Niji Style XL

v2.1版本介绍： 这个版本对于2.0版本的一些bug进行修复，2.1版本的手部表现更好，减少肢体bug，同时这个版本融合了琥珀佬Epsilon_rev2，同样的对于epsilon rev2的其他内部画风没有污染，可以使用其他画风串，大家在使用这个版本的时候一定要加上触发词啊！impasto和flat coloring！此模型仅用作交流和学习，不可商用，在各个平台开放下载，如果融合请标注原模型以免不必要的麻烦。 Version 2.1 Introduction: This version addresses some bugs from 2.0 and improves hand performance while reducing limb bugs. Additionally, it integrates kohaku Epsilon_rev2 without contaminating other internal art styles of Epsilon rev2, allowing the use of different artistic styles. When using this version, be sure to include trigger words like "impasto" and "flat coloring"! This model is for communication and learning purposes only, not for commercial use. It is available for download on various platforms. When integrating, please attribute the original model to avoid unnecessary complications. v2.0版本介绍： 更新了训练集分开练进去了厚涂（impasto）和平涂（flat coloring）两种画风，同时基本没有污染琥珀佬Epsilon_rev1里的其他画风（其他画风词去琥珀佬模型介绍查 ），由于是基于琥珀佬Epsilon微调出来的模型保留了其内部3600个人物的数据，此模型在手部准确度上稍差，请您在使用时放弃掉那些写了一大堆质量词的tag，masterpice和best quality 已经足够了多写反而会增加炸图的概率，采样器推荐使用DPM++ 2M Karras，步数给到30即可，其他的大家自行体会吧！ v2.0 Version Overview: This update focuses on separately training for two painting styles: impasto and flat coloring. It minimally interferes with other painting styles present in Kblueleaf Epsilon_rev1, as explained in the model documentation (for details on other styles, please refer to the Kohaku XL Epsilon model introduction at ). Derived from fine-tuning Kblueleaf Epsilon, this model retains the dataset of approximately 3600 characters. While there is a slight decrease in hand rendering accuracy, users are advised to refrain from using excessively detailed quality tags when utilizing the model. Tags like "masterpiece" and "best quality" are sufficient; adding more may increase the risk of producing undesirable outputs. For sampling, it is recommended to use DPM++ 2M Karras with a step count of 30. Feel free to explore other settings as per your preference! V1.0版本介绍： 训练集为niji6出的图，练这个模型的初衷是还原一些niji才能出的风格，目前尝试很多次以失败告终，最后决定先练一个niji的基础风格，后续的构图视角包括多画风用lora来“打补丁”，这个模型更新版本不会出现多概念多画风的版本，可以省省大家的硬盘空间，训练底模用的是animagineXL base，遵循animagineXL的条款，此模型仅用作交流和学习，不可商用，在各个平台开放下载，如果融合请标注原模型以免不必要的麻烦。 V1.0 Version Introduction: The training set consists of images from niji6. The original intention of training this model was to replicate some of the styles unique to niji. After many unsuccessful attempts, it was decided to first train a basic niji style. Subsequent composition perspectives include using Lora to "patch" multiple art styles. The updated versions of this model will not feature multiple concepts or styles, thus saving everyone's hard drive space. The base model used for training is animagineXL base, following animagineXL's terms. This model is strictly for communication and learning purposes and cannot be used commercially. It's available for download on various platforms. If mixed, please attribute the original model to avoid unnecessary trouble. 使用说明： 1.XL的出图分辨率要达到1024左右，此模型建议使用928 1232（3：4）直出加放大 2. masterpiece,best quality,一定要加上哦 3. 采样器使用3M系列 步数35左右即可 4.XL的模型不需要添加vae 很多小白出紫色噪点图都是因为加了1.5的vae 5.关于负面提示词 不需要写太多的负面提示词我个人在用的就是 (worst quality, low quality:1.4),(malformed hands:1.4),(poorly drawn hands:1.4),(mutated fingers:1.4),(extra limbs:1.35),(poorly drawn face:1.4),bad leg,strange leg, Instructions for Use: 1. The output resolution of XL should reach about 1024. It is recommended to use 928x1232 (3:4) for direct output and then enlarge. 2. Always include "masterpiece" and "best quality." 3. Use the 3M series sampler with around 35 steps. 4. XL models do not need to add VAE. Many beginners produce purple noise images because they add 1.5 VAE. 5. Regarding negative prompt words, you don't need to write too many. Personally, I use: (worst quality, low quality:1.4), (malformed hands:1.4), (poorly drawn hands:1.4), (mutated fingers:1.4), (extra limbs:1.35), (poorly drawn face:1.4), bad leg, strange leg.