Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,27 @@
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: apache-2.0
|
3 |
---
|
4 |
+
|
5 |
+
# Anime AI Art Detect
|
6 |
+
A BEiT classifier to see if anime art was made by an AI or a human.
|
7 |
+
|
8 |
+
### Disclaimer
|
9 |
+
Like most AI models, this classifier is not 100% accurate. Please do not take the results of this model as fact.
|
10 |
+
|
11 |
+
The best version had a 96% accuracy distinguishing aibooru and the images from the imageboard sites. However, the success you have with this model will vary based off of the dataset.
|
12 |
+
|
13 |
+
Here are some biases I have noticed from my testing:
|
14 |
+
|
15 |
+
- Images on aibooru, the site where the AI images were taken from, were high quality AI generations. Low quality AI generations have a higher chance of being misclassified
|
16 |
+
- Textual inversions and hypernetworks increase the chance of misclassification
|
17 |
+
|
18 |
+
### Training
|
19 |
+
This model was trained from microsoft/beit-base-patch16-224 for one epoch on 11 thousand images from imageboard sites, and 11 thousand images from aibooru.
|
20 |
+
|
21 |
+
You can view the wandb run [here](https://wandb.ai/saltacc/huggingface/runs/2mp30x7j?workspace=user-saltacc).
|
22 |
+
|
23 |
+
|
24 |
+
### Use Case
|
25 |
+
I don't intend for this model to be more accurate than humans for detecting AI art.
|
26 |
+
I think the best use cases for this model would be for cases where misclassification isn't a big deal, such as
|
27 |
+
removing AI art from a training dataset.
|