Type | |
Stats | 479 134,536 |
Reviews | (101) |
Published | Jul 3, 2024 |
Base Model | |
Usage Tips | Clip Skip: 2 |
Hash | AutoV2 F7C6480435 |
Please give revies and comments then I can make the model better.
PLEASE READ DESCRIPTION FOR A GUIDE + how to prompt, for best images
WHAT MODEL SHOULD I CHOOSE?
Vanilla: Kinda ok, OK SFW and NSFW
Sakura: good, good SFW and NSFW
Niji models: Pretty good, Pretty good SFW, Ok NSFW
Pear: Very good, Very good SFW and NSFW
Fused: Really good, good Lora and Lycoris support, Really good SFW and NSFW
Lemon Version A: Excellent, BAD LORA AND LYCORIS SUPPORT, pretty good SFW and NSFW
Lemon Version B: Excellent, Good Lora/Lycoris support trained on Nai or Anylora Great Lora/Lycoris support trained on fluffyrock, Excellent SFW and NSFW
Grapefruit: Seems to have the best so far SFW and NSFW (OUT OF DATE: only for previous models), Seems to be great with Loras trained on fluffyrock and the NAI leak/Anime_full
Clementine: has better style, versatility and NSFW compared to the last few models Plus, I'm using new base prompts for the best results, You can expect this model Very soon.
REBOOTED: The model which has tooken the place of all previous models. it has a few kinks, but when I add VPRED to it it should become better...
https://files.catbox.moe/oeax9f.gif
Boring RANDOM notes:
A continuation of my other model- CrossLawlasMix, but with more models mixed in selectively to make sure it would be the best quality AND I would be able to set the permissions to what I thought would be good. Also wanna find more of my models? Go here: HuggingFace, My-Profile.
NOTE: My HuggingFace does have more models that are not published here but most are experimental SO their names maybe misleading. Also I work on LLM stuff there mainly LLAMA models because they are easy to quantize when I can. Also I have been having troubles uploading files to hugging face so I guess no uploading for now :'(
VANILLA IS THE FIRST STYLE/VERSION meaning the next models MAY or MAY NOT improve on quality.
I Do encourage people to upload my models to other websites, the only thing I ask is that you link back to here.
Also Vpred models will have VPD or just VPRED in their name.
All of the images were created using Adetailer
Here is the HuggingFace Repo -> Here! <-
DOCUMENTATION ON MODEL (how to use the Model)
SECTION #1 Prompting:
Prompting for this model is quite versatile but using e621 tags will give you better results and using the negative and positive prompts here, Also UPDATE: The model supports Artist tags now!!! Yiffy mix tags to be specific.
(masterpiece, highest quality, 16k, high quality, best quality:1.3, highres, digital painting \(artwork\), furry, uploaded_on_e621:1.4), absurd res, RAW photo, (by zackary911, by canson, by theawesomefoxguy), by null-ghost, by s1m, (sharp focus:1), ((cel shading)), (realistic, kemono), BREAK
[PUT E621TAGS YOU WANT HERE]
And For the Negative prompt I recommend this:
low res, lowres, blurry, bad anatomy, letterbox, deformity, mutilated, malformed, amputee, watermark, signature, unusual anatomy, username, sketch, monochrome, (worst quality:2),(low quality:2),(normal quality:2),((monochrome)),((grayscale)),watermark, blurred, bad face, old, (cub:1.2, child:1.2, young:1.2),bad anatomy,(long body :1.3), BadDream, badyiffymix41, boring_e621_fluffyrock_v4, boring_e621_v4, easynegative, EasyNegativeV2, FastNegativeV2, fluffynegative, negative_hand-neg, ng_deepnegative_v1_75t, NGH, SimpleNegativeV3, updn, <lora:neg4all_bdsqlsz_V3.5:1>
Also Ps, you are going to need the negative embeddings, unless you want bad results.
I have also put all the embeds and Loras in the following link so you guys don't have to get them yourself: https://mega.nz/folder/8y0mFRBT#r2SaP2k-Gm__30AX74JhwQ
SECTION #2 Generation settings:
Higher Steps = better image
Lower CFG = more creative BUT more inaccurate
I suggest you use:
Steps: 40+ for best results, Sampler: DPM++ 2M SDE Heun Karras but DPM++ 2M Karras also works, CFG scale: 10
but if you want you can use:
Steps: 20-28 , Sampler: Euler a, CFG scale: 7
SECTION #3 Upscaling:
Upscaler: R-ESRGAN 4x+ Anime6B (Or really any anime style upscaler), Denoising strength: 0.3,