Type | |
Stats | 345 397 |
Reviews | (35) |
Published | Aug 20, 2024 |
Base Model | |
Training | Steps: 3,150 Epochs: 5 |
Usage Tips | Clip Skip: 1 |
Trigger Words | east face |
Hash | AutoV2 C38213B311 |
Flux is very good, but the Asian women's faces it generates are relatively simple and fixed, so I trained this lora. It is good at generating realistic pictures, but it can also be compatible with the hints of fantasy department. Generally speaking, when the seed value is the same, the weight of 025-0.4 will not change the content and composition of the original Flux image too much. I tested that the weight of 0.3 is the most balanced, and in many cases, it will almost only change the face in the image. Between the weight of 0.45 - 0.8, the picture will still follow the prompt words, but the details will be slightly deviated, but the aesthetic embodiment is still excellent. I think this is still due to the powerful ability of Flux model. With 12 billion parameter training, lora is more used for guidance.
————————
Flux很好,但它生成的亚洲女人面孔相对单一和固定,所以我训练了这个lora。
它擅长生成偏真实的图片,但也可以兼容幻想系的提示词。
通常来说,在seed值相同时,025-0.4的权重不会过多改变Flux原版生图的内容和构图,我测试下来0.3权重是最平衡的,许多情况下,几乎只会改变图像内的脸部。
0.45-0.8的权重之间,图片还是会遵循提示词,但细节会稍有偏离,但美学体现上依然出色。我想这还是得益于Flux模型的强大能力,120亿的参数训练量,lora更多只是用于引导。