大家好，我是chosen鲷。更新了更稳定的V4.1版本，起因是因为我实在是受不了V4的手部崩坏问题了，出十张图能崩一半，剩下的也经常很抽象，实在忍不了了，所以用V4版本和之前的稳定版模型混合得到了这个V4.1版本，他比V4版本更稳定，手部崩坏概率较低，但是在某些画风上不如V4那么有风格化，大家可以根据自己的需求选择版本。具体支持的新画师串还是那些个，并没增加新内容，只是让他更稳定了。<s>更新了V4版本，在原有的V3版本进行加强训练而来，主要是增强了眼睛部分的细节，以及增加了更多画师串的支持，具体的请看我的例图。不过V4版本的稳定性上面确实不如V3版本，有时候会出现手部崩坏的情况，后续如果有机会我会去改善这个情况，这里也提供一个V3.5版本，这个版本更稳定，但是支持的新内容不如V4多，大家可以根据具体需求去使用不同版本的模型。</s><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/40c80243-1a46-44c4-abe4-8244bf36c4b6/width=525/40c80243-1a46-44c4-abe4-8244bf36c4b6.jpeg" /><s>是我自己使用的一个融合模型，主要是用noobal 0.5和一些其他的乱八七糟东西融的，经常用来和我训练的一些lora搭配使用，和noobai一样，画师串和人物可以直接识别，但是具体效果和nai3有些微妙的不同，不加任何lora，出图也挺稳定的，手基本没啥问题。</s><s>新的融合模型，用了新版的noob和wai nsfw,还有一些其他的模型，和1.0一样，出图很稳定，人物、画师串都可以直接识别，手也挺稳定的，和我的各种lora适配性也很好。这个模型主要升级的地方在于，V1版本在厚涂，以及韩风方面的效果不好，现在的V3版本对这两种风格的效果进行了优化。你可能要问了，为啥没有V2版本直接V3了呢，其实是因为我犯懒没发V2，现在有V3了，所以也不需要V2了。</s>具体的tag可以看我例图。感谢大家的下载使用，如果喜欢的话欢迎关注我，我的bilibili pixiv ID均为chosen鲷，欢迎大家关注我的频道<a target="_blank" rel="ugc" href="https://space.bilibili.com/3867754?spm_id_from=333.337.0.0">https://space.bilibili.com/3867754?spm_id_from=333.337.0.0</a> <a target="_blank" rel="ugc" href="https://www.pixiv.net/users/102907778">https://www.pixiv.net/users/102907778</a> 谢谢。The following content is from machine translationHello everyone, I'm chosen鲷.The more stable V4.1 version has been released. The reason for this update is that I simply couldn't stand the hand deformation issues in V4 anymore. Out of ten generated images, half would have broken hands, and the rest were often highly abstract. I couldn't tolerate it any longer, so I mixed the V4 version with previous stable models to create this V4.1 version. It is more stable than V4, with a lower probability of hand deformations, but it may not be as stylized as V4 in certain art styles. You can choose the version based on your needs. The supported new artist prompt sequences remain the same—no new content has been added; it's just been made more stable.<s>This is a merged model I created for personal use, primarily blending noobal 0.5 with some other miscellaneous elements. I often use it in combination with the LoRAs I've trained. Like noobai, it can directly recognize artist strings and characters, but the specific results have subtle differences compared to nai3. Even without any LoRA, the output is quite stable, and hands are generally well-rendered.</s><s>The new merged model incorporates an updated version of noob and wai nsfw, along with some other models. Similar to version 1.0, it produces stable outputs, directly recognizes characters and artist strings, and maintains good hand stability. It also has strong compatibility with my various LoRAs. The main improvements in this model are in addressing the shortcomings of V1 in thick-painting and Korean-style art—the current V3 version optimizes performance for these two styles. You might wonder why there's no V2 and we jumped straight to V3—well, it's because I was too lazy to release V2. Now that V3 is here, V2 isn't needed anymore.</s><s>For specific tags, please refer to my example images.</s><s>Thank you all for downloading and using it. If you like it, feel free to follow me. M</s>y Bilibili and Pixiv IDs are both chosen鲷. You're welcome to check out my channels: <a target="_blank" rel="ugc" href="https://space.bilibili.com/3867754?spm_id_from=333.337.0.0">https://space.bilibili.com/3867754?spm_id_from=333.337.0.0</a> <a target="_blank" rel="ugc" href="https://www.pixiv.net/users/102907778">https://www.pixiv.net/users/102907778</a> Thanks!

Chosen-mix_XL

chosenMixXL_v41.safetensors

Files with SHA256: bc3d335767c805a754eb4c0a3a039820b589d05fff9a586fde6418a19e2f66ef