(日本語説明は後半にあります)
About this model
[Updated v2.0 : Refined color, paintings, face]
"PansyBreed" is a SDXL(Pony) checkpoint, pursuing "kawaii", cuteness. This checkpoint has 2.5D art style.
Sister models are available:
VioletBreed : 2D style
RoseBreed : 2.2D style
I feel that model merging is similar to the breeding of flowers (to improve their appearance), so named "Breed".
These models have special features :
Include many originally trained LoRA's to adjust painting style (saturation, tone), eye shape, and quality (finely detailed pupils, finely detailed outlines, fine painting, reducing grain noise etc).
Highly detailed backgrounds.
Support NSFW outputs powered by Pony.
More various other LoRA's and Checkpoints are released on my CivitAI models list or HuggingFace repository. Please check it out.
Recommended output settings
CFG Scale: 7 - 10
Clip skip: 2
Sampler : DPM++ 2M SDE Karras / DPM++ 3M SDE Karras
Sampling steps: 30 - 50
Resolution: mainly tested on 768x768, 768x1152, 832x1216, 800x1200 and 1024x768
HiRes Fix: use SwinIR_4x, R-ESRGAN 4x+ or 4x-UltraSharp with denoise strength 0.3-0.7.
Negative prompt: Negative quality prompts are not needed especially.
VAE: v1.0A and later : embedded. (v1.0 : not embedded.)
Quality prompts sample
This and base checkpoints are derived from PonyDiffusion, so use quality prompts for Pony models.
newest, masterpiece, best quality
BREAK
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up
BREAK
cute round face, slender, ultra detailed eyes, ultra detailed hair, ultra beautiful
BREAK
(PUT YOUR PROMPTS HERE)このデータモデルについて
[v2.0：陰影や色彩、顔や眼の調整を行いました。]
PansyBreedは、2.5Dイラストレーション生成（Pony準拠のNSFW対応）に特化したモデルで、可愛さを追求しています。
姉妹モデル2種類も公開中です。
2Dスタイルの「VioletBreed」
2.2Dスタイルの「RoseBreed」
私はモデルのマージはお花の交配（品種改良）と近いものを感じているため、品種改良を意味するBreedと名付けました。Breedシリーズの特徴として、作者が独自に学習させた多数のLoRAをマージすることで、画質・色味・陰影や眼の形などを調整したり、輪郭線および眼の質感向上や粒子ノイズ低減、HiRes Fixの高品質化などを図っています。
推奨される出力設定や品質指定プロンプトなどは上記を参照してください。
VAEは焼き込み済みですので、別途用意する必要はありませんが、WebUIの設定でVAEを変更することで任意のVAEを使用して出力することができます。v1.0はVAEの焼き込みに失敗しているため、別のVAEを指定してください。v1.0Aでは修正されています。
他にもさまざまなLoRAやデータモデルを公開していますので、CivitAIの配布モデルリストまたはHuggingFaceリポジトリをご覧ください。
Models used to merge
Original trained styling/eye shape/painting/detailing LoRAs
XL_VAE_C F2 VAE
Description
FAQ
Comments (18)
its faded with a bunch of white artifact, while the other "breed" model works fine
Thank you for your interest in my model.
Strongly strengthened (or contradicted) prompts may trigger bad results.
Please try "(some element:1.5)" -> "(some element:1.2)" or "some element"
@JujoHotaru I'm seeing this same problem, even with short and simple prompts; I fixed it by manually including a VAE during generation. This model doesn't seem to have a VAE properly embedded.
@SecretH will give it another try with a vae
@SecretH wich one did you add?
@lunarpumpkin @SecretH Thank you for information. This model embeds "XL_VAE_C F2" VAE. You can use another VAE instead.
@JujoHotaru
Hello! Thank you for the amazing model!
I am encountering the same issue. The overall colors appear to be washed out. I am using ComfyUI with the recommended settings and prompts.
VioletBreed is working correctly.
@jj0808 Thank you for your interest. Do the prompts on my sample images work correctly?
And, please provide some prompts which does not work, then I'll try to investigate them.
@JujoHotaru
THX
I generated the image using the same settings (seed, prompt, etc.) as the one provided below, but the colors seem to be faded after removing the BREAK structure from the prompt. (Since I'm using ComfyUI, I can't use BREAK.)
positive prompt: newest, masterpiece, best quality, score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, cute round face, slender, ultra detailed eyes, ultra detailed hair, ultra beautiful, 1girl, solo, wearing a kawaii fashion outfit, outfit features a pink blouse with black lace trim and a large bow at the collar, The blouse is paired with a black high-waisted skirt with lace-up details, The person is also wearing black knee-high socks with lace tops, The overall look is cute and elegant, smile, open eyes, closed mouth, bag, sitting on bench, in a park, black hair, dark purple eyes, long hair
There is no major issue with the picture itself, but the colors have faded to a sepia tone overall, and there is a slight mix of white noise.
@jj0808 Thanks for information. I tried without BREAK, but still got good result on WebUI-ReForge.
Sorry I'm not good at ComfyUI, only WebUI tested...
Or, try with other sampling method ("Euler a" etc).
[to all]
I found that the VAE embedded in is model was broken, so sepia color and white artifacts generated.
I'm going to fix it and update. Thanks for your reports.
v1.0A updated. please try this version.
@JujoHotaru Thank you for the excellent work. I was able to confirm that it has been successfully fixed.
@JujoHotaru v1.0A works perfectly now! Thank you so much for all your hard work. Your models have always been some of my favourites.
When I use Fast LoRa (like TCD/Hyper-SD), the result looks more like a standard 2.5D anime model, rather than the usual style of your "breed" series. To get that distinct "breed" look, I have to use prompts from your SD1.5 breed series, like: "cute round face, slender, kawaii, perfect symmetrical face, ultra cute girl, ultra cute face, ultra detailed eyes, ultra detailed hair, ultra cute, ultra beautiful."