StablyDiffused's AniMagic
詳細
ファイルをダウンロード
モデル説明
Introduction
Wow... It's been a while since I uploaded anything. I've been working on this project for a long time. It has been scrapped, redone, retheorized, rebuilt, more times than I'd like to count. This really shouldn't be version 1.0, but it is the first one that I've released so.... The goal with this model was to build something that was very NSFW, but had an animation style that was similar to RealCartoon-Pixar. RealCartoon-Pixar is capable of NSFW, but not the greatest. it's CLIP just doesn't understand some of the more naughty terms that I wanted the model to be able to do. It has some really cool aesthetics though and unfortunately some of that gets lost in this merge. But the merge essentially built up a lot of different NSFW models (and some others) attempting to improve different sex positions, anatomy, and etc while adding in some good background aesthetics and that sultry, seductive style that I like to get from my models. Then after building that up I merged RealCartoon-Pixar on top using the CLIP from my base model and a MBW Merge using the WRAP16 preset pulling in the outer layers from Pixar. Voila! it worked. There are still some things I'd like to improve here, and I contemplated not releasing this one, again.... Ultimately I decided to release this model though because there is a lot of really good stuff in it. All Models used are credited below.
This model is very NSFW inclined. It also tends to prefer putting the subject in a pose where they are facing away, looking over shoulder, and sticking their ass out (something I'd like to improve in a future version). If you don't want this you'll likely need to prompt it out by adding terms that encourage the subject to be viewed from the front angle. I designed it for NSFW and it is good at that, so most of the Example images will be NSFW ones. But I included a couple that are not just to show that the model is capable of more than just NSFW.
Disclaimer
All prompts in the example images use InvokeAI Syntax NOT Automatic1111 syntax. The key thing you need to know is that 'term+' is equivalent to '(term)' ++ is the same as (()) and so on. All example images were generated in InvokeAI with xformers enabled. InvokeAI also converts all safetensors or ckpt files into diffusers models before generating images. For these reasons you will likely not be able to reproduce my example images 1:1.
Prompting & Generation
Scheduler: Euler
Steps: 20 - 30
CFG: 8 - 12
CLIP Skip: 2
For Using HiRes Fix I'd recommend using GFPGAN. 4x or Anime should work fine. Invoke doesn't let me change which version I can use unfortunately but I did test 4x in Auto1111 and it works pretty well. For Strength I'd recommend 0.45
Definitely recommend using the following TIs in your Negatives. All my example images use all 3 of these.
I tried EasyNegative and BadHandv5 as well, but these ones just seemed to give me the best results. I use them in the order they are listed here and it works pretty well.
Pretty much all of my prompts start like this. Then I fill in the blank with what I want to see and it works very well:
Positive Prompt: (masterpiece)+, (best quality)+, highres, high quality digital art of _____, high contrast, soft lighting, backlighting, bloom, light sparkles, chromatic aberration, smooth, sharp focus, (detailed)+, 4k, 8k uhd, (perfect anatomy)+, sharp, (intricate detail)+, (detailed background)+
Negative Prompt: boring background, plain background, censored, (worst quality, low quality), greyscale, monocrome, multiple views, blurry, text, signature, <bad_quality>, <ng_deepnegative_v1_75t>, <negative_hand-neg>
Eyes: If the Eyes look washed out, make sure you mention them in the positive prompts. you can add things like 'blue eyes, detailed eyes, looking at viewer' of course use whatever color eyes you want.
VAE is baked in and comes from one of the other models used in the merge. Not sure which VAE is used, but it looks nice and produces vivid colors with good saturation so I'm not worried about it. In the toolkit it lists as SD v2 VAE
This works very well with many of the LoRA's from Kojimbomber and I'd suggest you try some of them out. No LoRA was used in any of my example images, only the aforementioned TIs.