V2.0

• This model mainly tests training specific layers LoRA, referencing the post: Flux Fine-Tuning with Specific Layers

• 50 images, 512 resolution, trained for 1000 steps

• Training speed increased by approximately 1.7x

V1.0

1. This model is designed to generate different facial types of Asian women with slight NSFW functionality.

2. It has undergone two training sessions:

- First, with 400 real half-body photos.

- Second, using 150 images mixed with NSFW content.

- Both trainings were done at a resolution of 768x768.

3. The model was tagged simply with "Asian woman" and "Nude woman." TE were not trained in the process.

4. Occasionally, the model generates distorted limbs. After the second training, the facial variety became limited, with faces tending to merge into a similar appearance.

5. Future plans include testing Joycaption for tagging and experimenting with TE training for potential improvements.

6. The trigger word "Asian woman" is required to generate images correctly.

7. The model was trained on my potato GPU. Using SDP Attention, the training speed was about 1.7x faster than with xformers, but after the same number of steps, xformers produced better results.