Some cover images are prompts created by you guys and alot from @GayPurpleSkunk (A pro at genning with furry models) If you notice you're prompt DM me for compensation!
WOW! You guys got both XXX-tended editions online from bidding! Y'all really like that series, eh?
Version VANTABLACK (READ ME)
Updated to MegaCock V2 and REED_FURRY_MIX
Color, vibrancy, and contrast are adjusted.
Prompt adherence is more sensitive
Clean, crisp, HQ output.
any blotchyness or blur removed.
Version XXX-TENDED v2.0 (READ ME)
This version uses an entirely different base model known as "Mega Cock"
it is appearently a checkpoint trained model focused on delivering highly detailed, cinematic, and creativity for NSFW, Fantasy and anime content. Still delivers that glossy, shiny, 2.5D look, but with more detail and better posing.
Anatomy & Poses - 9/10
Seems to get creative with poses, dynamic angles, shockingly good at getting full body or 3/4th body shots with good eyes.
Also appears to do duo very well with eyes, and full body.
fingers, mostly gives furries 4 or 5 fingers. Mostly 5, but here and there 4 fingers in springled in. Chance of finger failure is pretty low.
Prompt Adherence - 7/10
Great adherence, can get good results with a basic detailed prompt, but detailed prompting results in a finer tuned result.
Version XXX-TENDED (READ ME)
Base aesthetic cleaned up.
Better anatomy,
Less male gender bias.
I know you guys love using Yiff in Hell for the huge glossy boobs.
and this model amplifies that affect.For my gay bros, thick shiny asses and glossy thick cocks are locked and loaded.
Clean, Smooth, Glossy, as ever.
Awesome NSFW knowledge.
High Customizable via LoRA's for NSFW concepts
Backgrounds are about the same as V3/V4 ish.
Further Improved chances of perfect outcome without inpainting
Extremely strong prompt adherence
This means more detailed the prompt, the better the outcome
This almost understands natural language in a scary accurate way sometimes.
No need for prompt strenghts of (1.3) or higher. Recognized tags are already potent
You can choose to extend the power of each of these baked in LoRA's to finetune your results.
Rule34 V2
Pornmaster add detail
Gay Porn Extender
Gay Painterly Mix (Bara/Kemono)
All of these work harmonously if you dont want to add more of them to the model, and wont affect other lora's you choose.
Try out these embeddings too if you hate writing quality tags!
Recommended Settings
Steps: 24
Resolution: Portrait(832x1216), Square(1024x1024) Landscape(1216x832)
HiRes Multiplier: 1, 1.2, 1.5(best on A1111 and its forks)
Hires fix steps: 12
Sampler: Euler A(Auto, Beta, SGM Uniform), LCM,
CFG 2 - 4
Supports 512~1536 resolutions from its Illustrious 2.0 base
Recommended Positive Prompts
2.5D Photorealistic / Fuller Realism
masterpiece, ultra-HD, photorealistic, impressionism, high detail, depth of field, dynamic lighting, best quality, very aesthetic, 8k, best quality, realistic, muted colors palette. anthro
2.5D Semi-Realism with impressionism(painterly)
masterpiece, ultra-HD, impressionism, high detail, depth of field, dynamic lighting, best quality, very aesthetic, 8k, best quality, anthro,2.5D-3D (Ultra Quality)
masterpiece, best quality, amazing quality, very aesthetic, high resolution, ultra-detailed, 32k, UHD, absurdres, newest, scenery, furry, anthro2.5D Semi-Realism (High Quality)
masterpiece, best quality, amazing quality, very aesthetic, high resolution, ultra-detailed, absurdres, newest, scenery, furry, anthro,2.5D (Simple Quality)
masterpiece, best quality, newest, absurdres, highres, anthro,__________________________________________________________________________________________________
Version 4
Total modification of base aesthetic.
YES - the rule 34 v2 is still in this.
Suprisingly creative, good dynamic poses
Cleaner, Smoother, AND GLOSSY
Background details are slightly decreased
Backgrounds are more coherent and fitting.
Prompt Adherence is pretty solid.
__________________________________________________________________________________________________
Version 3
Big imporovements on base aesthetic.
Increased creativity and poses
Cleaner, smoother, 2.5D look
More detailed backgrounds
__________________________________________________________________________________________________
Version 2
One of the big things I wanted to announce with this version, is the addition of the all new Rule 34 dataset V2.
This will greatly improve prompt adherence, aesthetic and NSFW knowledge.
