这个模型是实验品，背景人物等效果很差，但是实际测试在同类型模型里面仍然是比较好用的。COS模型训练难度不是一般的大……
This model is specifically used to generate Cos graphs.Since it is trained from the animation base model, the body and hand effects of generating images are relatively poor. If you want a better effect, you can try MBW processing.I also tried to use a real human model to train, but the hair will always be black, can not achieve the desired effect, so give up, can only use anime base model training.
Train
CosXL is trained using 6000 high quality cos diagrams, using Lokr training with the base model AnimagineV3, the difference between A and B is that they use different LBW.
dataset
The training dataset used consists of over 6000 randomly selected images from Cosbooru, sourced from an ***** ******* website. Among the selected 6000+ images, 95% of them feature cosers from East Asia. The training images utilized are a collection of images after excluding NSFW content.
Training dependent parameter:
******* Model XL
GPU 1 x RTX4090
dataset 6450
Num Epoch 34
Learning Rate 1e-3
TE Lr 1e-5
Effective Bs 2 x 64
Optimizer AdamW
Lr_scheduler cosine_with_restarts
Lr_warmup_steps 50
Picture size:
(768 1312), count: 36
(768 1344), count: 111
(800 1280), count: 45
(832 1216), count: 159
(832 1248), count: 4404
(864 1184), count: 198
(896 1152), count: 21
(928 1120), count: 3
(1024 1024), count: 12
(1184 864), count: 24
(1216 832), count: 24
(1248 832), count: 1272
(1280 800), count: 39
(1344 768), count: 60
LoRa Block Weights:
CosXL_A: 0.5,0.8,0.3,0.3,0,0.3,0.3,0,0,0.5,0.5,0,0,0.5,0.3,0.3,0.5,0.6,0.6,0.5
CosXL_B: 0.8,0.3,0.3,0.0,0.3,0.3,0,0,0.5,0.5,0,0,0.5,0.3,0.3,0.5,0.6,0.6,0.5,0
Recommended settings
To guide the model towards generating high-aesthetic images, use negative prompts like (Please do not use the artist's name as a prompt):
nsfw, lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract]
Negative tags can include common negative tags, but it is best not to assign too high of a weight to their content, for example (ugly:2.8).
Most of the generation parameters of the example graph are:
euler_a | 20steps | no hires fix | CFG7
Acknowledgements
The development and release of CosXL would not have been possible without the invaluable contributions and support from the following individuals and organizations:
TusiArt:Our collaboration partner and sponsor.
HCP-diffusion:The previoustrainer
秋葉aaaki:This time use the training UI
Kohya SS: For their comprehensive training scripts.
Disclaimer:
All images generated by the model are created by the users themselves, and the model author cannot control the images generated by the users. The model author will not be held responsible for any potential copyright infringement or unsafe images.
License:
CosXL now uses the Fair AI Public License 1.0-SD, compatible with Stable Diffusion models. Key points:
Modification Sharing: If you modify CosXL, you must share both your changes and the original license.
Source Code Accessibility: If your modified version is network-accessible, provide a way (like a download link) for others to get the source code. This applies to derived models too.
Distribution Terms: Any distribution must be under this license or another with similar rules.
Compliance: Non-compliance must be fixed within 30 days to avoid license termination, emphasizing transparency and adherence toopen-sourcevalues
