使用florence2-large-ft-gufeng_v3:<a target="_blank" rel="ugc" href="https://huggingface.co/F16/florence2-large-ft-gufeng_v3">https://huggingface.co/F16/florence2-large-ft-gufeng_v3</a><a target="_blank" rel="ugc" href="https://huggingface.co/F16/florence2-large-ft-gufeng_v3（这个对汉服类、团扇、油纸伞等chinese元素识别比较好）">（这个对汉服类、团扇、油纸伞等chinese元素识别比较好）</a>感谢newtypezx177大佬赞助的buzz这个flux lora模型是使用civitai在线训练，410图，花了6000buzz使用kohya训练，这个lora是第30个小时，效果不错。提前替你们踩下坑。推荐参数：<h2 id="-e03ocddty">所以这个模型会概率坏手脚，但大部分还是好的</h2>采样方法(Sampler)： Euler simple采样迭代步数(Steps)：20提示词相关性(CFG Scale): 3.5训练flux需要更加严谨的训练集tag，好好处理。 florence2-large-ft-gufeng_v3 model:<a target="_blank" rel="ugc" href="https://huggingface.co/F16/florence2-large-ft-gufeng_v3">https://huggingface.co/F16/florence2-large-ft-gufeng_v3</a>(This model excels in recognizing Chinese elements such as Hanfu, group fans, and oil paper umbrellas.)I am grateful to newtypezx177 for sponsoring the buzz.The flux lora model was trained online using civitai, with 410 images, and it cost 6000 buzz to train using kohya.This lora model was completed at the 30-hour mark, and the results are quite promising. I've gone ahead and tested the waters for you.Recommended parameters:Please note that this model may occasionally produce faulty hands and feet, but for the most part, they are good.Sampling Method: Euler simpleSampling Steps: 20CFG Scale: 3.5Training flux requires a more rigorous set of training tags, so it's important to handle them carefully.

古风女孩汉服FLUX

First and foremost, I extend my heartfelt gratitude to DukeG for their technical support and guidance.I am also deeply thankful to fok3827 for their fine-tuning of the florence2-large-ft-gufeng_v3 model:<a target="_blank" rel="ugc" href="https://huggingface.co/F16/florence2-large-ft-gufeng_v3">https://huggingface.co/F16/florence2-large-ft-gufeng_v3</a>(This model excels in recognizing Chinese elements such as Hanfu, group fans, and oil paper umbrellas.)I am grateful to both Chiren Shuo Meng and newtypezx177 for sponsoring the buzz.I also appreciate the efforts of Huan Yue Ling for assisting me with testing.The flux lora model was trained online using civitai, with 410 images, and it cost 6000 buzz to train using kohya.This lora model was completed at the 30-hour mark, and the results are quite promising. I've gone ahead and tested the waters for you.Recommended parameters:Please note that this model may occasionally produce faulty hands and feet, but for the most part, they are good.Sampling Method: Euler simpleSampling Steps: 20CFG Scale: 3.5Training flux requires a more rigorous set of training tags, so it's important to handle them carefully.

Files with SHA256: d88fca98df5168be4bf48b544a5f70dd7d19156399a26b5baa6d220bc12f3a04