Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,30 @@
|
|
| 1 |
---
|
| 2 |
license: creativeml-openrail-m
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: creativeml-openrail-m
|
| 3 |
+
tags:
|
| 4 |
+
- StableDiffusion
|
| 5 |
+
- Anime
|
| 6 |
---
|
| 7 |
+
|
| 8 |
+
This model is for the two characters Anisphia and Euphyllia from Tensei Ōjo to Tensai Reijō no Mahō Kakumei (転生王女と天才令嬢の魔法革命, aka The Magical Revolution of the Reincarnated Princess and the Genius Young Lady in English).
|
| 9 |
+
|
| 10 |
+
To prompt the two characters use **AnneSophia** and **Euphilia**.
|
| 11 |
+
|
| 12 |
+
**Training details**
|
| 13 |
+
|
| 14 |
+
- Dataset: around 100 concept images (30 for Anisphia only, 30 for Euphyllia only, and 30 for the two) + around 100 regularization images
|
| 15 |
+
- Trainer: [Kohya trainer colab](https://github.com/Linaqruf/kohya-trainer)
|
| 16 |
+
- Text encoder is trained, total training step is 26000 but probably the checkpoint at 16000 steps is already good enough
|
| 17 |
+
- This time I try training with clip skip 2.
|
| 18 |
+
However strangely it turns out that using clip skip 1 for inference seems to give better results as shown below (for the final checkpoint)
|
| 19 |
+
|
| 20 |
+
- Training at clip skip 2 may help separate the two characters, but there is not enough evidence now
|
| 21 |
+
|
| 22 |
+
**Example generations**
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
|