请支持游戏及画师，本模型仅供研究使用！
配合ADetailer插件或许有更好的面部效果。
推荐权重：0.4-0.8，具体请自己调整。
以下是一些可能有用的正向提示词，可参照我已上传的图片：
jianxin,1girl,solo,gloves,jewelry,earrings,black hair,multicolored hair,long hair,white gloves,black gloves,long sleeves,clothing cutout,bangs,ahoge,yin yang,强调建模风格：
3djianxin,Description
For Pony
FAQ
Comments (5)
This is a pretty good model! 20mb too! Great job, how long did this take to train?? When looking through the metadata it looked like it was 21 epochs per folder with a combined total of 60 repeats from all the different folders used for each style (I'm guessing that's what you did anyway) I tried it on my machine (3090) those specs just to see and it said 20 hours for 10 repeats, 20 epoch... Thta's crazy if true. Unless you have a beefy gpu or there was optional args that somehow lowered the overall time.
Translate using ChatGPT.
Hello! I am just a beginner and still exploring many areas. Thank you for your recognition.
Regarding model size, I learned from @Chenkin. In short, a size of 20MB is sufficient to accommodate several two-dimensional characters without much change in image quality.
I don't have many optimization strategies, just relying on high configuration, long time, and passion. I use a cloud server platform with a configuration of 24GB/4090, and it took approximately 5 hours for training.
The evaluation criteria for model effectiveness vary due to different needs. Some pursue "better outfit changes" or "more art styles." Currently, the demand for training my model is to "faithfully recreate characters" (mainly based on A3.1). Much of the processing revolves around this requirement, so other criteria are not my main concern. To be honest, as you mentioned, the biggest workload is indeed collecting the training set. When searching for the training set, I aim for "high quality, multiple angles," primarily based on the "official design of the characters." Also, to ensure "faithful recreation of characters," I processed the labels. (I am still exploring more effective ways in this regard.) During training, I personally adhere to the belief that "overfitting is better than underfitting," so I try to train as many rounds as possible. A refined training set and reasonable training are crucial factors in improving model quality.
By the way, I believe that for model publishers, a comprehensive and thorough usage document is also an important factor in improving model quality. For many reasons, the training set is usually not made public. However, to ensure image quality, I usually write out notable prompts from the training set and provide selected display images covering character features from different perspectives. This may make the prompts look not very concise, but I believe it is beneficial for both character recreation and more detailed feature adjustments. Although outfit changes and art style shifts are not my main criteria, I also conduct some testing when releasing the model and provide explanations for any special cases.
The above is just my personal understanding. Thank you for your comments.
@Elusa Thank you so much for in-depth response! For a beginning you're on the right track for sure. As I said, 20mb and it being very good at the character is fantastic. Too many on here have the mindset of 217mb being the norm when that absolutely should not be the case.
Though I feel that become the norm when the first, and now largest people were on this site at the beginning of sdxl and all of them made 217mb and people just followed without really learning about even WHY it has has to be larger. You did a great job with finding good quality images too. For me, all too often I have to upscale the ones I find to look more clean and crisp, otherwise I get pixelation around the edges sometimes.
I've heard upscaling though can also get rid of small artificts within the image though, so now I always upscale regardless. Giving thorough usage documents and explaining your workflow, I'm sure you will become more popular and well received in no time! Best of luck!
i really like your model, it's coming out very good.
are there any chance you will try to model the other characters from wuthering waves ?
Thank you! I am still in the process of learning. Currently, I don't have any plans to model the other characters from wuthering waves.