Type | |
Stats | 6,001 |
Reviews | (1,274) |
Published | May 14, 2023 |
Base Model | |
Trigger Words | long sleeves,beard, hanfu |
Hash | AutoV2 EEF461398C |
注:不同模型权重不一样,有的0.8都没事,有的不能超0.5,需要自己调试,另外画女的普遍稳定些,原因各位懂得,男的还得看底模水平,比如我用的的大模型是anything,Counterfeit,linghuaInkJX3Game_v10还有专门画男生的一些模型,一般而言二次元或者本身就水彩水墨风的人物的大模型比较好;用Lora分层搭配其他lora更可控,效果也更好,我展示的就是和一个中国恐怖风lora结合的结果。
根据太平宫神仙壁画训练,后续马上会更新新的版本,该版本只是初步测试,但是发现效果还挺不错,因此也发上来。具体使用后续更新。暂时使用可以权重0.8-1.2。后续会逐渐完善该Lora,敬请期待
v1.0发布了,但是我发现还有一些问题,比如笏板official's tablet board可以触发,但是冠冕imperial crown触发几率很小,我分析是材料不足,下一版本会着重改善这类问题。还有我在使用中发现提示词太多效果不一定好,建议\(word\)这种词效果尤其脸部会明显改善。但我对这个不是很满意,建议直接v2.0。
v2.0更加稳定,泛化性也更加好,建议0.8以内。
According to the fairy mural training of Taiping Palace, a new version will be updated soon. This version is just a preliminary test, but it is found to be quite good, so it is also posted. Use subsequent updates. Temporary use can weight 0.8-1.2. The Lora will be gradually improved in the future, please look forward to it
v1.0 was released, but I found that there are still some issues. For example, the official's tablet board can trigger, but the imperial crown is very unlikely to trigger. My analysis is that there is not enough material. In addition, I found that the effect of too many prompt words is not necessarily good in use. The effect of suggestion \(word\), especially for the face, will be significantly improved. But I'm not very happy with this and suggest going straight to v2.0.
v2.0 is more stable and has better generalization, and is recommended within 0.8.