Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,26 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- danbooru
|
| 5 |
+
---
|
| 6 |
+
|
| 7 |
+
# Image Restoration
|
| 8 |
+
|
| 9 |
+
This marks the 10th month anniversary of the previous image restoration model.
|
| 10 |
+
|
| 11 |
+
The decoder model redraws the masked area based on the booru tags.
|
| 12 |
+
|
| 13 |
+
The input and the training target are based on:
|
| 14 |
+
- LAB color space
|
| 15 |
+
- FFT
|
| 16 |
+
- FDL
|
| 17 |
+
|
| 18 |
+
As in another [encoder](https://huggingface.co/nightknocker/sdxs-1b-image-to-longclip-encoder) model, the image was split to 4 equal-sized squares, and then one of the squares was masked by less than 30%.
|
| 19 |
+
|
| 20 |
+
The image goes through the PatchEmbed module. The transformer blocks receive the masked, patched image embeddings along with their Fourier transform, and the [Camie tags](https://huggingface.co/Camais03/camie-tagger-v2) as input.
|
| 21 |
+
|
| 22 |
+
Finally, the latent is [upsampled](https://github.com/zelaki/eqvae/blob/18f2016dd7e9a934c2fb175d6ff54f52422d8928/train_eqvae/ldm/modules/diffusionmodules/model.py) to the original image size.
|
| 23 |
+
|
| 24 |
+
## Datasets
|
| 25 |
+
|
| 26 |
+
- pixiv rank
|