in the words of the creator in the unstable diffusion discord "- Reworked the entire recipe multiple times - Smaller file size thanks to reworking things coincidentally (from 6GB to 4GB) - Added more model merges - Added an anime merge without losing realistic textures (bring on those danbooru tags) - Fixed the CLIP. This will allow better prompting - Way better anatomy. Anatomy issues will still happen, but a lot less often - More porn merged in, allowing for more zoomed out shots (when prompted right) - Better faces when looking behind - Better fingers, and less double assholes - And so much more that I forgot to mention"

Click on the "Show more" under the tensor and description.

+ 0.1

@saftle Sorry, I already fixed the problem, it was caused by the SD trying to use an incorrect VAE

you have to train it sets of images that are captioned with "small cock" "big cock" etc.

i've tried using the ckpt and the safetensors btw

Have you tried both the normal 1.3 version and the inpainting 1.3 version?

@HollowStrawberry yes, the normal works fine but the inpainting just produces noise

@saftle i was using the stable-diffusion-webui localhost as well as painthua. however I got it to work perfectly eventually by using a clean install of stable-diffusion-webui. thanks

If you are using AUTOMATIC1111 have you updated it to the latest version?
right click anywhere inside stable-diffusion-webui folder, click on "git bash", then write "git pull"

you can use colab.

@saftle yes，i try，but not useful

you can

Try stable-diffusion-webui，you can use it with your mobile phone

maybe draw black colour on legs and then set denoise to 0.4-0.6

"thigh highs"?

Lingerie? Panties?

@beignets @whywhynot @ai_ai_ai_ai_aiiiiiii I try again and agin , then the best prompt are still ((black pantyhose:1.5)), but can only reach about 30%(29 of 100 results can be recognized as pantyhose, more are ordinary silk stockings) corret rate.And when disable "skirt", rate rise to 70%(72 of 100), however it is sad to see this as most of time we need skirt on output image.

Try (full length pantyhose)

@saftle Yeah I've done that, but while the embedding looks perfect in SD 1.5, it looks slightly off in URPM and other models I've downloaded from this site seem to work the same way. I mean I can recognize certain features in their face, but it still doesn't look quite right, even when trying other tags to fine tune that. I'm not sure what's so different about the downloaded models that causes issues in embedding training. I'll keep trying other things including LORAs, but I'm just curious about why this doesn't work.

@fudefrak i got some good results with LoRas trained on SD 1.5. You can try a few things:
- don't use prompts that describe the face (like "perfect eyes, pure eros face, angelic face" etc.) this can break LoRas
- generate images at 600x600 instead of 512x512 (for some reason this improves my LoRas, works on other SD models as well)
- this model uses VAE, try generating without VAE (this can make a big difference in some models)
- to avoid "anime" look like getting big eyes & small nose, use negative prompts: cgi, render, cartoon, anime, drawing, painting, child, girl, boy

- any type of face fixing can break LoRas as well, like Restore Faces, GFPGAN or Codeformer. If you want to fix a face: send image to Extras -> Upscale 2x -> send image to inpainting -> inpaint face area and set resolution to 512x512 or 600x600 and generate using the same prompts from text2img until you get a good face

I have the opposite experience, i train on 1.5 base and then render in this model, even pretty bad training images get very good here. Its my favorite model and that's mostly for non porn.

@olej1964 It's true the images look better overall, in terms of detail, proportions, composition, etc, but they don't always look quite as much like the subject I trained on

@fudefrak  You have to first understand how embeddings work and then you will understand why it's off. TIs reverse engineer their way into tricking the base model to produce a specific output based on the tokens you provide. This means when you try to use an embedding on a model that the TI was not trained on, you will not get the same results. And, the more "fine tuning" that occurs in the base model, the more obvious the differences will be. That's why an embedding trained on SD 1.5 will not work on 2.1 or other custom/scratch models.

Search YT for "LoRA vs Dreambooth vs Textual Inversion vs Hypernetworks". It will be 20 minutes well spent.

@dabeckham Oh yeah I completely understand why it happens. That's why I was trying to train an embedding on other models like this one but was not able to.

@saftle If you could then that would be awesome :) my potato laptop does not have a lot of storage and only 4GB VRAM so your pruned models have always worked for me without any problems ^_^ Thanks anyway for this model and for considering my wishes

@saftle thanks a lot 😊❤️

@saftle That's i actually do ! maybe i have to increase the strengh of those one !

@saftle ok ok thanks! so only patrons get that or we common men can also get a chance to download it?

@saftle ahh, my bad! tis happens when we smoke too much :P, Doubts cleared!

