>>722
Good work. Impressive.

You can re-train with 768x768 using this google colab

https://colab.research.google.com/github/Linaqruf/kohya-trainer/blob/main/fast-kohya-trainer.ipynb

>>722
Good work, not a lot of people sharing loras for this genre. If you share the dataset+tags I (and maybe other anons) could try to do a 768 bake with NAI for wider compatibility.

If not, give a try to adjusting your training settings. Like go from 1 epoch to maybe 6-8 and turn the image repeats way down because 120 repeats is way too much for 450 images. Like 12 repeats, 8 epochs or something like that.

Here is the dataset in fullres
https://anonfiles.com/n1V3z7adz6/DataSet_zip

i'll try to make some models on some other specifics, thanks for the training advice

>>730
>>730
Fucked around with it for a bit, decapitation_anon is yours and right is mine in pic. Much less fried, lr2.4-e4, telr4e-5, unet2e-4, batch 4, epochs 6, repeats 6. You can look at the training info in additional networks tab to see more details.

Having hundreds of images gives a lot of leeway for training so probably could have baked even longer. Works okay at 1.0 weight, up to 1.2 even but I didn't test it that much. Target tags seem to work at least a little bit if you give them more weight (guillotine:1.3). Probably didn't catch them all since I only had keep_tokens=4.

https://files.catbox.moe/0iggb5.safetensors

Some more notes for self-help:
- enable ss_shuffle_caption and keep_tokens=# (the number of tags from the beginning of the tag file you want to keep per image for every iteration = your targeted activators), shuffling tags makes it less likely to fry/flexible. Same reason you want more epochs than 1 since epochs shuffle the image order around.
- always train on the novelai model, most models and mixes are rooted in it and the lora will work much better with most models this way
- https://rentry.org/RentrySD/#1162-lora-guides-and-info has quite a lot of information, wherever you learned that 100+ repeats is okay is garbage.

impressive results, thanks for the advice ill use it in future models.

So as someone fairly new to this:

This model file is basically just containing the training you did, and I have to merge it with another model to use it? (You said you used Grapefruit_v4, but in theory it could be merged with a plethora of other models)

Do you/anyone else plan on training different guro models in the future? It would be interesting to get more head crush, half decapitation (cut at jawline), spilling entrails, etc. working.

Also, as for hosting on civitai or other model hosting sites, considering its possible to generate stuff like head destruction or eye socket penetration with several more "normal" models, I wonder if maybe you could generate some images of knights, orcs, goblins in battle and maybe market them as "Combat/Action RPG" models for illustrating D&D campaigns and Tolkien - esque Fantasy battles. But then *we* know what the model is actually used for here. Lol.

>>733
I plan on making some more models when I feel like it
I dont want to dedicate any serious amount of time to very obscure AI generated guro pornography
but with the very helpful advice of >>731 ill probably make a head/ear/skull fucking model next

>>733
No, it's like "plugin", do you just need to put it to appropriate folder and include in promt.

thanks a lot to >>722, that a great advance of AI guro art work.

it gives surprisingly good results.
Thank you

>>737
Thank you OP

Fantastic tool and capable of reliably creating guro content, thanks a lot for the work here... I have had the best success with the mixedmixedmixedv2 model, grapefruit and "uberrealisticpornmerge" for more realistic images, the grid greated below was using these settings:
Positive: masterpiece,best quality, (body_pile:1.4), multiple girls, 6+ girls, sportswear, sports uniform, <lora:0iggb5:1.2>, sunny day, arms behind back , (head_flying:1.4),(severed_head:1.4),(decapitation:1.4),(neck_spray:1.4), (high detailed skin:1.2), 8k uhd
Negative prompt: (worst quality:1.2), (low quality:1.2), (lowres:1.1), (monochrome:1.1), (greyscale),blurry,easynegative,(bad anatomy:1.2),cencored, censored, moasic,low quality,worst quality,bisexual,futa,futanari,insert,missing hand,missing arm,missing leg,missing foot,low quality, worst quality,ugly,fused fingers, extra fingers, extra toes,malformed hands, mutated hands,mutated fingers, mosaic censoring, (weapon_trail:1.1)
Steps: 50, Sampler: Euler a, CFG scale: 7, Seed: 2002124325, Size: 1024x1024, Model hash: 0469912235, Model: mixedmixedmixed_(1)

