This time, I set the trigger word to natural language：photos of a real female cyborg with realistic skin texture and hair.In the example, I only tried 8 steps and cfg=4 because that would be faster for testing. I didn't use turbo lora, as it makes the realistic image style look oilier.This setting might cause damage to your hands or structure. Perhaps you could try other parameters later, such as increasing the number of steps.All images used in the training were generated by the sdxl model; no real images were used.The WebP format image I uploaded should contain a workflow.这次训练我尝试将触发词设置成了自然语言：photos of a real female cyborg with realistic skin texture and hair.在示例中，我仅用8步和cfg=4进行了尝试，因为这用来测试会快一些。我未使用turbo lora，turbo lora会使得写实图片风格变得更油。这个设置可能会导致手部或结构崩坏，也许你之后可以试一下其他参数，比如增加步数。训练使用的素材图片均为sdxl模型生成，未使用真实的素材。上传的webp格式的图片中应该是带有工作流的。基于preview3训练，在base1.0运行测试。

Files with SHA256: f14ef02e9a0ad197ded3a81216236c8844aa36a3d64873b21758ba7c837ecde0