stable value. For max detail 0.4-0.5 is the spot.
The problem with 0.4-0.5 is it almost always creates random errors.
I get around this by plotting 3 or more gens (x/y/z plot -> denoise at 0.5,0.3,0.2) and then compositing: meaning I soft crop the good areas on the error ridden high denoise gens by erasing a layer on top of another.
You typically want this on high attention areas like face, and also small sections where it omits detail, can be nipples, fingernails, labia-clit, a small object, stuff like that.
This is more an OCD perfectionist thing. I don't think 99% of people will scrutinize your work pixel by pixel.

>Is there something in the pic that leads you to believe I may not have had it ticked? :)
No, just making sure since it's a new tool for you, Mutltidiffusers without mixture of diffusers is just slightly worse

>Still, even I can see that the overall quality is improved, despite me usually beeing rather blind to this sort of thing. (So again many thanks for showing me the light! :D )
You're very welcome

>On a related note, by doing this pic I came to notice, that apparently Adetailer is working far worse in Forge in some cases for me. At least concerning hands. I tried for hours but no matter what I tried, the output never came even close to the results from a1111.

I don't think its behavior should be any different on forge. Forge handles loras differently but the tools should be exactly the same.
Adetailer hands was always shit for me, even on 1.5

Personally I don't use adetailer anymore, face or hands. I just let multidiffusers enhance face and hand detail automatically. It does a good enough job and being a separate part of the process I don't need to worry about adetailer screwing up the gen, something that can always happen no matter the model

>>2667

>The problem with 0.4-0.5 is it almost always creates random errors.

Yeah, I get stuff like a second nose on the forehead and in some cases deformed mylittlepony-esque faces when I try anything above 0.35ish. Also very common are random dwarf versions of the char scattered around the pic.

>I get around this by plotting 3 or more gens

Well, that's the next noobishness I'll have to admit to; I've tried making x/y/z plots a while ago (allthough way back when I had to work with a 1070) but even though as far as I can tell I followed the guide(s) to the 't' I didn't manage to get any plot gens to work. I just accepted defeat back then and have not tried again so far. I'll have to look into it again.

>and then compositing: meaning I soft crop the good areas on the error ridden high denoise gens by erasing a layer on top of another.

yes, I basicly do the same in some cases, just with separate gens instead of a nice convenient plot and thus most likely not as extensive as could be.

>This is more an OCD perfectionist thing.

Well even though I may not have a well trained eye for it (and standards always differ from person to person) I too would still consider myself to be somewhere in the OCD-perfectionist category. So wherever I can improve (for reasonable costs and returns) I'm happy to do so. :)

>I don't think its behavior should be any different on forge. Forge handles loras differently but the tools should be exactly the same.

In theory yes, but according to some discussions on the git page of adetailer there also have to be at least some minor differences between the way weights are applied in forge vs a1111 when using adetailer. Adetailer -hands and even more so -feet worked relatively great for me with the right prompts (I'd say for around 80-90% of gens in SD1.5 and even in pony with a1111 I got somewhere around 60-70% straight useable results.)

But like I said, this is mostly just a convenience issue and at least faces still work good enough for me, as long as I keep the weights down.

> I just let multidiffusers enhance face and hand detail automatically. It does a good enough job and being a separate part of the process I don't need to worry about adetailer screwing up the gen

I'll have to experiment and see how well that'll work for me. More reasons to get x/y/z -plots working, I guess.

Allthough no matter how that'll work out, I dread what will happen when I finally get back to work on my biggest project yet, again. I've got a rather large multipanel story in progress that I've beeen working on for months prior to my old gpu breaking down. I'm around 2/3rds done with it but I'll have to finish it with SD1.5 and the older methods I used, not just in order to keep the style consistent but also because the whole project depends on some Loras only available for 1.5. If adetailer doesn't even work with those models like it used to I'll be in inpaint-hell, as you might put it. xD

>>2670
It's easy, just add this to your multidiffusion process
Batch count on 1, or it'll do this 3 more times per batch count.
Denoise setting above will be ignored

>>2678

Thanks! Works like a charm! :)

(>>2695, >>2696)

>>2697
That's it man, flawless work. A+ job

>>2698

Thanks dude! :) Now if I could finally stop forgetting to change the 'resize to' values when I pick a pic from the right while inpainting I might even take less than 2 days for one of these... >.< (whenever I do forget and let it inpaint on full resolution it takes forever and of course results are mostly shit. Then when I realise it's running full res and I try to interrupt forge likes to freeze on me for 5 - 10 minutes. At least that way it is very punishing, so I should hopefully learn soon. xD)

I'm currently working on expanding the dataset for the Lora. I've set the arbitrary goal of 42 images per bladed weapon type that (in an ideal world) I want the Lora(s) to be capable of reproducing. 42 because its a nice number and also because the dataset that was used to create the pussystab_v1 lora contained exactly 42 pics of knives and imo knives so far work the best. So out of 8 planned weapon types I've got at least this category already covered...

