Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,75 @@
|
|
| 1 |
---
|
| 2 |
license: other
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: other
|
| 3 |
+
datasets:
|
| 4 |
+
- p1atdev/pvc
|
| 5 |
+
language:
|
| 6 |
+
- en
|
| 7 |
+
library_name: diffusers
|
| 8 |
+
tags:
|
| 9 |
+
- text-to-image
|
| 10 |
+
- stable-diffusion
|
| 11 |
+
- safetensors
|
| 12 |
---
|
| 13 |
+
|
| 14 |
+
# PVC v3
|
| 15 |
+
|
| 16 |
+
This model is a latent diffusion model finetuned on Waifu Diffusion v1.5 beta 2 with PVC figure images.
|
| 17 |
+
You can use Danbooru tags to generate images.
|
| 18 |
+
|
| 19 |
+
## Downloads
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
## Prompt guide
|
| 24 |
+
|
| 25 |
+
### Trigger words
|
| 26 |
+
|
| 27 |
+
- `pvc` means the pvc material style but not needed always.
|
| 28 |
+
|
| 29 |
+
- `figma` is the figure style that has joints, and more tend to be product thumbnail images.
|
| 30 |
+
|
| 31 |
+
- `nendoroid` means the style of chibi figures. Use with `chibi` to get better results.
|
| 32 |
+
|
| 33 |
+
### Tips
|
| 34 |
+
|
| 35 |
+
The PVC figure style is closer to the anime style than to the realistic style.
|
| 36 |
+
So, it is recommended to put `anime` to **positive** prompt and `realistic` to **negative** prompt to get better results sometimes.
|
| 37 |
+
If you want to avoid too realistic faces, try this! Setting `realistic` to negative does not affect the pvc style, only the faces of the characters.
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
## Examples
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
## Training information
|
| 46 |
+
|
| 47 |
+
<details>
|
| 48 |
+
<table>
|
| 49 |
+
<thead>
|
| 50 |
+
<tr><th>Parameter</td><td>Value</th></tr>
|
| 51 |
+
</thead>
|
| 52 |
+
<tbody>
|
| 53 |
+
|
| 54 |
+
<tr><td>Service</td><td>Runpod</td></tr>
|
| 55 |
+
<tr><td>GPU</td><td>A5000</td></tr>
|
| 56 |
+
<tr><td>Notebook</td><td><a href="https://github.com/Linaqruf/kohya-trainer/blob/main/kohya-trainer.ipynb" target="_blank">Linaqruf/kohya-trainer</a></td></tr>
|
| 57 |
+
<tr><td>Cost</td><td>about $2</td></tr>
|
| 58 |
+
<tr><td>Hours</td><td>about 6 hours</td></tr>
|
| 59 |
+
|
| 60 |
+
<tr><td>Dataset</td><td>7467 images from p1atdev/pvc</td></tr>
|
| 61 |
+
<tr><td>Resolution</td><td>896</td></tr>
|
| 62 |
+
<tr><td>Epochs</td><td>5</td></tr>
|
| 63 |
+
<tr><td>Optimizer</td><td>Lion</td></tr>
|
| 64 |
+
<tr><td>LR</td><td>4e-7</td></tr>
|
| 65 |
+
<tr><td>Scheduler</td><td>cosine_with_restarts</td></tr>
|
| 66 |
+
<tr><td>Train Batch Size</td><td>1</td></tr>
|
| 67 |
+
|
| 68 |
+
</tbody>
|
| 69 |
+
</table>
|
| 70 |
+
</details>
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
## License
|
| 74 |
+
PVC v3 is released under the Fair AI Public License 1.0-SD (https://freedevproject.org/faipl-1.0-sd/). If any derivative of this model is made, please share your changes accordingly. Special thanks to ronsor/undeleted (https://undeleted.ronsor.com/) for help with the license.
|
| 75 |
+
|