Dreambooth training Stable Diffusion on a human subject with 900+ training images

dreambooth-training.md

Dreambooth training Stable Diffusion on a human subject with 900+ training images

Details

Training

• Model: Stable Diffusion v1.4
• Training repo: JoePenna/Dreambooth-Stable-Diffusion
• Learning rate: 1E-6

Sampling

• Prompt: portrait of a spanish poet MY_TOKEN person, colourised, face portrait, epic, tragic, military art, fantasy, dieselpunk, hd shot, digital portrait, beautiful, artstation, comic style, by artgerm, guy denning, jakub rozalski, magali villeneuve and charlie bowater
• Seed: 535185327
• Sampler: Euler a

Unmodified network

500 steps

1200 steps

1500 steps

2000 steps

The 9-up really was what was generated for the third image in the batch, it's not an editing error. Strange.

2500 steps

3000 steps

3500 steps

4000 steps

4500 steps

5000 steps

5500 steps

6000 steps

6500 steps

7000 steps

7500 steps

8000 steps

8500 steps

9000 steps

9500 steps

10000 steps

10500 steps

11000 steps

11500 steps

12000 steps

12500 steps

13000 steps

13500 steps

14000 steps

14500 steps

15000 steps

Beyond

Unmodified network

500 steps

3000 steps

5000 steps

7000 steps

9000 steps

11000 steps