Used embeddings: easynegative [119b]

unfortuantely I will be away from my PC for some months in the near future so will not be able to tinker and contribute as much as I would like here (with a little more time I would have loved to teach myself a little more on the training side and create a hanging/ garrote/ bagging SD Asphyxiation Lora) but I will try and at least drop some more prompts that I have had success with to get some other people started

just a quick demo set, this prompt could do with some fine tuning but does the job:
Positive: masterpiece, best quality, (body_pile:1.4), multiple girls, 6+ girls, sportswear, sports uniform, spread legs, rape, raped, cum, creampie, cum inside, scared, angry, helpless<lora:0iggb5:1.2>, sunny day, arms behind back , (head_flying:1.4),(severed_head:1.4),(decapitation:1.4),(neck_spray:1.4), (high detailed skin:1.2), 8k uhd,
Negative prompt: (worst quality:1.2), (low quality:1.2), (lowres:1.1), (monochrome:1.1), (greyscale),blurry,easynegative,(bad anatomy:1.2),cencored, censored, moasic,low quality,worst quality,bisexual,futa,futanari,insert,missing hand,missing arm,missing leg,missing foot,low quality, worst quality,ugly,fused fingers, extra fingers, extra toes,malformed hands, mutated hands,mutated fingers, mosaic censoring, (weapon_trail:1.1)
Steps: 50, Sampler: Euler a, CFG scale: 7, Seed: 2999542804, Size: 520x512, Model hash: c590550ea5, Model: grapefruitHentaiModel_grapefruitv41



Used embeddings: easynegative [119b]

looks like img2img has some uses

>>739
very thanks, that help me a lot, truly amazing.

>>742 That is good to hear, I should give credit to https://www.pixiv.net/en/artworks/106012056 who got me kickstarted by including some of their prompts in their descriptions, you can find some more ones there :)

This one has an odd habit of transporting itself outdoors for me when I bump the resolution up from 512*512 to 1024*1024 (so I just ran at 512 and upscaled), apart from that I am quite happy with the trophy displays it makes:

(tested on mixedmixedmixed and grapefruit, good results in both but somewhat more reliable in mixedmixedmixed as that is what I mainly used when tuning the prompt)

Positive: masterpiece, vibrant, best quality, (multiple heads mounted on trophy plaques:1.4),(severed_head:1.4), (head_trophy:1.3), (fantasy creature:1.4), trophy room, well lit, beautiful day, clear facial expression, pale skin, taxidermy, wall decoration, Exquisite details and textures, clear detail, crying, terrified, (weapon_trail:0.4),<lora:0iggb5:0.8>, female focus, 8k uhd
Negative prompt: Negative Prompt:
(worst quality:1.2), (low quality:1.2), (lowres:1.1), (monochrome:1.1), (greyscale), mutated, blurry,easynegative, Torso, Arms, Legs, bad anatomy, censoring,mosaic,low quality, worst quality, duplicate face, clone face, bisexual,futa,futanari,insert,missing hand,missing arm,missing leg,missing foot,low quality, worst quality,ugly,fused fingers, extra fingers, extra toes,malformed hands, mutated hands,mutated fingers, mosaic censoring, weapon_trail, blood

Steps: 15, Sampler: DDIM, CFG scale: 6, Seed: 2874633958, Face restoration: CodeFormer, Size: 512x512, Model hash: 0469912235, Model: mixedmixedmixed_(1), Denoising strength: 0.7, Hires upscale: 2, Hires upscaler: Latent

The low step count and use of DDIM here was as this was meant to be a test run for some settings and it came out quite well, if you have the time I find some of the DPM Karras variants can really make details pop at 40-60 steps... it does really like to take its time though so you might want to queue up 100 or so for an overnight job on a dialled in prompt and see what it gives you in the morning.

