This model is specifically designed for fine-tuning cosplay portraits.
This is my first model release, trained on over 12,000 images. It took approximately 50,000 training steps.
Recommended intensity setting: 0.6~1.2. This model might not be trained very well, mainly because it was my first time training with such a large dataset and I didn't thoroughly clean the data. The model's performance differs significantly from the results on my actual dataset. I haven't properly cleaned the dataset or used more training methods yet, I apologize. However, you can definitely use it in conjunction with other models from the community; the results should be excellent. If you provide positive feedback, I will consider retraining a second version. Thank you!
The main reason I trained and released this model is because I've obtained many excellent models on Civitai, and I wanted to contribute to the community. Hope you have fun! Thank you!
Description
This is my first released model, trained using over 12,000 images. My main reason for training and releasing this model is that I've obtained many fantastic models on Civitai, and I wanted to contribute my small part to the community. I hope everyone has fun!
FAQ
Comments (7)
Thank you so much!! I also often download LoRAs from here. You’re someone with a great open-source spirit. Wish you a happy life, and may God bless you!!!
Thank you, and have a wonderful day!
模糊的原因是你训练集图片分辨率的关系，你愿意修改最小边长为统一的1024，就会是一个很好的lora了
Okay, I didn't notice the resolution issue. Could you please tell me if the minimum side length of each image must be greater than 1024 or equal to 1024?
@3193975562yu312 我的经验是等于1024，这样就能保留更多细节，我的训练集一般都是先用seedvr2放大或者edit类模型先洗一遍尺寸。有不清晰的地方或者有问题的地方都提前修改了。
@Ray3780 Okay, thank you very much. I will handle it as follows~
@3193975562yu312 Huge props for sharing this openly. You could try using an gemini or qwen3.7 type LLM to generate more detailed captions for your training data — with such a large dataset and that many training steps under your belt, this LoRA makes a fantastic catalyst for the checkpoint, bringing in tons of new concepts and fixing all those built-in tag mix-up issues. Can’t wait to see what you come up with on your next training run. 🫶