Update README.md
Browse files
README.md
CHANGED
|
@@ -11,7 +11,7 @@ I`m still in the state of learning, so any comments/feedbacks are welcom!
|
|
| 11 |
|
| 12 |
## Characters from Genshin Impact
|
| 13 |
|
| 14 |
-
###
|
| 15 |
#### Brief intro
|
| 16 |
LORA of Sangonomiya Kokomi, with her default costume in game.
|
| 17 |
civitAI page [download](https://civitai.com/models/9186/sangonomiya-kokomi)
|
|
@@ -37,3 +37,39 @@ Trainig process consist of two phases. The first one with default parameters of:
|
|
| 37 |
Then, for phase2, all three learning rate were decreased to 1/10, and trained with another 5 epochs.
|
| 38 |
|
| 39 |
#### results
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
|
| 12 |
## Characters from Genshin Impact
|
| 13 |
|
| 14 |
+
### Sangonomiya-Kokomi / 珊瑚宫心海
|
| 15 |
#### Brief intro
|
| 16 |
LORA of Sangonomiya Kokomi, with her default costume in game.
|
| 17 |
civitAI page [download](https://civitai.com/models/9186/sangonomiya-kokomi)
|
|
|
|
| 37 |
Then, for phase2, all three learning rate were decreased to 1/10, and trained with another 5 epochs.
|
| 38 |
|
| 39 |
#### results
|
| 40 |
+
V1.0 samples
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
## Characters from Honkai Impact 3rd
|
| 47 |
+
|
| 48 |
+
### Raiden Mei adult ver / 雷电芽衣
|
| 49 |
+
#### Brief intro
|
| 50 |
+
LORA of the adult Raiden Mei from Honkai Impact 3rd, Post-Honkai Odyssey, with her default costume in game.
|
| 51 |
+
civitAI page [download](https://civitai.com/models/13023/raiden-mei-adult-ver)
|
| 52 |
+
#### Training dataset
|
| 53 |
+
96 images of Raiden Mei:
|
| 54 |
+
* 36 illustrations. both SFW and NSFW, 3 of them are with other costumes.
|
| 55 |
+
* 30 360degree 3D model snapshots for accuracy.
|
| 56 |
+
* 30 360degree 3D model nude snapshot to ensure the costume is removable/replacable.
|
| 57 |
+
Since only one costume is included, all 96 images are placed inside one folder.
|
| 58 |
+
#### Captioning
|
| 59 |
+
WD14 captioning instead of the danbooru caption was used, since the former one will not crop/resize the images.
|
| 60 |
+
Threshold are usually set to 0.75-0.8. since I don't like to have a very long and sometimes inaccurate caption for my training data.
|
| 61 |
+
After captionin is done, I added "raiden mei" after "1girl" to every caption file generate as the triggering prompt. Some of the caption files were empty so I have to manually type the words.
|
| 62 |
+
#### Training setup
|
| 63 |
+
Trained with Kohya_SS stable diffusion trainer
|
| 64 |
+
Base model was [Anything V3.0 full](https://huggingface.co/Linaqruf/anything-v3.0/blob/main/anything-v3-fp32-pruned.safetensors)
|
| 65 |
+
Trainig process consist of two phases. The first one with default parameters of:
|
| 66 |
+
* learning_rate: 0.0001
|
| 67 |
+
* text_encoder_lr: 5e-5
|
| 68 |
+
* unet_lr: 0.0001
|
| 69 |
+
20 repeats, and 3 epoch
|
| 70 |
+
Then, for phase2, all three learning rate were decreased to 1/10, and trained with another 8 epochs.
|
| 71 |
+
|
| 72 |
+
#### results
|
| 73 |
+
V1.0 samples
|
| 74 |
+
|
| 75 |
+
|