The sample images are generated by my friend "聖聖聖也" -> his PIXIV page.
I have set my models forbidden to be used for commercial purposes, so please report to me if you find anyone using my models for commercial purposes.
"I would like to express my gratitude to Philip_ovo for helping me publish the model on Furry Diffusion Discord. If possible, please also support the Lora model MJFurryLora And NJFurryLora created by Philip_ovo."
This English version of the description has been modified based on the translation by marbrdf on ChatGPT.
This model is designed to generate kemono-style furry using booru tags and can generate human-like figures without decreasing the quality of the output. The model is designed to produce images without blurring the background.
See the model description for more details.
Models are named alphabetically based on the order in which they were uploaded, and the later models are not necessarily better. Choose Model A, B, C, D, E, F or G based on your needs. This is similar to orange and 7th, and you can intuitively compare them with reference images to see the differences.
If you only want to use a short prompt without adjusting the weights and without using the kemono-style Hypernetwork or Lora to generate kemono, choose Model B.
If you have a basic prompt foundation and can adjust the weights but do not want to use the kemono-style Hypernetwork or Lora to generate kemono, choose Model D.
If you want a very different style and can accept using the kemono-style Hypernetwork or Lora to generate kemono, choose Model C.
If you want to generate attractive human-like figures and kemono-style furry that are close to the current mainstream anime style, and can accept using the kemono-style Hypernetwork or Lora to generate kemono, choose Model A.
If you are looking for a more realistic texture similar to AOM2 or AOM3, and can accept not using the kemono-style Hypernetwork or Lora to generate kemono with a more realistic painting style, then choose Model F.
If you want clear lines and celluloid-style coloring, and are willing to accept that the characters generated by Model E may appear very young and petite, then choose Model E.
If you are looking for a highly impasto coloring style and incredibly detailed and accurate scenes with a rich and colorful palette, and are willing to accept that the generated kemono poses in Model G may have a more fleshy feel, such as a more visible navel, then choose Model G. Please note that this model requires the use of kemono-style Hypernetwork or Lora to generate kemono, as otherwise, unnecessary human ears and incorrect animal noses may be generated.
This series of models combines five Lora styles I trained myself with five kemono-style furry Lorastyles in a hierarchical U-net model that I repeatedly fine-tuned through layer merging.
Like Counterfeit and Pastel, this model uses a hierarchical U-net model combined with Lora. However, because it includes two completely different species with different body structures, humans and kemono furry, in the merging process, its stability is not as good as that of a general model.
The models used in the process include but are not limited to the Anything series, Orange series, Pastel+ (Yohan series), Counterfeit series, and 7th series, totaling more than 50 models combined, plus my own personal Lora model, so I cannot share the recipe.
Although this series of models can generate NSFW content, it is recommended to use any Lora related to NSFW for overlaying. The example images do not use Lora related to NSFW, but overlaying any of them will significantly improve the quality.
This series of models was generated with the default HI-RES setting, and I am not sure how much the image quality would be affected without it. However, based on the images generated by others in the discussion area without HI-RES, the results seem to be good.
I will not upload more example images because the style is very diverse, combining more than 50 models, including some realistic style images created by others in the discussion area.
2023/2/24 CrosskemonoC
2023/2/24 Deleted most of the example images, including NSFW, and replaced two images. Following Civitai's rules, example images with minors and NSFW are not included.
