The pursuit of perfect balance between realism and anime, a semi-realistic model aimed to achieve beautiful realistic faces with sexy hentai bodies.
This model aims to be flexible with all kinds of NSFW/hentai concepts, so a large portion of the merge will be hentai models.
Description
Even better faces than v3 and perhaps NSFW capabilities too, thanks to 50% of v4 being ChikMix v2 ;)
Face changing LoRA/TI might be redundant for v4.
Model tends to draw nsfw/females anyways so there's no need to try too hard for that. Face/style changes depending on prompts, feel free to experiment.
Model has some major side effects such as freckles/bad anatomy/hands, hard to draw dark skin/squatting, and other unknown effects. You may use v2 or v1 for a more stable experience.
Although not shown in examples, model can do males, scenery, etc. quite well too.
Use "freckles/mole" tag(s) in negative to reduce the dots on face/body side effect if needed.
All sample images were generated from nothing but prompts and the model.
FAQ
Comments (95)
How do I add a male figure? All of the nsfw loras I tried are missing male bodies :(
Not sure what you mean, but try hetero or 1boy or muscular male tags?
Thank you for your hard work. This model is excellent.
GLHF :)
Have the hands improved from V3 to V4?
Dunno, maybe, maybe not :^)
What is the difference between V3 and V4?
V3 和 V4 的区别是什么?
Check on the right "About this version", it's the same for every model on the site now.
Description on the left is the model introduction, "About this version“ on the right is notes for each version.
I'm using the latest v4 version, and my settings are the same, but why do my images always look too realistic without the blur effect seen in the examples? Thank you.
Hm.... shouldn't be, if you just PNG Info then "Send to txt2img" should give you a very similar result.
Can add me on Discord I can try to resolve your problem: Bloodsuga#6201
@Bloodsuga It was a problem with face repair I found the reason, thx !
@luldue OMG, NEVER USE RESTORE FACES EVER, NO MATTER WHAT!
我感觉女性阴部层峦叠嶂的有些过分了
小细节很难画的，凑活看吧 XD
加个专门的pussy Lora不就好了😎
不知道为什么，我用v4模型生成的图的人物脸部过于真实了，达不到示例的那种动漫效果
I don't know why, the character faces I generated using the v4 model are too realistic and cannot achieve the animation effect shown in the example
Certain prompts give certain styles because certain weights are biased, try other things.
Can add me on Discord I can try to resolve your problem: Bloodsuga#6201
@Bloodsuga好的，谢谢大佬，我再去试试其他tag和采样方法什么的
我用V4（不带vae）发现颜色会很奇怪，大家用的什么vae？
I used V4 (without vae) and found that the color would be strange, what vae did everyone use?
That make no sense. V4 was baked in 840000 vae so you can choose “None” for vae and the colors would look good. A vae is not necessary when a model has been baked with a vae, unless you want to try a different vae like kl-f8-anime2 or orangemix vae.
@Bloodsuga I found the reason, I always choose "default" before, this time choose "none" and there is no problem. thx !
@thevenomsnake Glad I helped :)
Hello, could you please explain the recommended setting of the picture size? When I tried to reproduce your results, I found that different picture sizes would produce different results
512 x 768 then x2 upscale with a latent denoise strength at around 0.6. I have no tried other resolutions or upscalers.
很奇怪，用了脸部的lora比如koreandolllikeness之后，v4版本跑出来的图中脸部会比较奇怪，而这些lora在前几个版本中都表现得很好
when I try t use the same model with the same prompt and params, instead of getting an amazing image, I get a horrible one. How does this work?
You most likely did not use highres fix. You should PNG Info then send to txt2img, that's how you replicate images.
For me the V2 version is still the best, i really prefer it over V3 & V4, this one is amazing
I copied all the generation data of the sample images but I can't produce them 1:1, did you change the eta too or is it not possible to produce the same images without the same graphics card?
Try using size: 512x768
For Upscale use Latent, Denoising strength: 0.6
If you still have issues try setting this in settings:
Clip skip: 2
VAE: None
But some of the posts from other users might deviate due to other factors outside of our control. For instance they could've posted using v3 instead of v4 or any number of textual inversions we might not have readily available etc.