ALSO IF YOU NOTICE YOUR PROMPT USED IN THE COVER IMAGES DM ME FOR COMPENSATION.
READ ME!
Includes rule 34 dataset
Includes training data from top rated furry images on Civitai and e6ai.
Highly detailed backgrounds
Has really good knowledge of sex acts and positions.
Can do 2.5D all the way up to semi-realism.
Recommended Settings
Steps: 30
Hires fix steps: 15
Sampler: Euler A(Auto), DPM++ 2M(karras), DPM++ 2M SDE (Karras)
CFG 4.5
Resolution: supports 512~1536 resolutions
__________________________________________________________________________________________________
Description
FAQ
Comments (20)
You know, i didnt even know e6ai even existed before i found this model lol. Besides that i love it. Easy to use, cheap and hecken versatile. Besides that its easy to get my favorite type of girls (gynoids).
A new update…hm…may just have to check this out. Will update when model is released next auction. let’s see if this can defeat Banana Splitz.
Its pretty hard to say which would be better. Banana Splitz has a flatter, less 2.5D, less glossy look. Yiff in hell follows the extremely popular trend of glossy. 2.5D, highly detailed look.
the new Yiff in Hell contains a newer and more updated version of the R34 dataset. So maybe in the sense of NSFW content, yiff in hell would be superior.
But honestly its very hard to compare which is better, as they both have different intended uses.
Ok, This Update Looks Nice, Can’t wait for it to be out! You may have outdone yourself AGAIN!
It's really great.
Do models that include the Rule34 dataset still understand e621-style tags for prompting? From my testing, they seem to, but I’d like to confirm.
This inludes the newest and best R34 Dataset (version 2)
As for the question about e621-style tags for prompting. The answer will always be yes for every model that uses Illustrious(v0.1 and up) or NoobAI.
Basically Illustrious XL is the base(root) model that all the knowledge comes from.
When you see in the model description box on the right side on the page, look for if it says CHECKPOINT TRAINED or CHECKPOINT MERGE.
Checkpoint training is how we get models like illustrious, NoobAI, Hassaku, even Pony XL from SDXL.
Checkpoint merging is how we modify the weights on models. to change how certain prompts are treated or how certain styles are treated.
This is how we get models like Nova Furry, WAI NSFW, or even my own.
Checkpoint Training is like an expansion pack or DLC
Chekcpoint Merging is like a texture pack or mod.
Lora Training is like a miniature version of checkpoint training(usually adding one singular character, style or concept)
LatteLeopard Thank you for the detailed answer! I always write my prompts using only e621 tags with your models, as well as with Nova Furry and Indigo models, and I'm getting excellent results. But since we're talking about a new dataset here, supposedly trained on different tags (the Rule34 ones), I was wondering if its inclusion in merged models might affect how they understand prompts. But I suppose more knowledge is always good, as long as the base understanding is preserved and we can still prompt using the good old e621 tags.
(On a side note, since you mentioned WAI, that one seems to only generate furry art when using Danbooru tags instead.)
By the way, this one and Banana Splitz are my favorite models.
Hildale its actually very difficult for me to explain how the Rule34 dataset lora works, but it can absolutely notice a difference and it is extremely potent. It has to be merged in weights under 0.3. The normal wieght suggested by the creator is 0.1
anyhow I believe it not only introduces some tags that e621/danbooru doesnt use BUT! It also re-enforces their concepts strongly. you can prompt NSFW stuff so much better with it.
The model is created by iamddtla but he doesnt really understand how it works either....if you can believe that..... yeah...
Hey, could you advise on the best prompting method? I played with different ways I've seen in the gallery and on other sites (book like description, worded description, BREAK usage, parentheses etc.) My results greatly varies from ok to worse. Any suggestion would be greatly appreciated.
Start out with someting like this
masterpiece, best quality, amazing quality, very aesthetic, high resolution, ultra-detailed, absurdres, newest, scenery, furry, anthro,
when you prompt, every 50 tokens are grouped together. Some people like to neatly choose which tokens are grouped together so they used break.
its pretty much equal to starting a new paragraph when writting. if you were told you could only have 50 words per paragraph. Parenthesis are used to increase weight like this...
(orange fur) = 1.1 weight
((orange fur)) = 1.2 weight
OR you can do this
(orange fur:1.2)
(orange fur:1.4)
(orange fur:-1.2) yes thats right you can use negatives too.
