更多款式可以来这里查看：新中式服装 vol.2
You can find more style here: New Chinese-style clothing vol.2
一些带中华古风元素的现代服饰。比如古风现代衣着，古风洛丽塔裙子等。
Some modern clothing with traditional Chinese elements. Such as Lolita dress with chinses traditional element.
诚挚邀请您尝试此LORA，作者很希望能得到您的反馈和建议。如果您能分享您生成的美图就更好了。感谢~
Sincerely invite you to try this Lora. I am eager for your feedback and comment. It would be great if you can share the beautiful images generated by this lora. Thanks~
该LORA的训练期望仅仅包含裙子，不包括人脸和背景等其他元素。您可以搭配其他lora或底模元素来生成完整图片。如果污染到其他元素请评论告知，我会尽量尝试修复。
The training expectation of this LORA only contains the dress, not including other elements such as faces and backgrounds. You can mix this LORA with elements (for example : face) of other lora or checkpoint to create a complete picture. If this LORA pollutes other elements, pls let me know by comment. I will try to fix it.
** 请使用ADetailer/Face editor来修缮脸部
** Please use ADetailer/Face editor to fix the face
推荐LORA权重：您可以先尝试0.7。如果感觉有污染到其他元素请酌情降低权重。不建议超过0.85。
Recommend LORA weight : You can try 0.7 first. If you feel that other elements are polluted, please reduce the weight as appropriate. It is not recommended to exceed 0.85.
触发关键词：请参考版本说明
Trigger word: Please refer to instructions for each version
Description
FAQ
Comments (12)
Wow, Actually looks great, i tried training Lora, I trained clothes 5 times, each time i have issue with face being trained, any idea how can i fix that, and also can you share parameters this model is trained on, Thanks
Maybe you can try regularization for training. Compared with training parameters, I think the organization of training materials is more important. My experience is that the impact of training material on model performance is far greater than training parameters. My training parameters are as below
--enable_bucket --pretrained_model_name_or_path=chilloutmix_Ni.safetensors --resolution=512,768 --network_module=networks.lora --max_train_epochs=20 --learning_rate=8e-5 --unet_lr=8e-5 --text_encoder_lr=1e-5 --lr_scheduler=cosine_with_restarts --lr_warmup_steps=0 --lr_scheduler_num_cycles=1 --network_dim=64 --network_alpha=64 --train_batch_size=4 --save_every_n_epochs=2 --mixed_precision=fp16 --save_precision=fp16 --seed=1337 --cache_latents --prior_loss_weight=1 --max_token_length=225 --caption_extension=.txt --save_model_as=safetensors --min_bucket_reso=256 --max_bucket_reso=1024 --keep_tokens=1 --xformers --shuffle_caption --clip_skip 2 --optimizer_type AdamW --noise_offset 0.1
@cyberAngel_ Thank u so much, btw you mean regularization images ?, how that works
@Astro9 There is a parameter called reg_data_dir of kohya-ss sd-scripts. Maybe you can search what it does and usage. It's something like class identifier of DreamBooth.
@cyberAngel_ Thank u so much, i'll definitely look it up, i already have many reg images for human faces, should i use that for clothes also or i have to get some clothes reg images
@Astro9 Actually I am not very sure how reg training works as well 😂 I just try it and found that reg seems can enhance the multiusability of the model (For example, change a face to wear it). But someone point out that my LORA is still hard to change a face. I'm also wondering whether reg works or not now😂
My current approach is to put the clothes without head in the main training set. And then put images with head wearing this clothes in the reg folder.
If I only put clothes without head in the main training set without reg, the head will be too large or too small when I use the LORA to generate images. I read a articles saying that if you put someone wearing this clothes to reg, reg will tell the model how the clothes should be worn, and where should the clothes be placed. It can improve the wrong proportion of clothes and head. However, I also found that the head in the reg will pollute the LORA. In my understanding , the element in the reg but not in the main training set should not be learned by the model. But it does. The head in the reg also be learned to the LORA, and hard to change it.
Maybe you can try and tell me your founding. I look forward to your experience as well~ Thx a lot
@cyberAngel_ goddamn that's a nice trick 👌, this is exactly what I was looking for, i will try that, even if it can change slightly for better is enough i can do rest in img2img, thank u so much 🙏
就冲你这个赛博天使我就的关注你一波😁
嘿嘿~感谢~ 名字出自崩坏3一首歌很好听可以去听听哦hhhh~
可问题是,触发关键词参考各版本说明,那么各版本的说明又在哪里呢？
版面右上角下载按钮的下面有个“Trigger Words” 哦~
好的,感谢解答
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.