Update for v1.3: Full model/weight changelog at bottom of Description.
This is a free and open source model.
Model now includes workflow for generating images on ComfyUI with baked in Lora support and difuser preset. Check files section of this models page for the zip containing a showcase img with all the embedding for generation, and a json workflow used to generate the showcase images. Both should be drag and drop compatible for comfyUI.
Update: Devil Diffusion
To further support for Linux and Mac machines, i forked ComfyUI and built a nix expression out of it that should run out of the box on any Nixified machine, including a windows machine running wsl in a shell.
Currently, only supports AMD cards but NVIDIA cards will be supported soon.
Devil v1.3 along with IPTables and CLIP vision will be automatically pulled and installed alongside the new UI.
This Merge is a combo of many other models and ideas, most notably:
Realistic and semi realistic Cartoon
Realistic
Splash of anime for style and flavor.
Optimized to run on Linux. Runs well on NixOS.
Unknown if potato PC's can run this model, please give me some feedback if you got a potato, though i dont think its much of an issue.
Usage Tip: This model has been rigorously tuned mostly for Euler, aligned steps scheduler type, and high step counts.
Example.
Euler - Align your steps OR Simple OR Karras - step count 60+ for best results, more steps will refine detail in the background and around tags for "perfect" or "beautiful" or "flawless"
Works aight with other non Ancestral sampler types.
Not really tested much on Ancestral scheduler types, please post feedback on other scheduler types if you are interested.
Suggested Positive prompt:
For realistic, start with "hyper realistic" or "realistic" and it will produce good results.
For hentai/cartoon, use "cartoon" or "anime" style, and it will produce good results.
This model will also produce NSFW frequently.
This model also produces good quality smut so its a positive.
Simply putting a style tag such as western art style or realistic will generate the wanted output in the base model alone without any additional resources.
blue hair, spiky hair,sharp bangs, vibrant blue eyes, icy blue iris, short hair, sharp bangs, blue nail paint, painted nails, colored skin, skin-covered oni horns, black devil tail.
BREAK
standing, flowing black dress, mischievous smile, dim lighting, dark background, holding snowflake cupped in palms, hands out in front, devil tail visible
They dont always come out perfect or as intended, but the results have been consistent and good for me.
Im still working on figuring a way to generate multiple toons in the same photo interacting naturally, this is also the main focus of the models training in the future.
Using Loras with Devil v1.3
TLDR
Simply adding the Lora into your generation without style/art specific keywords is enough to produce the desired style even if it isnt realistic.
Works for realistic, cartoon, anime, and hentai Loras.
Long Read:
Typically when you add a Lora into prompt you sometimes also need some extra tags in the prompt to get the style correctly applied, depending on the model you are working with.
IE trying an anime style Lora on a realistic model normally produces shit results.
Devil is a bit more resilient when it comes to this, a lot of time and attention was taken to ensure the current weight set was flexible and fine-tuned for generating EXCELLENT quality images for a broad range of NSFW and SFW interests.
The blending of realistic anime + realistic cartoon + realistic porn makes Loras appear to act a bit nicer when they are simply added to a prompt like so:
solo,
1girl, mix of katara, meru, wrenchmeru, merunyaa, and sinon, 21 years old.
BREAK
view from behind,
wet,dancing alone in the rain, smiling, enjoying the cool breeze, head back, looking at sky, eyes closed in bliss, rain hitting face, rain splash on face, splash wet hair, steam rising from skin
BREAK
black gown,strapless gown, skin tight gown, icy blue hair, spikey hair,sharp bangs, vibrant blue eyes, medium length hair, sharp bangs, black nail paint, painted nails, red skin, colored skin, skin-covered oni horns, black devil tail
<lora:katara-book2-ponyxl-lora-nochekaiser:1>
<lora:Katara_Avatar_The_Last_Airbender:1>
<lora:sinon:1>
<meru_succubus_ponyxl_v1_unet_resized:1>
<lora:meruXL-08:1>
<lora:meru_succubus_ponyxl_v1_unet_resized:1>
Without any specific activators for the style you are trying to impart.
These 2 images: One is with Lora, and one is Without Lora were produced with this prompt.
ComfyUI Specific tip: Skip if you are not using ComfyUI
Using my workflow provided, adjust the CLIP strength and Model strength in the Loras individually for smoother blends, I'm not sure how to represent CLIP weight in a prompt directly so please DM me if you do know.
Ive noticed that setting clip higher and model weight lower in the Lora loader settings produces significantly smoother blended imgs compared to using the same value for both like you normally do by simply declaring it in the prompt like lora:meru_succubus_ponyxl_v1_unet_resized:1
To produce exact copies of the showcase images, use my work flow provided.
I also prepared the 2 mentioned showcase images workflow specifically with a frozen seed in the model files zip file.
img's also include the embedded generation data, so you should be able to use them on most UI's without issue.