...I plan on also having the following types, all as stand-alone-loras:

-classic medieval swords (only 3 images to go),
-viking/early medieval/late antiquity swords (shorter but wider blades with a stubby guard, 25 pics to go),
-roman gladius/generic ancient swords (the peppermill, 25 pics to go)
-katanas (10 pics to go, no problem)
-sabres (16 pics to go, arrr!)
-rapiers (I'll only try to get the less complicated cuphilt-rapiers to work, 29 pics to go)
-shamshir/scimitar - style arab/persian/etc. (least likely to be completed obviously, more than 35 pics still needed)...

...So just in case anyone else is even interested in these Loras becoming a thing, I'd be really happy over any useable pics anyone may want to contribute, especially for the categories that are missing the most pics as of now. :)

how hard would it be to make a somewhat more generic stabbing lora? even with all the available (pony) lora on the rentry, i am having a hard time to have a simple scene of 2 characters where one is stabbing/slashing the other. even with generous inpainting it's a huge pain.
sorry if that's a bit offtopic, didnt think it warranted its own thread.

>>2750

That's a good question to which I don't have a definitive answer. But from what I understand there are multiple problems that will make it hard to train a working Lora for generic stabbing. You need to either focus on a single type of weapon or a specific variety of weapons (like I'm trying to do for just pussy stabbing), otherwise at least the hilt/handle if not the whole weapon will in most cases be a melty unrecogniseable blob of AI weirdness. And therefore for each kind of weapon you will need a certain amount of pics in the training data. (at least 20) Then you'd need a comparable amount of pics with all kinds of camera-angles/poses (unless you want one of those Loras that can only do one angle/pose), a bunch of pics to cover for stabbing or slashing different bodyparts and for whatever else interaction between the characters you want (for example holding the stabbed subjects hair from behind vs. holding the stabbed subjects arms from behind). I think there would be even more things you'd need to consider and therefore cover in the dataset, but all this is just laymans speculation on my part.

>>2752
thank you for your detailed answer. i suppose i could curate a dataset of just a gut stab or so to try it out.
i have trained styles using Onetrainer, but never specific concepts so I'm not sure how to go about that.

>>2762

As far as I know there shouldn't be too much difference between style- and concept- training. Instead of using a triggerword for your style (or leaving the Lora completely triggerless) you'd have to specify a triggerword for the action ("gutstab" perhaps?) and the kind of weapon you end up using. In case you didn't have to worry about tagging at all so far, this guide may be helpfull: https://civitai.com/models/351583/sdxl-pony-fast-training-guide ...also some of the settings should hopefully apply to Onetrainer, too. But again, all I've said is laymans speculation at this point. I may have curated/created the images for the pussystab_v1 Lora's dataset, but it was Entropy who was so kind to tagg and train the Lora with it. He also gave me the link to that guide. But I'm still a few weeks away from having enough data for me to try to train a _v2 (multiple _v2's) on my own. So hopefully someone with greater experience in creating Loras could correct what I've said so far or add relevant info I've missed.

EeL. I love you work. Very excellent work. Do u publish work on different sites.
Do you have any more pics coming

>>3064

Thanks, much appreciated! :)

I had originally planned to post stuff on pixiv, but their new ToS has me second guessing that. I was going to wait a little now that these new rules apply over there and depending on how things develop I might start posting there. Not sure if there are other more relaxed sites where my stuff would fit in.

But in general there'll be more stuff from me, don't worry. :) I'm just currently working on a large project for the /fur ai-thread as well as the extended dataset for a v2 of the pussy stabbing lora.

