This version takes into account the sky proportion problem and depth of field. And the training set is three times larger, with each image magnified using the tile model. The results are impressive. Lowering the weights and not using trigger words should result in just scenery (which didn't work in an hour of testing), cyberpunk, surreal, etc.
该版本考虑到了天空占比问题以及景深。并且训练集扩大了三倍，每张图片都使用tile模型放大过。效果很可观。降低权重、不使用触发词应该可以只得到风景（但在一个小时的测试中没有成功过），赛博朋克，超现实等应该都可以生成。
This version takes into account the sky proportion problem and depth of field. And the training set is three times larger, with each image magnified using the tile model. The results are impressive. Lowering the weights and not using trigger words should result in just scenery (which didn't work in an hour of testing), cyberpunk, surreal, etc.