This was an earlier stage of the above, focused on creating single head trophies,this one plays a little better with grapefruit:

Positive: masterpiece, vibrant, best quality, trophy room,(severed_head:1.4),(wall mounted hunting trophy:1.2), severed head mounted as trophy, hunting trophy, (fantasy race:1.4), woman, (clear face:1.4), well lit, beautiful day, clear facial expression, pale skin, taxidermy, wall decoration, Exquisite details and textures, clear detail, crying, terrfied, (weapon_trail:0.4),<lora:0iggb5:0.8>, female focus, 8k uhd


Negative prompt: (worst quality:1.2), (low quality:1.2), (lowres:1.1), (monochrome:1.1), (greyscale), mutated, blurry,easynegative, Torso, Arms, Legs, bad anatomy, censoring,mosaic,low quality, worst quality, duplicate face, clone face, bisexual,futa,futanari,insert,missing hand,missing arm,missing leg,missing foot,low quality, worst quality,ugly,fused fingers, extra fingers, extra toes,malformed hands, mutated hands,mutated fingers, mosaic censoring, weapon_trail, blood,
Steps: 30, Sampler: DPM2 a Karras, CFG scale: 5.5, Seed: 50682516, Face restoration: CodeFormer, Size: 512x512, Model hash: 0469912235, Model: mixedmixedmixed_(1)

Same prompt as above

A few odd tips for prompt crafting: you can test prompts quickly using the DDIM sampler even at low steps, it may not be the prettiest but should let you quickly check if it is generating what you want and tune the prompt before you use higher iterations or a slower sampler

The reason I have sword flash in a lot of my negatives is because this LORA sometimes confuses neck blood splatter and sword flashes, resulting in them being placed where you do not need them/ putting white lines where you do not want them around the neck, if you have that issue doing the same may solve it.

doubling edge size (512 - 1024 for instance) quadruples the amount of work, if things are going slow in testing "halfing" resolution will SIGNIFICANTLY increase speed

in many cases (especially if you are trying for a very specific look to your image and need to go through a lot of iterations, doing the above and just upscaling the result or using a "nearly there" result and using a little tweaking/img2img work may get you there faster than just letting 100 renders of the maximum resolution and quality you can run so you can find that one that is just perfect

Try different models, and LORAS a lot are available here: https://civitai.com/ and can change the style of an image totally, below is a quick prompt using a realistic model:

Positive: RAW photo, 1girl, photograph, (female police officer:1.1), kneeling, bra, <lora:0iggb5:1.2> , solo, kneeling on road , seiza , arms behind back , beautiful day, (head_flying:1.4),(severed_head:1.4),(decapitation:1.4),(neck_spray:1.4), (high detailed skin:1.2), (blood splatter:1.1), 8k uhd, dslr, soft lighting, high quality, film grain, Fujifilm XT3
Negative prompt: (worst quality:1.2), (low quality:1.2), (lowres:1.1), (monochrome:1.1), (weapon_trail:1.4)(greyscale),blurry,easynegative,bad anatomy,cencored,moasic,low quality,worst quality,bisexual,futa,futanari,insert,missing hand,missing arm,missing leg,missing foot,low quality, worst quality,ugly,fused fingers, extra fingers, extra toes,malformed hands, mutated hands,mutated fingers, mosaic censoring, (weapon_trail:1.4)
Steps: 40, Sampler: DPM2 a Karras, CFG scale: 6, Seed: 1412703140, Face restoration: CodeFormer, Size: 512x512, Model hash: f93e6a50ac, Model: uberRealisticPornMerge_urpmv13
https://civitai.com/models/2661/uber-realistic-porn-merge-urpm

Now this one could do with a little tuning but tinkering and finding out may get you some fun results.

Feel free to remove prompts - trimming down your prompt list can help in many cases, too many prompts for facial expressions can result in odd faces as it tries to make three espressions at once for instance

PROMPT ORDER MATTERS: prompts near the start of your prompt will have additional weight to them over those at the end, if you want a specific pose, outfit or object in the scene and it is not appearing put it near the start of your long prompt list instead of the middle and leave the petty tweaks for the end.