The specific settings for producing the show case images linked below MUST use the workflow mentioned previously that's included with this version release, and are preset for the showcase images so you can see what I'm talking about.
The short answer is CLIP being adjusted UP with model strength adjusted DOWN still activates the tags needed to use the Lora BUT reduces many of the other side effects that come with Lora's in general, such as shit eyes and poor edges, melting clothes and skin, etc.
Long Read
Why: Loras are just trained image sets that more or less get patched onto a model during generation. They are a collection of both CLIP interpretations of Tags AND images. But the tags often times have multiple descriptors/tags to describe the same object/character, object/character such as red skin, devil girl, devil tail, etc which all have their own correlations the model pulls inspiration from to generate them without Lora. You would describe a devil girl as red skin, but you could also describe a red skin girl as a devil girl or demon girl, or (stretching a bit here) Native American/Indian for example, even though this is isn't necessarily the best descriptor or even the right descriptor, but its enough for CLIP to identify your idea well enough using the Lora CLIP to generate something desirable.
From what i understand and have experienced personally, this is due to how CLIP encodes and interprets tags in the prompt itself without taking the reference photos into account as much, just using the tags and weight modifiers introduced by CLIP from the Loras.
The Lora actually modifies the CLIP from the model during generation by merging in the CLIP from the Lora into the Model CLIP, this modifier is is the CLIP strength slider. 1.0 == Full weight towards Lora, 0.0 is Full weight towards model.
NOTE: This modifier can also be negative.
Lora Model Strength on the other hand, effects how much weight/inspiration the generation takes to produce an image, things like textures, color pallets, are styles, etc. and often the img output like your Lora img references, clothes and accessories included which may not be what you want depending on what you are working with.
Turning up clip independent of the Lora model strength pads the CLIP output with more descriptors and tags of the Lora and fewer images to draw on, fewer images to draw on usually ends up with better generations if the clothes, edges, eyes, feet, hands or face are all fucked beyond repair in your previous generations.
This is especially effective if you have a shit trained Lora or one that majorly apposes in style from what you are working with as a base model or other Loras that might not get along, but you still REALLY want to use a specific Lora.
This is a free and open source model expanding on the idea behind the devil SD 1.5 model i previously released but this time using Pony. Still under development.
This model will produce good excellent quality NSFW images, this model excels in realistic and semi realistic cartoon in particular, and does a fair job at artsy styles and Anime.
The main focus of this model was to produce a blend of elements that would produce quality images of Devil girls and Succubi in particular. Such as Meru the Succubus.
This will be an open source model, anyone can take and use this model, merge it, tinker, etc
I only ask that any and all merges, and/or content is held to the spirit of open source.
You may sell works derived from this model so long as you are not using another resource alongside it that disallows this.
I also ask that any merges made from this model also follow this guideline and be made freely available to anyone either here on civit, or somewhere else publicly available without restriction.
I want anyone to be able to use my models freely without fear of retaliation so long as the spirit of this request is upheld.
Though being credited in any works/merges yall might produce would be greatly appreciated, it is not necessary.
The only thing i ask in return is a star on some of my work and perhaps a tip. ;)
V1.3 Changelog:
Heavy weight modification to INPUT blocks 0-8, towards anime/drawn. These blocks are mainly responsible for generating character poses and background, 4-8 in particular handle latent compression.
Varying between 13-40% adjustment to weight leaning towards realistic anime/hentai.
Middle Block 0-1 fully merged towards anime/hentai, adds the feel at the edges but keeps core realism intact.
100% adjustment to weight leaning towards realistic anime/hentai.
Middle Block 2 -- slight weight adjustment towards anime, not really sure what this does by itself but seemed to be a positive change during testing.
36% adjustment to weight leaning towards realistic anime/hentai.
OUTPUT Blocks 0-8 -- Very minor adjustment leading towards anime/hentai, removed alot of doubling limbs, patchy skin, and poor lighting details in final output images and yet still retain core realism.
10-23% adjustment to weight leaning towards realistic anime/hentai.
'out.' block -- Unmodified.
No adjustment to weight.
CLIP ratio -- Moderate adjustment favoring hentai/anime esk style tags and tag interpretation. Makes Lora compatibility with Anime types more flexible.
58% adjustment to weight leaning towards realistic anime/hentai.
Time Embedding block -- Not really sure what this does exactly but without this aggressive tinkering all images came out looking like shit.
76% adjustment to weight leaning towards realistic anime/hentai.
Label Embedding block -- 78% merge towards realistic anime/hentai -- Not 100% sure what this does on its own, but sliding it with clip ratio seemed to produce more consistent images overall both with and without Loras.
78% adjustment to weight leaning towards realistic anime/hentai.
