sfy | a character lora trained by only one picture版本v1.0 (ID: 131806)

2023.8.3 update:

• I use kohya_ss's svd_merge_lora.py to change the lora's dimension from 128 to 32 and make it become much smaller. In some cases, the character can't keep same with the original picture, but most time it still works well.

---

Original:

• This is a character lora that trained by only one picture on SDXL1.0(a full body picture generated by a merged checkpoint which is SD1.5 version, then cut it into an upper body picture and a head picture, I think it can be called "trained by only one picture").

• It's a test to find out if we can get a character lora who don't have many pictures or even just generated by AI and you want to keep him/her.

• You can change the character's action and clothes by prompts. In SD1.5, it's much easier because there're many high quality checkpoints in the same style, and you can also use Clothing lora and controlnet to change it (I have already tried it and it work well). But maybe SDXL will do it better in the future.

• Now this lora is not pretty well, the clothes can't be replaced completely, but maybe I'll push other version which can solve this problem in the future. (You can follow the lora's metadata to train another one in SD1.5, it must be better.)

• I recommend using DPM++ SDE Karras with weight 1.0.

描述:

训练词语: black footwear, black cape, black pantyhose, short dress, black dress