I still have some unpublished pics ready to upload from the Lora - dataset. Originally I was planning to upload most of them on pixiv or something similar once I was certain where I'd end up. But if you tell me which of my pics here in this thread you like most, I'll be happy to take a look and see if I've got something similar already waiting to be uploaded. :)

Do you have any with red heads. Do you deal with body piles, or girl on top of another girl

>>3069

Surprisingly I have only a few red heads (that I'm comfortable to upload). Really thought I had done more. There are plenty redheads in the dataset in general, but most of them are by other artists. But there's at least this generic redhead here I made with SD 1.5...

>>3069

...as well as this Jessica Albert pic here, also made long ago with SD1.5. Not at all gorey but I hope you'll like it anyways.

I have tried to do body piles/2girls/3girls on top of one another in the past but so far I didn't get any great looking results. But that's something on the bucketlist sometime when my backlog has been reduced a little.

Thanks for the redheads. Asking for my favorites. That's a different choice. So many great ones. But if I had to choice. I like girls with eyes closed.

These two of my favorites. https://boards.guro.cx/ai/src/1720061187876.jpg
Can't wait for more from her. And this series. https://boards.guro.cx/ai/src/1717975305967.jpg
I want to play with her boobs all the time. Thanks

>>3074

>And this series. https://boards.guro.cx/ai/src/1717975305967.jpg
I want to play with her boobs all the time.

Yeah I friggin' love this exact kind of extremely pushed up corset/bustier tittays, too! Was really proud of that one, for sure! :) And especially with ebony girls, its just lovely I think!...

>>3074

>https://boards.guro.cx/ai/src/1720061187876.jpg
Can't wait for more from her.

I'm with you on that one, but we'll have to be very patient for more from her, I'm afraid. Once I've finished the project for /fur that I mentioned I will focus on the dataset for the v2 lora for a few weeks until I can try to train at least 2 or 3 different weapon-versions. When that is done I'll work on yet another project that I've started waaaay back. That project is the largest I've started so far but it was interrupted when my old gpu died. It's come along too far for me to abandon it. (Since there is a coincidental connection between it and the project I'm doing right now for /fur I'll most likely elaborate a little on that project when I upload the /fur-thing there. So if you're interested you might want to check that out in a few days or so).

Then finally I'll start the project with the girl from >>2518. I already have a complete handscribbled storyboard for what will happen to her. But I already know that it'll be a few months before I can start on it. But since we're talking about nice plongeant decolletes I hope this teaser of her will please you. ^^

>>3072

Sry, no closed eyes for now, but I'll keep it in mind. ;)

Just seeing if you are still alive. Do you have anymore religious one. Like the Amish. Or any more of the black girl.

>>3154

Oh, didn't see your reply until now. Sorry for the late response. Only saw it now because I have finally finished that project on /fur and wanted to shamelessly plug it here...

>Do you have anymore religious one. Like the Amish. Or any more of the black girl.

I only redid this ebony girl here from the training data the other day when I was experimenting with multi diffusion settings.

I'd love to do do some amish or even quaker or mennonites etc. But sadly I haven't found a lora or a way to just prompt for that kind of outfit, yet. If you happen to know a char-Lora with that kind of attire for pony please let me know. :)

Now that I'm done with that other project I'll take a break for a few days and then I'll work on the dataset again as planned. So you can expect more uploads from me in a week or so. :)

Now since it entails pussy stabbing let me shamelessly plug that little project I did for /fur here real quick: https://boards.guro.cx/fur/res/15075.html#q15212 ...also as promised I go into a wee bit more detail about that no.1 project I mentioned which I'll finish sometime later this year.


Also *PSA TO SD-FORGE USERS*:

Forge has finally gotten an update/overhaul again: https://github.com/lllyasviel/stable-diffusion-webui-forge

I haven't updated myself yet, because atm I lack the neccessary backup space and I wanted to wait a little to see if folks report any major issues. But since I haven't seen the overhaul/update announced anywhere I thought it might be of interest to some here.

...and now for something completely different...

>>3290
>>3426

...So yeah, hope this and the other pics in the debreast thread satisfy your buttloving cravings for now. :)

As I've said, the Anna&Elsa thing will also have butcheeks cut off, but it is coming along slowly. Multichar work with multiple Loras is really hard. Takes time to not look shitty in my eyes. So no date when that'll be done. Besides it's too friggin hot for me to continue non stop genning at the moment and I'm still a bit salty and demoralised for that flopped project on /fur. So for the time beeing please refrain from immediate requests.

>>3465
No worries on the work that you're doing for us, take your time, I am in No rush. Thank you for your quality, very nicely done.

