johnowhitaker
commited on
Commit
•
ef89c82
1
Parent(s):
16a3c99
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
VQGAN is great, but leaves artifacts that are especially visible around things like faces.
|
2 |
+
|
3 |
+
It's be great to be able to train a model to fix ('devqganify') these flaws.
|
4 |
+
|
5 |
+
For this purpose, I've made this dataset, which contains 100k examples, each with
|
6 |
+
- A 512px image
|
7 |
+
- A smaller 256px version of the same image
|
8 |
+
- A reconstructed version, which is made by encoding the 256px image with VQGAN (f16, 1024 version from https://heibox.uni-heidelberg.de/d/8088892a516d4e3baf92, one of the ones from taming-transformers) and then decoding the result.
|
9 |
+
|
10 |
+
The idea is to train a model to go from the 256px vqgan output back to something as close to the original image as possible, or even to try and output an up-scaled 512px version for extra points.
|
11 |
+
|
12 |
+
Let me know what you come up with :)
|
13 |
+
|
14 |
+
Usage:
|
15 |
+
```python
|
16 |
+
from datasets import load_dataset
|
17 |
+
dataset = load_dataset('johnowhitaker/vqgan1024_reconstruction')
|
18 |
+
dataset['train'][0]['image_256'] # Original image
|
19 |
+
dataset['train'][0]['reconstruction_256'] # Reconstructed version
|
20 |
+
````
|
21 |
+
|
22 |
+
|
23 |
+
|
24 |
+
Approximate code used to prepare this data: https://colab.research.google.com/drive/1AXzlRMvAIE6krkpFwFnFr2c5SnOsygf-?usp=sharing (let me know if you hit issues)
|
25 |
+
|
26 |
+
I'll be making a similar dataset with other VQGAN variants and posting progress on devqganify models soon, feel free to get in touch for more info (@johnowhitaker)
|