This page holds realistic versions of eroDeity. For anime series / 这是艳神之仪的写实版模型页面。在这里下载二次元模型:[Anime] eroDeity | 艳神之仪
Experimental | 此模型为探索版
This means that they are not 100% stable. Offen generate bad anatomy, bad hands and bad feet. Since I want specific concept to be emphesized, other concepts out of scope were ignored and therefore will have poor quality. If you are looking for a 9/10 checkpoint then this is not for you, but if you are tired of stable experience (what a horrible idea) then please do try them out. I have a lot of fun merging them so I hope it bring you fun using them as well.
这意味着它的表现并不稳定,我在混合的时候只进行部分感兴趣的概念测试,因此你不能期待它能够在其他概念中也保持良好的表现。图片崩坏、肢体残缺可能是比较频繁的。如果你需要一个稳定不出错的模型,那么这不是你的最好选择。但如果你想要惊喜、喜欢更多的玩具,可以尝试,或对他们进行进一步融合来改善我没有改善的缺陷。
Use these settings | 使用说明
Please also read "About this version" tab. Or don't and tread wildly and have some fun.
"About this version"中还有每个版本的进一步说明,看不看都随意。
If you find an image you really like, but with multiple legs / mutated hands: Fix broken image by using Variation Seed, while keeping the original composition.
使用变异种子功能来修复图片错误,并大体保持图片原本的构图。

VAE:vae-ft-mse-840000-ema-pruned
Sampler | 采样器: 2M Karras or 2M SDE Karras
Sampling Steps | 采样步长: 24 on t2i and >26 on i2i
Clip Skip: 2
ADetailer on img2img | 在图生图后使用ADetailer修复脸部
ADetailer Settings

after img2img,check both enable and skip.

Detection set to 0.5

mask blur = 8, denoise strength <=0.4, check "use seperate steps" and "use seperate width/height".
Description
Adjusted the merging recipe basically merge from the very beginning. The chance of producing mutated hands, bad images and watermarks are reduced, however the legs are not that long anymore. It's all give or take, so you decide which version you want to use.
重新微调配方,降低了坏手、坏图和水印出现的概率,代价是什么呢,代价是腿没有x0那么夸张了。请自行取舍。
Feature: A slightly more charming face, long and slender legs, oilly skins and shiny fabric textures, curvy body.
特征:微调过的脸型,以非常低的权重融合了我以前训练过的脸模;修长的腿;油腻的皮肤和光滑面料质感;较夸张的身材曲线。
Limitations:
1. Hard to get good images, bad hands, watermarks, limited knowledge of many concepts from anime.
2. About LoRA: clothings, faces and background LoRAs works fine, however sex pose LoRAs don't work. You can use ControlNet to help with that.
3. General pose: from behind will generate bad anatomy, bad hands quite often. Wink (one eye closed) doesn't work (like many other realistic checkpoints).
4. Composition (Framing): close-up need high weight to work, but full body will work without issue, you just need to be careful with the prompting (e.g. if you are generating close up profile image, don't describe other part of body or clothings). 5. Zoom Slider LoRA doesn't work, it brake the image. See example image for horrible things (you have been warned lol)
缺陷:1. 坏图概率较高、手、水印、对二次元的概念了解有限;2. LoRA - 各种服装、脸模、场景等基本上能够比较好的支持,色色LoRA尤其是sex姿势基本上无法生效(Ahegao之类的表情是可以的);3. 姿势 - 背对姿势手部极容易残疾,wink(one eye closed)不生效(与大多数写实模型一样);4. 构图 - 近距离(close-up)需要拉大提示词权重,全身图仅需要写full body并正常描述从头到脚的特征即可,不可使用zoom-slider,会导致你的精神受到伤害。
- Positive prompts: DO NOT put a shit ton of quality tags in your positives and negatives. "masterpiece, best quality" is enough and put them towards the end of your prompts. I use this Wildcard Collection. Optional trigger words are: long legs, curvy, wide hips, narrow waist,
Negative prompts: (negative_hand-neg, Realisian-Neg:1.2), (low quality, worst quality:1.2), FastNegativeV2, NegfeetV2, ng_deepnegative_v1_75t, watermark, username, signature, logo, badhandv4, bad hands, missing fingers, fused fingers, extra legs, bad feet, malformed hands, mutated hands, poorly drawn hands, extra digits, fewer digits, sticks, wire, wrinkled skin, EasyNegativeV2
正面:不要使用大量的“画面质量提示词”, "masterpiece, best quality" 已经足够,将他们放在提示词靠后的位置。我使用这个通配符合集。可选触发词:long legs, curvy, wide hips, narrow waist,
负面:(negative_hand-neg, Realisian-Neg:1.2), (low quality, worst quality:1.2), FastNegativeV2, NegfeetV2, ng_deepnegative_v1_75t, watermark, username, signature, logo, badhandv4, bad hands, missing fingers, fused fingers, extra legs, bad feet, malformed hands, mutated hands, poorly drawn hands, extra digits, fewer digits, sticks, wire, wrinkled skin, EasyNegativeV2
FAQ
Comments (4)
现在很少有你这种原创又好看的底模了, 大部分别人的底模都是各种混合, 一测试跟之前下载的相似度99%
目前的缺点是不能很好的还原LORA的服装
混合的东西多了就会不稳定,大部分相似度高的模型都是为了稳定舍弃了一些多样化的元素。这个没办法彻底解决,除非自己训练。
第二版脚上全是关节
用这个 FEETMASTER - v1.2 Remaster | Stable Diffusion LoRA | Civitai Suggested Resource 里的NegFeet也试一下
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.