It should be easier than ever to replicate images with the new webui overwrite features. Choose PW4 as your model, None for vae, you don't need to load anything extra. Upload the image to PNG Info, send to txt2img, it should load up all the prompts and settings, click Generate and you'll get it, because I didn't use anything extra but model and prompts. The result will not be exactly the same, even on my own computer I cannot replicate my own images EXACTLY because everyone uses --xformers now, and yes different GPU will produce similar but different results as well. I'm using MSI GAMING X TRIO RTX 4090 if you wanna go buy one and try to replicate better XD
血糖老师我身体受不住了🥵🥵🥵🥵🥵
多吃蛋白质，鸡蛋，鸡肉，豆腐，喝牛奶！
@Bloodsuga 有Discord讨论组吗 想来膜拜一下😋
@GoldenMiocola 我不知道，找别人要吧。想跟我聊的话可以加我 Bloodsuga#6201
@Bloodsuga 这个是哪个平台的账号鸭😲
@GoldenMiocola Discord.
为什么我画出的图偏暗呢？有什么解决办法吗？
i was wondering, where did you get so many hentai models to merge, and how many of them you have merged?
NSFW elements are mostly thanks to AbyssOrangeMixes. V4 has 50% ChikMix v2 so I don't know where those NSFW elements came from.
I like this model a lot, but i have 2 questions:
1. When you create it, have you used another models as base? i want to merge this model with AzoyvaRPG tools to make more fantastic themed images, so i want to see if it's better to sum or add diff the merge.
2. Any good lora you recommend to get a more painterly/traditional style or finnish to the image? you recommend some loras but i find them too anyme for my taste.
If you wanna create actual art this model is not recommended. Base model is probably AOM3, with some add difference, then mixed with ChikMix v2. I don't know anything about style LoRAs.
@Bloodsuga It's for a porn game, so i think it will service good
I'm an idiot and can not get it to use this style properly, everything seems like one step behind it. What is the keyword its looking for since I clearly input the wrong one
You probably didn't use highres fix. It's better to show an example image when you're asking for help.
Which sampler is recommended for this model?
Check the sample images, usually Euler a or DPM++ 2M Karrass.
Do i put the model in Stable-diffusion model folder directory or lora folder
This is a checkpoint, not a LoRA, so you put the checkpoint in stable-diffusion-webui\models\Stable-diffusion folder.
@Bloodsuga do you know how to fix this error
Traceback (most recent call last): File "D:\AUTOMATIC1111\Stable Diffusion\modules\script_callbacks.py", line 141, in model_loaded_callback c.callback(sd_model) File "D:\AUTOMATIC1111\Stable Diffusion\extensions-builtin\Lora\lora.py", line 145, in assign_lora_names_to_compvis_modules for name, module in shared.sd_model.cond_stage_model.wrapped.named_modules(): AttributeError: 'NoneType' object has no attribute 'cond_stage_model'
@awipslinuxuser175 All I can tell is this error stems from the built-in LoRA function of webui, I never had this error. You should probably ask people on the Civitai Discord or Stable Diffusion Reddit. Maybe a clean reinstall would fix it: https://www.youtube.com/watch?v=onmqbI5XPH8&ab_channel=KevinStratvert
Stable-diffusion
Lora goes in Lora folder, everything else goes in model. Summon loras with the commands explained on them. Loras can influence existing models
For some reason it just gave me prompts that doesn't make any sense, maybe there are some problems with my settings. Problems are like unbelievable ugly faces and also some fingers that don't make sense. Is this usually because my tags or my prompt settings? Hope can get a solution and thank you for reading my problem🥹
Put an example image that'll make your problem a lot clearer. You probably didn't use highres fix.
@Bloodsuga I'll try it, thank you very much🥹
negatif promts more important than promts it self.
@gambikules858 yeah man, agreed
Well thank you very much!