No bro its not for smartphone only high end specs pc can handle this huge rendering process

https://runpod.io?ref=cbh8ctqe

Go here... sign up. You can run the program as cheap at .20 an hour. Once you sign up, add money. Select community cloud, pick rtx 3080 pod (should be good enough unless you want to generate pictures faster), select template "RunPod Stable Diffusion v1.5", run the instance. Select "connect" then jupiter lab. Navigate to workspace/stable-diffusion-webui/models/Stable-diffusion. open new tab "launcher" and then select "terminal." type "wget https://civitai.com/api/download/models/4971" Rename the downloaded file and make sure you add the extension .safetensors. Now, go back to where you selected "jupyter lab" and click "connect via http" At the top, there is a drop down menu with the .ckpts. Select the file you just renamed. Start generating. The power of Automatic1111 run on a gpu through your android phone. =)

u can use google colab with automatic1111

反向可以不填。正向直接用你能想到的一句话来翻译成英文，另外别忘记加点真实，照片，大师级之类的tag在前面。配合你想要的一些lora生成你想要的特定人物。捏效果很差应该是没有填写VAE。在设置-stable diffusion里选择合适的VAE，这些在B站都有教程。

老铁 你们用的哪个封装的启动器呀 我用秋叶那个 生成成人图都变彩色马赛克 。。。。

@TTAI 不用启动器，原装最好用

@xiaoyao5715 启动器会有什么缺点吗

我一直想着怎么把衣服穿的多一点，你……

@YuntaoHu 我也是,沒衣服很快就厭倦了, 還是若隱若現的最讚

https:// bbs.imo utolove.m e/read.php?ti d=17 39126 把空格去掉，这个楼主专门做的教程

CFG参数别太大

这个模型是不是亚洲脸孔都效果不太好？

应该是附加网络里没有载入lora，要把你看到的图片标题名称在stable diffusion的civitai里下载，然后在前面的附加网络里加载一个lora的model

@3072854529910 感觉是

Thanks for the quick reply! In the Dreambooth colab there is a line to point to a hugging face model. I add saftle/urpm there but it does not find a model. Any idea what I'm doing wrong? Thanks!

@saftle It appears to have been a VRAM issue. The pruned version works fine. Thank you!

@saftle Yeah it looks like that model is just close-up images of vulvas and nothing else.

Can yo walkthrough how this editing stuff works? Which at is used to generate these images?

I usually type in the negative prompt, "large breast".

我用lora就會變成同一個臉，但是用默認韓國/日本會出現幾張不同的臉 默認的臉質量還不錯 就是lora不管用

prompt加beautiful可以让脸和发型更符合亚洲审美，如果不喜欢欧美系亚裔脸，可以尝试"beautiful Caucasian, (Asian 0.4)"这样自定义的比重，最后再混入你喜欢的lora

那是tag没写好

是不是可以merge其他asian的checkpoint解决这个问题？

你可以試試提示詞(((Chinese girl)))

@saftle you are amazing

Yeah it really struggles to do a realistic penis or testicles, the examples I've seen on here have been really weird.

Try a penis Lora, that might help

@saftle thankyou for your efforts, new here - realised it too late. Tks

use the --no-half alone

and watch if ur python crashes or not before starting

same here and --no-half is added to commandline args and upcast cross attention layer to float32 is checked in settings

@saftle oh wow, thanks for responding, i hate how hard it is to find any info, seems the tech is just moving too fast for us humans to catch up

@saftle I re-downloaded the inpainting model. Runs without any errors, thanks. Drove me mad, you saved my day.

little tip for in-painting if you havent tried it, change the inpaint area setting from 'whole picture' to 'only masked'

Results are far, far better. Maybe this is common knowledge hah but just in case

@eye_am_bored inpainting is a bitch, i need to fiddle so much

the majority of the reason for LORAs being used is because stable diffusion in general has issues drawing the complexities of penis and vagina in general, and I feel it's likely that the problem is that without a LORA, the typical process of SD in general without running special models and LORAs won't get you anything remotely adult oriented most of the time. it's extremely rare that you get anything adult content at all. This model is typically used for adult porn renditions of sexual acts, but even then, poses and positions and other details are still likely necessary in order to produce the outcomes you specifically are pursuing.

@umisguy348 Thanks for your feedback. Yea, i figured this is more of a base you have to mod on top off. But as soon as i prompt erotic outifts or explicit poses i do actually get NSFW compositions without using any embeddings or LORAs, you just have to "squeeze" it a bit, just the genitals usually are ridiculous

use this, it does all the work of installing the webui for you, it might break sometimes but thats the price you pay to get this done easily, otherwise follow Automatic1111's webui installation instructions

I now realise Ive encouraged people to be lazy idiots and come here to aks these questions, next time use google

you most likely dont have xformers installed.
to fix it, connect your python path to where xformers is installed.
OR another way to fix, is open the aunch file in the main directory and add this line

commandline_args = os.environ.get('COMMANDLINE_ARGS', "--xformers")

You'll need a controlnet for that

Ok, any idea when there might be an updated 1.5 model? The improvement from 1.2 to 1.3 was really good, so I'm interested to see where it will go from there.

These are designed to be used on a local stable diffusion

@anhelmlife439 Yeah, I guess since it's an IOS app, it tends to blur out nsfw? That would make much more sense, I think

@yoyoyoyoyoa i guess the issue was with import somehow. I deleted the file and downloaded it again and it worked nicely

Yes please!

