FKEY 画风模仿 STYLE DREAMBOOTH

【本模型训练过程未征得素材来源作者授权，如不恰当将立即下架。】【请勿将本模型及生成的图片用于任何商业用途！】【请勿利用本模型生成及传播不恰当的图片！！】训练方法：该Dreambooth采用了github上的kohya训练包。训练设备为RTX A6000。训练素材为70张FKEY老师的作品。并将其中脸部和上半身较为清晰的图片进行了二次裁剪，构成了总共134张图片的训练集。图片全部裁剪为10241024。其中，全身和半身图为裁剪成7681024或1024*768后，添加黑色背景填补剩余空缺部分。脸部图片（44张）repeat次数为25；上半身图片（56张）repeat次数为10；全身图片（34张）repeat次数为6。打标采用了Danbroou(0.7)+swinv2_tagger_v3(0.35)。自动打标后仅删去了与boy相关标签，其他标签未进行进一步处理。随后添加了触发词 fkey70。似乎是SDXL版本DB训练如果没有触发词，训练结果难以收敛。开启了shuffle caption。但将fkey70固定。全局学习率设置为1e-6。采用了constant的学习率调整策略。AdamW8bit优化器。没有开启噪声偏移。关闭了enable bucket选项。REG图片为从相关网站获取。共使用320张，略大于素材集的2倍数量。保存方法为按照step保存。每隔2500步保存一次。计划最大步数为150000步。50000-100000步时画风还原程度较高且繁华性较好。100000-120000步左右过拟合对画面产生影响。120000-150000步区间内出的图已经完全无法使用了。经过测试后，发布时所采用的70000步和75000步为原有特征表现和泛化性表现较为均衡的版本。clip skip在训练时为1。但跑图时采用了2。采用了混合精度训练。跑图：非常建议采用Adetailer。原始图片采用Euler A，step 28左右。当step过高，容易导致肢体错乱。CFG 5-7之间。而Adetailer时，部分参数与原始图片不同。DPM++ 3M SDE Karras，step 40左右。脸部高清化需要较高的迭代次数。CFG也需要开到8左右。因为原始素材中猫耳元素较多。如果生成的角色不用猫耳时，最好在负面提示词中添加animal ears和cat ears发布的两个版本间仅有细微的特征表现差别。请根据自己的偏好进行下载。"This model was trained without the consent of the authors of the source material. If it seems inappropriate, I will immediately delete it.""Please do not use this model or the images it generates for any commercial purposes!""Do not use this model to create or distribute inappropriate images!!"Training Methods:This Dreambooth implementation uses the Kohya training package available on GitHub.The training was conducted on an RTX A6000.The training dataset consists of 70 artworks by FKEY. The images where the face and upper body were clear were further cropped to create a total of 134 images.All images were cropped to 1024x1024. Full-body and half-body images were resized to 768x1024 or 1024x768 and then filled with a black background to adjust the remaining space.For the training, in each epoch, the face images (44 total) were repeated 25 times; upper body images (56 total) were repeated 10 times; and full-body images (34 total) were repeated 6 times.Labeling was done using Danbroou (0.7) + swinv2_tagger_v3 (0.35). After automatic tagging, only tags related to 'boy' were removed, with no further adjustments to other tags.The trigger word 'fkey70' was added. It seems that without a trigger word, the training results in the SDXL version of Dreambooth are hard to converge.Shuffle caption was enabled, but 'fkey70' was kept constant.The global learning rate was set to 1e-6 using a constant learning rate adjustment strategy. The optimizer used was AdamW8bit.Noise offset was not enabled, and the enable bucket option was turned off.REG images were obtained from related websites, totaling 320 images, which is slightly more than double the training dataset.The saving method was set to save by step, every 2500 steps.The plan was to run up to 150,000 steps. Between 50,000 and 100,000 steps, the style restoration and richness were quite good. From 100,000 to 120,000 steps, overfitting began impacting the visuals. From 120,000 to 150,000 steps, the images produced were completely unusable.After testing, the versions at 70,000 and 75,000 steps were chosen for release as they balanced original feature representation and generalization best.The clip skip was set to 1 during training but was increased to 2 during rendering.Mixed precision training was utilized.Generate Images:Highly recommend using Adetailer.For the full images, I prefer Euler A at about 28 steps. If the step count is too high, it can lead to disordered limbs. Set the CFG between 5 and 7.When using Adetailer, some parameters differ from those for the original images.I prefer DPM++ 3M SDE Karras, at about 40 steps.For high-definition facial enhancements, a higher number of iterations is required. The CFG should also be set to around 8.Because the original source material has more cat ears elements. If the generated character doesn't use cat ears, it's best to add animal ears and cat ears to the negative wordsThere are only minor differences in the performance of features between the two released versions. Please download according to your preference.