>>3465
Holy shit these are INCREDIBLE

>>3564

Both TehDave Loras also seem to enhance Katanas. This Pic was made with the first Lora I've linked ( >>3570 ).

Judgeing by the triggerwords the second one I linked ( >>3571 ) must have had a lot of pics of Princess in the dataset.

With both Loras her Outfit gets even better replicated when using "slingshot swimsuit" instead of "slingshot Bikini". Only found that out after this pic was done.

any more new stuff coming soon. haven't heard from u in a while

>>3702

I'm a bit busy irl currently. Only made a few random Lora tests, not much to show. Will hopfully have more time in a few weeks again.

...also >>3723

Do you have any new stuff to post. Are you doing a project

>>3724
I love this~

>>4242

I was doing a lot of testing, working on the Lora dataset and I was also asked to help on a few non guro projects. I'll have a few pics to share in a few days though. Some don't fit too well into any existing threads here so I'll most likely start a more general thread for my stuff. However just this morning my boot-SSD decided that today is a good day to fail. Replacement is on its way but may take until next week to arrive. Most of my gens, models and Loras are safe and have backups, but about the latest 8-10 days worth of gens are in danger of beeing lost. So wish me luck, please, T.T ...But even if I loose those recent gens I still will have a few things to share once my main rig is running again. So stay tuned! :)

>>4243

Thx! :) Can I ask what you like in particular? The pussy stab concept itself? The style? The char? (asking so I know what more to post of. ^^)

I like your stuff, too, very much. :D Would've commented a bit more often on your thread if I wasn't so busy as of late. While I'm not into every concept you've got there (like tentacles ain't my thing for example) I really dig your artstyle! Hope to see some pussy stabbing and more debreasting from you sometime. :) Really have to try out that Krita thing you're using. Alas, didn't get around to it, yet.

>>4244

Definitely the style and character, subject matter is a given on here~, but very nice piece overall~🖤

>>4247

thx! ^^ Let me know if there's something else you'd like to see me do with her (or any other char that can be done with Pony+Loras).

>>4242
>>4244

Finally fixed my hardware issues and recovered my data, so as promised its time to post some pics. First a little variant of >>3454...

...and then some pony remakes/variants of pics from my training data set...

...just to compare/for fun; the old "original" done well over a year ago with some old SD1.5 checkpoint.

...trying out some checkpoints for pixar styles in combo with pussystabV1 and Anytest also gave some interesting results

same deal here... wish the Lora would be this good on its own already! ...soon™ :D

...and again, testing that chckpt. ...She should not have taken a shortcut through some backalleys... not in this part of town!

(@mods: To me these look like a mix of a normal drawing/oilpainting with a very faint hint of pixar-ish (almost-) 3D. Also the proportions are clearly cartoony imo. So I assume these are okay to post. However, if I'm wrong with that assessment please don't completely nuke me. ^^)

...anyways, gotta have some simple backgrounds in the training data, too, right?

...another SD1.5 to pony port...

...been mostly porting some pics from my sabre swords -folder for the v2 Lora lately. So here, have some more sabre! :3

But before I'll finish this little posting frenzy with two more sabre sword pony remakes; this is a wholy new one for the katana dataset.

...I think somebody once asked for more redheads, so I hope you like this generic redhaired princess.

...aaaand the last one for now. As always, hope somebody likes these.

I'll need a bit longer before I can start that generic thread with non pussy stabbing stuff, so stay tuned.

>>4292

Ah, fuck it, posted the wrong file again. >.< ...here is the correct one... differences are just the fixed fingers and a few minor details.

Do you do request? Here's my thought. Have you seen the Ritz Crackers Christmas commercial? The one where the girl is sleeping on the couch. She wakes to the see from the Nutcracker. She makes peace in it. It's annoying 😑 can you kill her.

>>4349

You mean the commercial where the nutcracker fights with a rat?

>>4301
particularly hot, this one.

>>4357
Yes the one with the nutcracker

>>4401

Hmmm, not sure if I can (or even want to) replicate her looks that closely tbh. But if you can live with a cartoony version of that scenario and with a more generic version of that girl I'll give it a go.

>>4373

Agreed! Very juicy! :3

Cartoon version is awesome. Thanks

>>4706
Thanks. It was totally awesome. I knifed it very much.

>>4706
Too bad you couldn't have Ritz crackers shoved in her mouth. I still love it.