Wasted Prompts: If the model you are using already creates beautiful faces by default a "beautiful face/ perfect face" prompt could be a waste of a prompt or even interfere with what you are trying to do (for instance if you want a blindfold on the head and the model decides that this conflicts with the perfect face prompt causing it to generate an odd blindfold that you can still see the victims eyes behind - if something is handled already less prompts may be more

Think outside the box - if a prompt that should work but seems to be playing up check if you have anything that could be misinterpreted in a prompt - I was trying to create a brightly sunlight scene in a town with "well lit" as a tag, it took 20 minutes of tinkering with a prompt to suddenly noticing that it had interpreted well lit as "an area with many streetlights illuminating it" - as street lights typically only light areas at night it kept creating night scenes for me despite other tags prompting for daylight.
Another example of this would be when I was having issues creating a "decapitated head on a spear" prompt that would reliably work and place the spears in the right position, swapping the spear prompt out for a flagpole got me a significantly more reliable vertical pole that stuck into the victims throat (below)

Positive: masterpiece, vibrant, best quality, trophy room,(severed_head:1.4),(severed head mounted on flagpole:1.4), young woman, perfect face, pale skin, blonde, (clear face:1.4), (bright fantasy background:1.2), sunny day, clear facial expression, pale skin, taxidermy, Exquisite details and textures, clear detail, crying, terrified, blood splatter, blood drip, (weapon_trail:0.4),<lora:0iggb5:0.8>, female focus, 8k uhd
Negative prompt: (worst quality:1.2), (low quality:1.2), (lowres:1.1), (monochrome:1.1), (weapon_trail:1.4)(greyscale),blurry,easynegative,bad anatomy,cencored,moasic,low quality,worst quality,bisexual,futa,futanari,insert,missing hand,missing arm,missing leg,missing foot,low quality, worst quality,ugly,fused fingers, extra fingers, extra toes,malformed hands, mutated hands,mutated fingers, mosaic censoring, (weapon_trail:1.4)
Steps: 40, Sampler: DPM2 a Karras, CFG scale: 6, Seed: 3959950693, Face restoration: CodeFormer, Size: 512x512, Model hash: 0469912235, Model: realisticVisionV13_v13

(obviously this could do with further tweaking but this makes a decent demonstration of that issue)

and finally, Tinker, Tinker and Tinker more, use other tools apart from TXT2IMG, expecially if you find a render you like, drop it into IMG2IMG and try seeing what it looks like after being run through another model or with a small tweak to the prompt, the four renders on the right are all img2img renders, the prompt and source image stayed the same but the model/ checkpoint and denoising strength were changed resulting in four quite different images that where fast to render and retained the core of the source image.

With a little luck I may be able to drop one or two more decent prompts in this thread before I go offline for a few months, hopefully my last 10 or so posts will help some people

>>749
Oh damn! So realistic. I love it!

Wow.
Thank you for a detailed write-up.
I tried with uber reaslitc porn merge and had very bad results.

The body is ok but hen it tries to draw a flaying head it fails, doesn't recognize face features at all.

>>753
Try https://civitai.com/models/4201/realistic-vision as well if you need something realistic, it may work better. from memory (I would need to retest but I think I encountered a similar issue) porn merge suffers with heads that are not vertical in the photo, what prompt/ settings are you using?

>>750
>>753

Okay, I just poked around to make a flying head focused prompt for testing and it seems fairly reliable (I am assuming you want a body with flying head), try a low batch count and change one or two things as a time to see if that helps narrow down a cause

For reference I am using the AUTOMATIC1111 stable-diffusion-webui: https://github.com/AUTOMATIC1111/stable-diffusion-webui

Positive Prompt: RAW photo, 1girl,(head_flying:1.4), (pretty female police woman:1.3),(severed_head:1.4), neck_spray,<lora:0iggb5:1.2>, beautiful day, (high detailed skin:1.2), (blood splatter:1.1),(decapitation:1.4), 8k uhd, dslr, soft lighting, high quality, Fujifilm XT3

Negative prompt: (worst quality:1.2), (low quality:1.2), (lowres:1.1), (monochrome:1.1), (weapon_trail:1.4)(greyscale),blurry,easynegative,moasic,low quality,worst quality,bisexual,futa,futanari,insertlow quality, worst quality,ugly,fused fingers, extra fingers, extra toes,malformed hands, mutated hands,mutated fingers, mosaic censoring, (weapon_trail:1.4)

Steps: 20
Sampler: DDIM
CFG scale: 7
Seed: Random
Face restoration: CodeFormer,
Size: 512x512,
Model hash: f93e6a50ac,
Model: uberRealisticPornMerge_urpmv13

>>755
The exact same prompt as above with the model swapped to realisticVisionV13_v13

If you are using AUTOMATIC1111 try matching the settings here and working from there DDIM/ 20 steps is nice for speed and can let you test things quickly

