Type | |
Stats | 1,354 |
Reviews | (239) |
Published | Jan 25, 2024 |
Base Model | |
Usage Tips | Clip Skip: 2 |
Hash | AutoV2 304E8B143C |
I'll start by saying that there's nothing super special about this checkpoint, it's just a simple merge of other models, it just happens to be a mix that's very targeted on what myself personally want to get out of my image generation. The idea behind this mix was to get a model that generates mainly beautiful women, but also has a soft painterly anime style. It is able to generate both SFW and NSFW art and while the focus during merging was getting the women to look nice, it is also able to generate men reasonably well. It does have a bit of a problem with hands, but that seems to be a wide-spread problem with SD that I personally have no idea how to fix with merges.
That said, I did not cherry-pick the example images too much. Other than rejecting distorted hands, the most cherry-picking I did was selecting my preferred composition among some quickly generated low-detail images and then generating that seed with more detail.
V3 INFO
I noticed that the model has difficulties with some things, in particular furry/non-human characters. The v3 merges were done to potentially improve those things while retaining the painterly aesthetic. The credit list of models used in the merges has been updated.
THE MERGE
This was mainly a succession of trial and error merges because I lack the actual knowledge to more target merges or training. I started with an idea of what I wanted and tried merging in models in Supermerger that had one or more traits I wanted to see if I could get it to stick. My starting point was was Eagelaxis' Cetus-Mix CodaEdition, since it already had a lot of things I wanted in it, I just wanted to accentuate or change some elements. There were lots of merges and re-merges in the process so I lost track of the actual merge formula, but I kept track of all models I added in at some point so I could thank the authors. Go check their models if you haven't already.
CHECKPOINTS
T-R-手绘-illustration-2D二次元+LCM - 手绘-illustration
Kizuki - Anime / Hentai Checkpoint - Anime / Hentai V3
Ultra-Illusi0n 2D - Ultra Illusi0n 2.5D V1(OLD)
richyrichMix - richyrichMix-v2.fp16
LORAS
Detail Tweaker LoRA (细节调整LoRA) - V1
锐化/柔和效果LoRA模型 sharpen/soften effect LoRA model - v1.0
Unlimited Blade Works (UBW) - v1.6 UBW
[Style] Hand-painted Brush Stroke / Disco Elysium - v1.0
SashaKHMEL - Style - v1.0 (Pixai)
RECOMMENDED VAE
Feel free to try the model with your favorite VAE if you have one, but I have tested it mostly with kl-f8-anime2 VAE and Pastel Waifu Diffusion.
RECOMMENDED SAMPLERS
For a more painterly feel: DPM++ 2M Karras, Heun, Restart
For a more 2.5D feel: DPM++ SDE Karras, DPM++ 2M SDE Exponential
MY WORKFLOW
I have a potato of a GPU, so a lot of my personal workflow focuses on getting around that. The metadata on my images may or may not seem a bit odd. because of that This includes using Hi-Res Fix with an "Upscale by" value of 1 just to improve overall detail (together with the aDetailer extension) in the original low-resolution generation and later upscaling in img2img with the Ultimate SD Upscale extension so I can upscale my image in tiles my GPU can handle. I also use the CFG Scale Fix extension because I was finding very hard to get satisfactory results with the regular CFG configuration slider in A1111.
PROMPTING
I like to do my prompts with a basic sentence describing the scene and then using NAI-style prompting to detail things. I use the "oil painting" prompt in most of my generations to enhance the painterly feeling and I like to use the word atmosphere a lot in prompts to set the mood of an image.