I like your comic. Are you going to do another one. I thought about a massacre. Like a sleepover, birthday etc...

>>4708

Pony doesn't "know" what saltcrackers are. If you let it loose it makes biscuits and/or donuts with toppings or slices of lemons instead 99 times out of 100. Otherwise I would have had her covered in crackers, that's what I had originally planned for her, but even just the pile of crackers on the table took hours (to not look completely shitty) so I was like "fuck it!" xD

>>4709

I already have a few massacre themed projects "under construction", but they're laying on ice for now. I've got a few more important projects to take care of first, before I can get back to those.

How do I find more of you or stuff eel?

>>4715
The cracker is cool. What do u mean by, find more of you

>>4715

All I've got can be found here on this board. Most is in this thread here but there's also some in the debreasting - thread; https://boards.guro.cx/ai/res/1864.html ...and some in the AI thread in /f; https://boards.guro.cx/fur/res/15075.html (in there is also my largest published project so far)

Other than that I haven't uploaded anything anywhere else.

>>4729

that wasn't me. The anon who posted that question may have mistaken the "subject"-line for a search bar or something xD (hence the red "EeL")

>>2399
I like that picture. When I go into https://civitai.com/models/545455?modelVersionId=606601

I get adult sensitive and can't view it. How can I view it.
I did sign up.

>>4732

On civit click on your profile name (upper right corner), that opens a dropdown menu. Go down and click on the gearwheel looking icon. A new menu will open up. Scroll down to the section "content moderation". Uncheck "Blur mature content" and check everything else in that section. reload civit and it should work.

But you'll be disappointed; that Lora you have linked never worked. There are no stabbed pussies at all, only one gay pic.

There's another Lora that works a bit better, this time there are also a few pics: https://civitai.com/models/584444?modelVersionId=652224

Do you have anything new for the holiday season?

>>4965

just a few, but no holiday specific ones. Anyways, enjoy! :3

(also >>5036)

alternate version of >>4301

...when Zarah's manager had landed her that last minute photoshooting gig just days before christmas she was annoyed at first until he told her about the amount of cash the client had offered. Besides, beeing flown out of her frosty eastern european homecountry in a private jet to spend a few days in some tropical paradise somewhere on the southern hemisphere wasn't such a bad thing either. Arriving late at the patreons mansion, the first day of shooting went smoothly with just some bland generic sunset shots of her on the beach. When she excused herself early, citing fatigue due to jetlag, they didn't even pressure her into staying for the ineviteable partying, she knew they'd later expect of her. The rich patron and his small circle of friends who had booked her and who were now acting as if they were professional photographers, were all just a bunch of typical boring nepo-babies, she thought. They all shared the same weird obsession for pseudo occultic fashion and accessoirs and all acted like hot shit about it. She had already kind of grown accustomed to their lack of taste, but when she saw the car that she was supposed to pose on and around on the second day it was just too embarrassing. She couldn't stop herself from rolling with her eyes and letting out a sigh of disapointment. She basicly checked out mentally and just wanted to get this over with. When she was ordered to lay on the heavily customised german cars hood her mind was occupied with thinking about how much cash she would demand, once her patron would muster the courage to ask for more than just a photoshooting. This wouldn't be the first time she'd blow some rich douchebag cocks for a few thousand dollars extra and maybe if she acted proactively she would also find a way to get home a little earlier, away from this cringefest. Had she not been distracted by these thougths she may have noticed what the photographer had in his hand and that his eyes started glowing red, when he suddenly stopped taking pictures and said: "There's something missing, stay like this..."

>>5041
I love this picture. How her guts ooze out her. Her pussy and her boobs are perfect. Very nice indeed.
You should if possible, put Eel on the weapon.

Hi just wondering how long it will be for the ass cheeks sliced off pictures

>>5138

>Her pussy and her boobs are perfect. Very nice indeed.

Yes she's really nice and juicy! Thank you! :3

>You should if possible, put Eel on the weapon.

Hmmm, you mean just for this pic or in general?

>Hi just wondering how long it will be for the ass cheeks sliced off pictures

hmmm, did I promise more of those? I don't remember. ^^ But I'll see what can be done. It'll be a while though, because I've got to finish a few other things first.

Cute redcoat gets it with the bajonet of a "Lebel model 1860" - rifle.

