模型说明：目前实际测试下来稳定度还欠佳，但是可以出一些好图，此模型并非练lora融底模制作，故制作成本巨大，如果喜欢请给我点五星，感激不尽

训练分辨率为768大数据洗一遍，后面用1024的小数据训练。

所以没有什么严格的出图分辨率要求，图集也有各种分辨率，不超过1280为最佳，推荐768-1024分辨率出图

cfg请控制在4-6之间，XL的base拟合度非常高，训练后非常容易把clip练爆了，请适当降低一些CFG。

重绘幅度控制在0.3-0.4，过大的重绘可能会造成人物偏2.5d。原因是因为XL的本身数据量多大，我这点训练量实在无法冲刷掉本身的许多概念。

使用建议：建议以此模型为基础，训练真人类的lora，在小概念中能大幅提升出图质量

训练过程：此模型又五万张图训练了近百万步，后又用小图集分别进行多次训练。总共应该跑了150w步数。由于训练时间过久，硬件成本过于庞大个人难以负担，故提前结束。再混合了百分之15的realvisxlV20_v20Bakedvae模型，百分之10的SDXL_xxmix9realisticsdxl_v10。目前观察下来训练量还远远达不到预期，没有庞大的数据量冲刷XL是非常困难做到1.5的真实照片稳定度。但目前来看，效果足够把玩一阵子

后期计划：由于XL训练成本过大，为爱发电难以支撑。故此模型短期内只能有这一版，后续如果本人发财或者有金主爸爸资助深入研究训练，会继续更新。家境贫寒，各位理解（当然如果有金主爸爸愿意资助请随时联系我）

b站id：咒术师老白

Model description: Currently, the stability is not good in actual testing, but some good images can be produced. This model is not made by practicing Lora fusion bottom mold, so the production cost is huge. If you like it, please give me a five-star order. Thank you very much

Train with a resolution of 768 big data, followed by training with 1024 small data.

So there are no strict requirements for image resolution, and the atlas also has various resolutions, with a resolution of no more than 1280 being the best. It is recommended to produce images with a resolution of 768-1024

Please control the CFG between 4 and 6. The base fit of XL is very high, and it is easy to burst the clip after training. Please reduce some CFG appropriately.

The redrawing range should be controlled between 0.3 and 0.4, and excessive redrawing may cause the character to deceive 2.5d. The reason is that due to the large amount of data in XL itself, my training volume cannot wash away many of my own concepts.

Usage suggestion: It is recommended to use this model as a basis to train the Lora of the human class, which can greatly improve the quality of the image in small concepts

Training process: This model trained nearly a million steps with 50000 images, and then trained multiple times with small datasets. I should have run a total of 150w steps. Due to the long training time and the high hardware cost, it is difficult for individuals to bear, so it ended early. Mixed with 15% realvisxlV20_ V20Bakedvae model, 10% SDXL_ Xxmix9realisticsdxl_ V10. At present, it has been observed that the training volume is far from meeting expectations, and it is very difficult to achieve 1.5 real photo stability without a large amount of data to flush the XL. But currently, the effect is enough to play with for a while

Later stage plan: Due to the high training cost of XL, it is difficult to support power generation for Ai. Therefore, this model can only have this version in the short term. In the future, if I become wealthy or have funding from the sponsor's father for in-depth research and training, I will continue to update it. I come from a poor family, everyone understands (of course, if you have a father who is willing to support me, please feel free to contact me at any time)

Station B ID: Spell Master Lao Bai