(obviously feel free to decrease batch count to whatever you like, that will not make a difference to the rendered images)

or possibly ignore me for a moment, I encountered exactly your issue when I tweaked that prompt to draw a full body, lemmie fiddle a little and see if I can make a more reliable base prompt

I'm using AUTOMATIC1111

Samantha Saint, , ((wide angle)), leash, collar, sexy, (submissive:1.3), (kneeling), (spread legs),( big breasts), bondage, fetish, (pigpen), (inside a barn), (cattle in the background), stable, (decapitation:1.3), (severed_head:1.4), neck_spray, (head_flying:1.4), sword, nude, blood, death, guro, <lora:Decapitation:0.8>
Negative prompt: cartoon, animated, pubic hair, mask, ((blurry)), duplicate, (duplicate body parts), chubby, fat, close-up, child,
Steps: 24, Sampler: DDIM, CFG scale: 7, Seed: 1839635961, Size: 480x640, Model hash: 40f9701da0, Model: uberRealisticPornMerge_urpmv13

>>760
But with some inpainting yu can get decent results

>>761

Unfortunately you may have to stick with that, the moment a body is in frame with a flying head at an angle it really likes to throw a fit and most of my straight txt2img ideas have not helped :( - still, if it works it works and this model is the best realistic one i have tested for this apart from the flying head issue

Still pretty good results.

>>762
But as you said you need to play in post-production.

I found it is easier to generate an image without a head, then inpaint an empty part of the image set denoising strength to 1 and draw a flying head over there.

Another things i found orks great is img2img from 3d renders.

Original

>>765
img2img



parameters

grill, spit roast, blonde, open mouth
Negative prompt: animated, pubic hair, mask, ((blurry)), duplicate, (duplicate body parts), chubby, fat, close-up, 3dcg, (closed eyes)
Steps: 16, Sampler: Euler a, CFG scale: 7, Seed: 2200185395, Size: 1512x992, Model hash: fcfaf106f2, Model: uberRealisticPornMerge_urpmv12, Denoising strength: 0.4, Mask blur: 28

Works quite well.

This was even before the Decapitation Lora

Best results ere ith amputee as uberrealistpornmerge worked badly ith blood.

I want to contribute to this.

I have tons of images, I need to sort through them and add tags.

SO far I can think of all danbooroou tags, realistic, 3d anime

ALl the ones mentioned by you.

how do you add tags to the images?

>>769

I posted the dataset above
>>730

the tags are in txt files with the same name as the image

besides the manually labeled tags, the rest of the tags were generated by deepbooru

Tha hanging SD produces much better results with UberRealisticPOrn Merge then decapitation.

Still the bloody neck stump and blood fountains come out nice.

>>722

Hi Guys, I've been out of the loop for a while.

I would like to know if anyone trained a good realistic gore model, with special emphasis on guts, brains, exposed muscle and bones since all my tries to make gore end up in a bit of blood and deformed red blobs without shape.

Is it possible? Anyone have heard of something similar?
Any help is appreciated Thanks.

(AI) Beheading Bea Snuff

Pixiv... https://www.pixiv.net/en/artworks/106595783

anyone have a new link for the lora? it seems to be a dead link/

>>774
https://www.pixiv.net/en/artworks/107340921
Made a new one.

>>774
The link still works fine for me

So HOT!

>>783
>>782
>>781
>>780
>>779
>>778

my work.

please check my pixiv
https://www.pixiv.net/users/1651150

>>778
I'll show you the one before I posted it on pixiv.

>>785
Love your works man!
Sorry I forgot to attach your pixiv link in the posts.
Could you please share the uncensored pictures which you posted on pixiv here? Thank you so much!

>>780
Especially this one!
It would be lovely if we can have the uncensored masterpiece.

>>786
an uncensored version ould be fantastic.
Or just give us the prompts you have used.

>>788
But I also fear the reason for the censorship is cause the model does not create a good wound/pussy

>>785, >>786

I am glad you like my work.
Thank you very much.

As for my method,
I start by making a regular piece first.

