使用了肌理感相对强的油画进行的训练，因此使用过程中不推荐对画面进行过多元素的添加或者描述，否则物体和肌理感叠加会使得画面看起来非常混乱，示例图片的树林其实有点那个意思- 非常适合对人物或者静物的特写，因为背景的部分虚化 + 人物的具象利于表达想要传达的氛围- 人像的高光很弱；- 人物、静物等的特写也可以生成，但未对人像训练集进行特殊的设计和考虑；如果只使用单个 lora，推荐权重 0.7- 1.0 注意：如果生成的场景有树林 ，建议降低权重，因为观察到上传的 image 里显示这个 lora 在树林的这个场景里发生了过拟合的情况 The training used oil paintings with relatively strong textures. Therefore, it's not recommended to add too many elements or descriptions to the image during the generation process. Otherwise, the combination of objects and textures can make the image appear very cluttered. The example image of the forest actually illustrates this point.<ul><li>It's very suitable for close-ups of people or still lifes, as the background blur + the concrete representation of the subject matter helps to convey the desired atmosphere.</li><li>The highlights in the portraits are relatively weak.</li><li>Close-ups of people, still lifes, etc., can also be generated, but the training dataset for portraits was not specifically designed or considered.</li><li>If using a single Lora, a weight of 0.7-1.0 is recommended.</li></ul>Note: If the generated scene contains a forest, it's recommended to lower the weight. Observations from the uploaded image show that this Lora overfits in forest scenes.

Files with SHA256: 6de4e6d451ad7690db7185cf84235bf9c15c80aaeb69793fb6647fab62bdd704