我尝试用以前的图集对SDXL进行了训练，尝试了很多次，这是目前我认为最好的一个模型。（但也许是SDXL本身很强大罢了，与我的训练无关）总之，推荐权重0.8-1，触发词是homo但是我觉得它有和没有真的差不多（SDXL它学不进去，气死了），因此你可以选择不采用触发词。顺带一提，我的电脑显存只有8G，所以例图是使用了TusiArt的网站在线产出的，也许它与本地部署的SDXL产出图片会有差别。希望你喜欢 ；）展示的图片提示词都采用了自然语言，负面提示词采用了我以往的例图的同一套负面提示词，封面的提示词是这样的（它没有把爱注入“胸肌”，哈哈哈）：A handsome man with short hair squatting on the top of the TV tower at night, colored hair, polygonal pupils, cyberpunk world, looking down on the boy and the TV tower from above, the boy throws winks at the camera, the boy is naked and has "love" tattooed in the middle of his chest muscles,I have tried training SDXL using a previous dataset many times, and this is currently the best model I believe I have achieved. (But perhaps SDXL itself is just very powerful, unrelated to my training efforts.) In any case, I recommend a weight of 0.8-1 and the trigger word "homo", but I feel that whether it is present or not doesn't make much of a difference (SDXL just doesn't seem to learn, so frustrating!). Therefore, you can choose not to use the trigger word. By the way, my computer only has 8GB of VRAM, so the example images were produced using the TusiArt website online, which may differ from images generated by locally deployed SDXL. Hope you like them ;)The displayed image captions are all in natural language, and negative image captions follow the same set of negative prompts as my previous examples. The cover image caption is as follows (without injecting love into the "chest muscles", haha): "A handsome man with short hair squatting on top of the TV tower at night, with colored hair and polygonal pupils, in a cyberpunk world. He looks down from above at the boy and the TV tower, while the boy winks at the camera. The boy is naked and has a tattoo of "love" in the middle of his chest."

Files with SHA256: 74f869a1f448450904c12f2fe5c6874f87bafae124a81ef67fb0fe3333dbcfaf