2023/2/24 CrosskemonoD
2023/3/1 CrosskemonoF
2023/3/7 CrosskemonoE
2023/3/8 CrosskemonoE_2
2023/3/13 CrosskemonoG
2023/3/14 CrosskemonoG_2
2023/3/23 Crosskemono2.0TEST
2023/3/24 Crosskemono2.0TEST_2
2023/4/1 Crosskemono2.0
2023/4/17 Crosskemono2.5
這個模型是以活用booru tag簡單生成kemono 風格的獸人，同時又能在模型中生成人類卻不降低生成質量，並且保持背景不會模糊的前提下製作的。
詳見型號說明。
按照上傳順序對模型以英文字母順序命名型號，並不是越後面上傳的模型就越好，請根據你的需要選擇A型、B型、C型、D型、E型、F型、G型，這有點類似orange以及7th，你可以從參考圖像直觀獲得對比，每個型號都有較大的差異。
如果你只想用簡短的prompt，不調整權重，也不使用kemono風格的Hypernetwork或者Lora來簡單生成kemono，那麼請選擇B型號。
如果你有一定的prompt基礎，可以接受調整權重，但不使用kemono風格的Hypernetwork或者Lora來生成kemono，那麼可以選擇D型號。
如果你想要一個很不一樣的風格，同時又能接受使用kemono風格的Hypernetwork或者Lora來生成kemono，那麼請選擇C型號。
如果你想要和目前主流的anime模型接近的畫風和構圖，想生成好看的人類又想生成kemono風格的獸人,同時又能接受使用kemono風格的Hypernetwork或者Lora來生成kemono，那麼可以選擇A型號。
如果你想要類似AOM2orAOM3的那樣更加立體的質感，又能接受不使用kemono風格的Hypernetwork或者Lora來生成kemono時畫風較為寫實，那麼請選擇F型號。
如果你想要清晰的線條的賽璐璐上色，又能接受使用模型E時，該模型生成的人物會顯得非常年輕和嬌小，那麼請選擇E型號。
如果你想要非常厚塗的上色風格，以及非常細緻且正確的場景，非常豐富多姿的色彩塗抹，又能接受在使用模型G時生成kemono的時候生成的姿態會非常具備肉感，也就是肚臍可能會比較明顯，那麼請選擇G型號。(這個型號需要使用kemono風格的Hypernetwork或者Lora來生成kemono，否則可能會生成不必要的人類耳朵和不正確的動物鼻子)。
這個系列模型是將我自己訓練的5個風格lora+5個kemono風格獸lora合併到我使用階層U-net合併的模型內，然後反復階層合併微調的模型。
和 counterfeit 以及 pastel 一樣 使用 階層U-net合併+lora合併的模型，不過因為在合併過程中加入了完全兩極的人類和獸人兩個身體構造不同的物種，穩定性較一般模型要差，望周知。
中間使用過的包含civitai和hugging這2個網站上有名的模型合併
包括不限於
anything series
orange series
pastel +（yohan series）
counterfeit series
7th series
等合計50多個模型合併，加上Lora模型為個人私享，所以不能分享食譜。
這個系列模型雖然可以生成nsfw,但推薦搭配任何跟nsfw有關的lora疊加使用，示例圖像沒有疊加nsfw相關的lora使用，實際上疊加任何一個都有很大的改善。
這個系列模型是在默認開啟 HI-RES 的情況下來生成的模型，我不清楚不使用HI-RES會對生成的圖像質量有多大的影響，但是看討論區其他人在不使用 HI-RES 的情況下生成的圖，感覺還不錯。
我不再上傳更多的示例圖像，因為基於我自己的模型並合併了50多種模型的緣故，風格非常的多樣化，評論區上也有人用來創作寫實風格的圖像，所以請參考評論區其他人創作的圖像。
2023/2/24 CrosskemonoC
2023/2/24 刪除大部分示例圖像，包括NSFW，並替換了兩張圖片，遵循civitai的規定在示例圖像上不放入包含有 minors NSFW 的內容，所有示例圖像的prompt 均不包含 loli or shota 的內容。
2023/2/24 CrosskemonoD
2023/3/1 CrosskemonoF
2023/3/7 CrosskemonoE
2023/3/8 CrosskemonoE_2
2023/3/13 CrosskemonoG
2023/3/14 CrosskemonoG_2
2023/3/23 Crosskemono2.0TEST
2023/3/24 Crosskemono2.0TEST_2
2023/4/1 Crosskemono2.0
2023/4/17 Crosskemono2.5
Description
+Better background.
+Better genitals, knot, sheath.
+Better E621 Furry characters and Furry artists support.
+Better E621TAG and Danbooru Tag.
+Better composition and fingers.
Other content is consistent with 2.0Test/2.0Test_2.
Thanks 4Chan / Trash / provided various prompts for reference.
FAQ
Comments (28)
Hi can you make version 2.0 with vae-ft-mse-840000-ema-pruned built in???
There is a reason for me not to bake VAE to Model, because VAE can change the color difference to a certain extent, and obtain different color comparisons by specifying different VAEs by options.
@toynya I get faded colors.
I can't connect VAE separately
Need a version with built-in VAE 😅
@kromarti Do not specify VAE as "auto", because VAE does not have the same name as Model, so it does not take effect directly when the auto is specified by default. Please specify the VAE of Crosskemono directly on the option to correct the VAE correctly.
@kromarti If you want to make it not fade without changing the setting, you only need to name the VAE name as exactly the same as the Model, so that it will take effect without changes.
@toynya I understand.
If I used the AUTOMATIC1111, then there would be no problems.
But I use the lite version (because the computer is weak) where you can only use baked wae.
@kromarti In https://github.com/camenduru/stable-diffusion-webui-colab notebook, the countdown second line adds this line
!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://civitai.com/api/download/models/32830 -d /content/stable-diffusion-webui/models/Stable-diffusion -o crosskemonoFurryModel_crosskemono20.safetensors
Then run it with one click, and then enjoy the full version on your computer.
@toynya Ok, thanks
the best series of models at the moment for SD!
Thank you, I think the models of other authors are also very good.
Alright quick lil comment here, how do I get rid of deformaties I get sometimes such as muliple limbs and body parts and bad hands?