Original Art by Pixiv user eroing (https://www.pixiv.net/en/artworks/70451010). With friendly permission. :)

...also made this one >>5238 ...enjoy ^^

anything new in this thread. How about any Spanish girls or more black girls.

>>5593

Got a few new ones. Somehow I prefer to upload a few pics at once in bulk in this thread here, don't ask me why. >.< ...so I was waiting till I had a few more ready to go, but I suppose these few here will have to suffice for now.

This girl here isn't spanish but rather french (at least the char-Lora is). Been testing some semireal Illustrious checkpoints and found one that not only works with most pony concept Loras but also does pretty nice swords and stabbing stuff in general...

...so yeah, still getting used to Illu at the moment...

...The checkpoint I'm using is knkLuminai. Like I said, it is really good with stabby and pokey stuff. So far it has managed to use all pony concept Loras I've thrown at it, at least 2 Loras that were trained on pony but barely worked there even magically started working with knk. Only (pony) char loras struggle a bit for the most part, but even clothes loras work just fine.

...also finally I can work with a model again that knows what a stone altar is. :3

...so once you get the hang of it you can get some really nice results.

...but still pony has its merits, too. Like for instance I've yet to find a model that can do this detailed painting-esque kind of semireal style like this here.

... <=== She's probably another sister of >>1862 and >>2517...


I've got a bunch of other pics that don't fit into this thread. As mentioned in the request thread I'm currently working on expanding the dataset for an upcoming version of the impalement Lora. But I'm nearly done, shouldn't take more than another 2-3 days to get everything in order. Once I'm done with that I'll prepare to start another thread for all my stuff that doesn't fit here. So stay tuned for a few more days, at least ~30 impalement pics and general stabbing etc. coming soon.

>>5596
This picture reminds me of something. But I cannot figure what it is.

>>5656

Yeah, I've had the same distant feeling that I had seen this (or more likely a very similar pic) before when I had finished it. But I don't remember where either (nor if it was cg or painted or whatever).

>>5658
Maybe is like that object which cotton candy is on. When you selling it in a football stadium 🏟. Instead it's a girl with a sword in her pussy

>>5659

That too. xD But seriously I could've sworn to have seen a very similar pic with a very similar girl with her kemono/furisode slipped from her shoulders and skewered on something like this before. :)

It is very rare that I get a really good pic just by prompting. Usually I'll have to do tons of inpainting etc., which takes a ton of time. But sometimes all the AI-stars and -planets align and the AI gods smile upon the dusty old rotten potato that is my PC; I check civit at least a few times a day to make sure I don't miss any Loras that might be of interest. Usually I just keep an eye out for obvious guro related stuff as that tends to be deleted within hours if not faster. There are however some characters (~Loras) that will always catch my eye, as was the case with this Bouquet Lora. Literally the only test prompt I ran with her gave me this base pic and all I had to do was to let multi res do its thing on it, so this pic here took less than 20 minutes to be made, most of it while I was doing the dishes. ...Though it is nice to sometimes just get what you want without the AI making a fuss about it, I don't know if I'd like it, when this were always the case... cause I hate doing the dishes! >.<

Any new stuff?

>>7146

Just a few...

...Some experimenting with G. Roberts - style Loras...

...really like this one! :3

...and the last one for now. ^^

>>7861
>>7862
simply awesome
>>7863
>>7864
also good

>>7864
LOVE the Roberts style!
even though the "pussystabbing" is not exactly my cup of tee, this picture is at least 9.5 out of 10.
RESPECT! keep going.

Do u have more coming soon

>>9215

Really not too much, just got a few I made in between the mammoth project that I'm still busy with. Let's start where we left off; some egyptian damsels! Here we go:

...and another...

...and some random dungeon damsel I already posted on ATF, so sry if you've seen her already.

...now let's take a look at how disney's Damsels fare against some blade-work (all in the spirit of/as a nod to/inspired by a certain old artist from rule34 known as col. Kink).

First we see Ariel feeling true love deep inside (more of her in my other thread, too), with...

...aaand without dialogue...

...auntie Belle is also a good sport...

...without dialogue again...

...and, oh boy! - Melody is a quick learner, too!

...and the last one again with no dialogue.

Th-th-th-that's all folks! :3

Wanted to fix a spelling error, that ended up blocking me from uploading this here, so let's try this again: ...and another cute little harem dancer's demise, pretty similar to one I posted before, but still I like her and it is in another aspect ratio. So I hope someone enjoys her, too...

now we're done! :3