Verified: 2 years ago
SafeTensor
Type | |
Stats | 481 |
Reviews | (102) |
Published | Apr 29, 2023 |
Base Model | |
Training | Steps: 6 Epochs: 40 |
Trigger Words | emilia (re:zero) |
Hash | AutoV2 DDCFBAFE92 |
1、This model is based on novel,but it de well on some models.
2、The lora is train with high resolution(1536x1536), so it can directly generate 1024 * 1536 images without producing any excess head or body( I suggest changing cfg to between 3 and 5 without higfix)
3、I suggest using both highfix and controlnet new tile model (cfg: 7-11, Denoising strength: 0.3-0.6,sd-webui-lora-block-weight:MIDD),.It performed very well with .
1、模型基于novel。但是也应该能在其余模型上面表现的不错。
2、训练lora的使用了1536*1536的分辨率,所以它可以在很少出现多余头部(头上加头)的情况下直出大图,但是这种情况下需要吧cfg调低(3-5)。
3(我比较建议的)还是老实开高清修复和controlnet最新的tile(不是原来的tile啊),这种情况下使用分层控制的MIDD,同时加上cfg 7-11,重绘幅度:0.3-0.6(你觉得不像就多加点)。在这种情况下表现的很不错。