Use Inpointsketch, and then repairs it with similar colors to fix it.DPM adaptive or DPM++SDE Karras ，Denoising strength: 0.25~0.3.During repair, only the prompts in the area are retained.
is it possible to get the vae in a different format? i use Shark (for amd cards), and it doesnt seem to know how to handle .pt files. i need a ckpt or a safetensor i can run to generate a vae it CAN use, or a vae that's a .bin
I'm sorry, I don't know how to convert vae to ckpt or safetensor, if you know, you can tell me how to do it.
important Notice2!
There was a problem with CIVITAI image review system, so the images uploaded by everyone from 4.4 to 4.6 disappeared.
The review system is being repaired to restore as many pictures as possible, not that I deleted the pictures you uploaded, I hope you understand.
上周我在這裡詢問關於如何獲取更高質量背景的方法，謝謝你的回答。
我根據你的建議使用Multidiffusion來分塊生成角色和背景。 整體提示詞只描述鏡頭和背景以及開啟hypernet，前景輸入角色提示詞，背景不輸入提示詞。 去噪過程前幾步可以正常生成kemono，但最後Hires fix將角色變成了人類，我嘗試降低Hires fix的Denoising值，但它會導致影像細節更少更模糊，提高Denoising會使角色更像人類。
我不知道這種情況下我應該怎麼做
你使用的是G系列的型號嗎？G系列的型號比較接近人類，它與更多的Lora兼容，但因為嘗試採用了Lora塊合併的新技術，對kemono的程度覆蓋不全面，hires會單獨對out層提取碎片然後生成，在out層我覆蓋的kemono程度在這個型號裡只有55%。
你可以搭配NJfurry lora 或者furrygirl lora這2個其他人製作泛化kemono的Lora，0.3~0.5的強度下微調模型。
或者使用 LLuL 擴展 單獨選擇臉部區域來細節增強,或使用 ddetailer 擴展 單獨使用臉部遮罩細節增強。
@toynya 我嘗試用MBW將G2與2.0合成（OUT以外的層取0），但只要經過Multidiffusion就會生成人類角色。
原版G2只要開啟Multidiffusion，無論Lora設定多少權重都無法生成kemono。LLuL似乎無法解决尾巴缺失的問題，我已經在聯系Multidiffusion的原作者
@Roro_ 嘗試Negative prompt:(human, human skin, human head, human hands,human ear,human ears:1.0)
也許，它可能會發揮作用。
@toynya 提示詞沒有起作用，提高權重只會生成場景的圖片。
我使用2.0模型開啟Hires fix可以正常與Multidiffusion互動，但D系列和F系列也會出現與G系列相同的問題(無論是否使用hypernet)，我完全沒有頭緒，你知道可能的原因嗎？
@Roro_ 我研究了很久還是搞不明白為什麼會這樣，真的非常抱歉┭┮﹏┭┮
@Roro_ 這是今天剛出來的新修臉擴展 https://github.com/ototadana/sd-face-editor 我還沒有試過，不知道可以不可以幫助到你
@toynya 我聯系到multidiffusion的作者了，最後發現這個問題是webui本身hires的bug。 我在t2i裏關閉hires然後將模糊的圖片送到i2i進行手動hires，應該算是解决了
@toynya 在t2i介面中hires的bug會導致hires過程沒有讀取multidiffusion裏的提示詞，而G系列模型在不輸入任何角色提示詞的情况下會默認生成人類
@Roro_ 原來是這樣，你有找到解決方法真的太好了！
Nice Model. I see some others mixed in.
Just noteing that you should read up on copyright law xD. As with everyone out there. Not many get it + the License from SD read it again.
Thanks for the heads up, but I don't get any commercial benefit from this, I don't run PIXIVFanbox, Patreon, Ko-fi.
As far as I know, many mod authors have Ko-fi donations, or Patreon backers unlocked, etc., and that includes PIXIVFanbox.
This is common outside of furry models, and my agreement with 聖聖聖聖 does not include any monetary benefit for me from him.
I only did it because he has been using my models for a long time and is very knowledgeable about it. So I wanted him to be responsible for future preview images of new releases, in exchange for my allowing him to commission or sell images generated from my models without paying me anything, if you have seen his PIXIV you will understand why I say this.
I just want more people to use my models better, so I let the best users generate images as previews, that's my original intention, I don't expect you to understand, but I want you to understand what I mean.
What Greenfox probably meant is that you can't just overrule CreativeML Open RAIL-M license (that is enforced on your mix due to it being a derivative of SD and its mixes) and claim that you "prohibit commercial use", because CreativeML Open RAIL-M allows it and enforces "share-alike" licensing.
It's not about you making any money from the mix, that would be perfectly fine according to the original license.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.