Train a textual embedding like this: tutorials/README.md at main · BelieveDiffusion/tutorials (github.com)

@jjpaq That was super useful. Thanks.

@jjpaq Thank you so much. Will try!

@saftle Thankss

as you pointed out, it evaluates to the same value. But with the first representation B-A acts like the delta. You can extend the first representation with multiple models, and m[i] would be the weights, you cannot elegantly present this otherwise.

@saftle 
Probably due to floating point rounding. Even if the two are not equal, they are practically equivalent.

Use the "realistic titfuck" LORA found on this site. If youre not familiar with using Lora's there are a few good guides on youtube to get it set up. its pretty simple.

@angusflooring458 yeah, I was using InvokeAI which doesn't support LORAs yet, so right now I'm making my own installation to get some LORAs because it's been a while and still no updates for InvokeAI. Thanks!

@saftle @saftle 

thank you now it works!

Works for me, you may have been trying to access it when Civitai was doing their stupidass updates that crippled the search and organized everything into their useless categories.

use DPM++ 2M Karras

https://civitai.com/models/135658/nipple-adetailer-model?modelVersionId=149588

LMAO

just add to negative prompt, tranny or long dicks.

maybe, "cold weather" to prompt?

It's never been a problem for me.

@saftle So does that mean you don't plan to upload 1.4 here then?

🤡

funnily enough urpm is great at beautiful faces with adetailer+loras. much better than popular alternatives

someone please train wedgies to any model plz <3

lets train that

what til you see what we are working on!

I have been using this as a training base for LoRA. I would suggest having a general "Sexy stuff" model like this and using swap-able and mixable LoRAS to achieve different results.
If you train a whole checkpoint, you could get one that was really good at "bondage" for instance, but it would lose it's ability to do other things for example food fetish.
With a "Bondage" LoRA and a well trained general model like this you can apply it when needed, but if you instead wanted to do food fetish stuff with hotdog buns and whipped cream you could swap it out for a "Food Fetish" LoRA and still retain the flexibility of the base model.
THEN if you wanted both, you can literally slap both LoRAs on there and get yourself some a lovely human platter bondaged to the table.
(You should be able to see why this is a bit better, full models/checkpoints should be used for large swath diverse themes while the details can be hotswapped in with LoRAs.)

how did you make a penis?

I'm also curious about this..

I downloaded and am trying to use it. Literally did the give her a hat prompt from the example image and it didn't do anything

8GB and it does literally nothing

У Меня другая беда, все сделал, все подключил. Но с промтами разобраться не могу) все время только нижнее белье прорисовывает) беда))

Poggies

on his discord he said, that he wont be aable to continue this model. So i think someone else need to step in

i should be able to take over on a new model for SDXL when i find a way to train it

@Commander1943 Whats stopping you?

@jacob42 SDXL models can't be trained yet only lora

@Commander1943 ? There's plenty of SDXL checkpoints on this site lol

also interested!

This is the way

@ubit34 any idea what the author is doing now?

why is this one Legacy?

@rogue8888992 probably cause the author is working for real money and has not time to update this project anymore.

@barlogspam176 Where can the new one be found/downloaded? Thanks

@barlogspam176 or what's his Patreon? 

@rogue8888992 he has a patreon? where?

@barlogspam176 where is he working for real money? do you have a link?

@rogue8888992 no. He mentioned many months ago on his discrod that eh found soemthing. No idea what

VAE is included in this model.

Use example prompts from the author and users gallery. Even then, expect to generate about 100 crappy images before you get something decent.

It's all in the negative prompts my man.

use a DPM++ 3M SDE Karras. Look up "ultimate guide to best photorealistic AI images & prompts" on google and follow it step by step.

I'm struggling too. Even with the most basic prompts (e.g. missionary), DPM++ 2M Karras was giving me absolute horrors like a woman with two butts and no face. I switched to Euler A after seeing that in a few of the good examples' info. It got MUCH closer to being okay, but I'm still looking for a better understanding.

use whatever you want to see

Try to inpaint the least area possible with a precise promt.

So if you want to clear up a hand with 4 fingers for example, draw around the middle two fingers and have "fingers" as promt. if that doesn't work use a larger area with "hand" as a promt for example

generally use less sampling steps and higher batchcount

darude - sandstorm

@saftle Am looking for work is there any remote jobs available

send us that

Send it if you can create penises just as easily with it.

You absolutely have something set wrong. This works fine.
Just play around with the settings until you get more what you want.

what model would that be?

Yo share the models ya dingus

But it can create penises.

The checkpoints I've seen so far can't do that.

what do you suggest then?

你所说的更好模型，请提供链接，让我们去判断。

To all askers, How about Pony models?

Guys, its pony, its always pony. Want animae waifu? Pony. Want realism? Pony_Realism. Want 2.5d? Pony styles. Want to take over the world? PONY!

what do you mean, morally outdated?

Better yet an SD3!

@Dragonbait SD3? No way, it's awful with nsfw