1gen published research results for how to train LORAs for characters with Extremly low real data (LORA-e).
Training LORA with very small datasets will cause the training to fail with many critical errors. It is a framework for algorithmically optimizing the usage of real data as well as synthetic data towards a higher quality successful LORA.
Our experiment showed promising results by using real data as few as 1%. it still generates meaningful photos with good quality, sometimes better. All results are shared at:https://civitai.com/user/okingjo, The Chief AI Artist of 1Gen Labs.
More code and paper to follow.
Dr. Ren@1gen labs
changelog:
- 2023/11 created
- 2024/3 updated