You probably are new to illustrious type models. so I would also suggest you are using a lower CFG then nomral. (CFG should be around 2 - 5) I love to use 4 and 4.5
LatteLeopard Thanks man appreciate it, I'll give these a shot. If you could take a look on my posts (especially the last two) you can see how I prompt the images with 1 character. Usually that works (some generations get the features mixed up, but overall I get the expected result), however if I try with 2 characters, most of the time it'll end up the prompted features switched, or merged into 1 character (I think I posted a few example of the switches, the merged results aren't, I consider those low quality to post anywhere).
Earlier I checked the models page and inspired from the pictures (=how they prompted theirs) others posted, but since the filter system seems to be not working, I can't do it anymore...
Crovidius7849 So I gave some of your images and their prompts a look over.
1. Dont_do_this_for_tags just write the tags with no underscore.
2. Don't use "BREAK," on civitai generator. it will automatically split your prompt into multiple chunks if you go over 50 tokens.
3. the further away the character or the more characters the worse the eyes will likely be. This is pretty universal for SDXL, but ive managed to get it stay clear as possible when making this merge. but its not perfect :\
3a. A lot of the amazingly done photos you see in the user gallery are done in A1111/Comfyui or some other external generator. They have access to inpainting/adetailer which give them perfect eyes no matter the situation.
3b.Civitai users have access to a feature on the generater called "hi-res fix" I reccommend using this for complex images along with lora's and embeddings that aid with civitai generating.
This info applies to ALL SDXL models, everything I taught you here will work from my models, to any other illustrious/noobAI model.
I think I might write a guide for civitai generating users to learn how to generate online with comyfui/A1111 by using services like vast.ai, google collab or thinkdiffusion.
for $2 you can rent a server using an RTX4090 for about 4-5 hours on vast.ai
-it can be paused, data can easily be transfered to a new server you rent.
LatteLeopard Thanks, this will help a lot. I recently switched to local generation since I'm not a native english speaker, and the on-site generator sometimes flags the prompt because I used the wrong word for it (but I know that has the same meaning, for example: adding a word to describe a character's height, and when I used "small" it was flagged...). Anyway, I'll try implement these suggestions next time (I'm using ForgeUI, that one I was able to make it work on my system). I'm still at the stage of figuring out the basics, and after that I'll try to use inpainting and adetailer.
Amazing generation
Really like using this model for most of my gens, but new version gives me some visual artifacts like some weird wetness, grainess, bad eyes and fingers. Here is my post with images generated both on V1 and V2 with same prompts and seeds https://civitai.com/posts/21212707
I guess it could be the style lora problem, but that only happens with V2 so yeah
This is 100% linked to DMD2 Speed LoRA which has extremely random compatibility with most illustrious models.
also DMD2 Speed Lora was built on the original SDXL 1.0 architecture.
While it CAN work on Pony and Illustrious it is EXTREMELY experimental.
in both of those comparisons. it looks very fried on Yiff in hell V1 and V2.
V2 is more detailed so its amplifying filigree like patterns.
Basically your frying the model with these loras.
You might have some success with other illustrious models that are more basic in structure, but this is really works best for SDXL 1.0 models.
(Even though it advertises it as working with Pony and Illustrious)
You can also try lowering the weight of the lora. Maybe trying 0.5 or 0.4
If you are looking for a lightning like lora I would suggest looking for one that is Illustrious compatible. I believe they do exist. I am just not fimilliar with them.
I do suggest you try out SPO, its not really a lightning model, but it will guarantee super nice aesthetic. you can use weights between 0.5 all the way up to 2.5 with good results.
https://civitai.com/models/510261?modelVersionId=567119
Love using it!
Having now used YiH 2.0 for a while, can safely say it's my personal favorite for what I like to do with my gens compared with some of the other "big furry" checkpoints like Banana Splitz or Nova Furry. Overall it best captures the semi-realistic but still stylized kind of look I prefer, and cooperates very well with a multitude of Loras from both Illustrious and Pony.
I've also found it has a general better adherence to prompting and posing than some of the other checkpoints. Taking the same prompt and running it through various checkpoints, YiH tends to hit closest to what I'm going for.
It's not flawless of course, as it often by default gives things a a more "wet" look unless you put wet/sweaty in the negatives. I've also noticed it has a tendency to add little white paint spots around the backgrounds, but that could also be a result of my preferred Lora mix.
Still overall my favorite Illustrious checkpoint by far and having a lot of success with it!
Awesome review! Thank you!
I know it might sound wierd, but the overly shiny, glossly look was intentional.
It seems to be a popular trend, but I can dial it back a bit if you think its really over the top.
Btw, the spot issue is related to issues with how your program is handling VAE....I believe.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.