(empty eyes:1.4), baby face, (nsfw, erotic, sexual:1.2), korean face, (naked latex black boots), (factory ruins:1.2), outdoor:1.2, (naked latex black thighhighs), (naked latex black gloves), (gun holster belt:1.2), (((masterpiece))), (cleavage,curvy:1.2), (Big eyes:1.2), (beautiful detailed face:1.2), ((glamor)), nude, (beautiful detailed nipple), (beautiful detailed pussy), (perfect detailed breasts:1.2), ((best quality)), ((intricate detailed)), ((Hyperrealistic)), (pussy juice:1.2), absurd res, perfect calves, perspective, highly detailed, illustration, (huge breasts), perfect hands, detailed fingers, beautiful detailed eyes, blue eyes, blurry background, perfect eyes, seductive eyes, pureerosface_v1, <lora:add_detail:1>, <lora:last:0.2>, mature female, milf, loli, beautiful legs, ((blood on face, nosebleed, blood splatter:1.4, torn clothes, blood)), the hurting, , (cuts), heavy bleeding:1.4, bleeding:1.2, blood:1.2, face cuts, the hurting;1.4, cuts on body:1.4, (blood on body:1.4), ((blood on body:1.4)), ((blood on body:1.4)), ((blood on body:1.4)), ((blood on body:1.4)), kneeling, short bob cut
Negative prompt: EasyNegative, badhandv4, (low quality, worst quality:1.4), poorly drawn hands, bad anatomy, monochrome, { long body }, bad anatomy , liquid body, malformed, mutated, anatomical nonsense ,bad proportions, uncoordinated body, unnatural body, disfigured, ugly, gross proportions ,mutation, disfigured, deformed, { mutation}, {poorlydrawn} , bad hand, mutated hand, bad fingers, mutated fingers, badhandv4 ,liquid tongue, long neck, fused ears, bad ears, poorly drawn ears, extra ears, liquid ears, heavy ears, missing ears, fused animal ears, bad animal ears, poorly drawn animal ears, extra animal ears, liquid animal ears, heavy animal ears, missing animal ears, bad hairs, poorly drawn hairs, fused hairs, bad face, fused face, poorly drawn face, cloned face, big face, long face, bad eyes, fused eyes poorly drawn eyes, extra eyes, bad mouth, fused mouth, poorly drawn mouth, bad tongue, big mouth, bad perspective, bad objects placement, ng_deepnegative_v1_75t, ghost, bed sheet, on bed, panties, (red, orange, yellow:1.3), pillow:1.

>>790

Next, use "img2img" "inpaint" to fill in the head and LoRA to remove the head.

The results will vary greatly depending on which model you use, so I suggest you use whatever is available in your environment.

>>791

But to be honest, this alone doesn't make for a very good piece of work, so I use gimp to modify it.

>>795
Thank you for the specific instructions and the prompt!
>>795

hello, im really new to this kind of stable diffusion thing and looking for how to make those kind of genre.

im sorry if my question bothers someone...

the grapefruit_v4
https://civitai.com/models/2583/grapefruit-hentai-model
goes to checkpoints

and
f608kn.safenensors
https://files.catbox.moe/f608kn.safetensors
goes to lora

am i right...?

using stable diffusion web ui, and even tho i put those in those section, it generates default models.
the decapitation itself seems working fine.

am i putting a wrong model or something...?

>>797
You need to select the model from the menu. After you put it into the folder

also the model iteself (grpfruit-hentai -model) goes into the the \models\Stable-diffusion Folder

I iscovered a breasts_on_plate LORA
and Im in love
https://civitai.com/models/7142/breastsontray?modelVersionId=8397

>>800
Nice.

>>802
Anyone tried generating guro stories with AI?
How hard it is?

>>803
If you mean text stories I generated plenty of them on lit

same thing as images you write prompt and you get a story or you can guide AI more precisely

How are you mixing models together?

>>731
I used your model and it won't actually have decapitated heads, just girls with blood around them. What am I doing wrong?

>>804
tell me which prorgams do you use?

Moor bdsm torture quad amputee

My goal is to achieve this head as a fleshlight idea:
https://www.pixiv.net/en/artworks/111063487

I installed the linked tensors but when I enter them I just get some girl staring at me with her tongue out.

Any help with getting this to work? I am using Automatic 1111.
I placed the file toubu_re-000010.safetensors in A1111\models\Lora

Am I doing something wrong? Can anyone help to get this working? Do I need some other dependcy? I copy paste what his prompt is but it just doesnt work, do i need to activate this or something I forget.