Your preview made me to leave my FIRST comment ever here...
and to turn the BLUR MATURE CONTENT switch right BACK ON.
You know you can just block me and view the rest of the mature content here.
Wow this model is even more variable than I expected at first ... you'll also get out some very muscular ladies without any lora, even made it to futa - broken about 50% but without any added lora a great result :D
Yes unfortunately that's an effect from the merge, same problem as AOM2H, I don't know if I can fix it in the future :/
@Bloodsuga Well I've seen it as positive as it mostly works fine :)
@keirdamone I was hesitating whether to release newer versions or not, but I think every nsfw enjoyer has tried v2 already, so I just released newer versions for people to try, but newer really doesn't mean better LOL
which upsclaer to use
Latent (bicubic antialiased)
Denoise 0.6 ish, if you're using Euler a I believe it can be as low as 0.5
@Bloodsuga hello, :) where do I specify that I would like to use bicubic-anti-aliased? i am using automatic1111. thank you. :)
@BigFatEvilCat https://files.catbox.moe/fist1s.png here :)
what upscaler do you use here?
You can just right click, save image to your device, then upload the saved image in PNG Info, you can see everything, same for all the images on the site that have metadata.
I used Latent (bicubic antialiased) at 0.5-0.6 denoising strength.
@Bloodsuga thank you very much sir
Any trick to for prompt to specific between anime, semi-realistic and realistic? sometime it hard to make specific one.
Might be long but these always helped me with my semi-realistic & realistic prompts.
((8k, RAW photo, highest quality, masterpiece), High detail RAW color photo professional photo, (realistic, photorealism:1. 5), (highest quality), (best shadow), (best illustration), ultra high resolution, highly detailed CG unified 8K wallpapers, physics-based rendering, cinematic lighting, photorealistic, photo, masterpiece, realistic, realism, photorealism, high contrast, photorealistic digital art trending on Artstation 8k HD high definition detailed realistic, detailed, skin texture, hyper detailed, realistic skin texture, best quality, ultra high res, (photorealistic:1. 4), high resolution, detailed, raw photo, sharp, d850 film stock photograph 4 kodak portra 400 camera f1. 6 lens rich colors hyper realistic lifelike texture dramatic lighting unreal engine trending on artstation cinestill 800), volume light, shadow
Here’s a few of mine for example, & these were all done using Pixai bc my ooor laptop crashed on me big time <|3
but doing a super descriptive prompt with the right model also helps a ton! These aren’t touched up or upscaled, just fresh off the generator 😅 I hope this all can be of help tho
I just did these 2 using only a kinda descriptive story prompt. If your not the best with that kinda stuff (bc I definitely wasn’t when I first started) there’s a few apps similar to chatGPT that are a huge help
Oh thank you for your advice. That is great help!!
PW has been one of my top 2 goto models. Right next to ReV <3
Why was it?
What happened to this model?
It means a "vae" has been mixed in with the model already, so a user can choose "None" for vae and colors will look good. A "vae" means a certain type of color correction, I baked the popular 840000 sd1.5 vae in the models.
@Bloodsuga Please tell me why this mozhel disappeared from the general list of models?
@ObiWan I guess it's because Perfect World is not popular anymore, you can sort by "All Time" and see it still.
Currently doesn‘t run on happy accidents
Can you tell me what has changed or improved in v4? I have been using v2 but besides bad fingers it takes a long time to load . Other than that it is perfect . Good thing about this model is you dont need to right a book of prompts like other models .
On the right side, "About this version".
Can we have thinner longer legs?
You can try, mayeb train a lora, but, most models like to draw THICC :)
try negative prompts to block out thicker body types
Use pos prompts like: long thin thighs,long thin legs,long thin calves. put neg prompts like: thick thighs, thick legs, fat thighs, fat legs. Adjust weight and emphasis until you get what you are aiming for.
很大呢❤️
Details
Files
perfectWorld_v4Baked.safetensors
Mirrors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
完美世界perfectWorld_v4Baked.safetensors
word_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
prfctwrld.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
Perfect_Worldv4.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
perfectWorld_v4Baked.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.