Gerold Meisinger commited on
Commit
f6fba87
1 Parent(s): 7a62860

eval, readme

Browse files
Files changed (2) hide show
  1. README.md +35 -0
  2. eval.zip +3 -0
README.md CHANGED
@@ -1,3 +1,38 @@
1
  ---
2
  license: cc-by-nc-sa-4.0
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: cc-by-nc-sa-4.0
3
  ---
4
+
5
+ **Restore missing RGB channels**
6
+
7
+ Restore a missing channel of a RGB image by using ControlNet to guide image generation of Stable Diffusion to infer missing channel from the other two channels.
8
+ * See accompanying discussion at [github.com - Channels RGB](https://github.com/lllyasviel/ControlNet/discussions/567) with detailed report and evaluations.
9
+ * To restore images with missing channels you can use [this space](https://huggingface.co/spaces/GeroldMeisinger/channels).
10
+ * For evaluation images see the corresponding .zip's at "files".
11
+ * To run your own evaluations you can use [this script at gitlab.com](https://gitlab.com/-/snippets/3602096).
12
+
13
+ # Training
14
+
15
+ ```
16
+ accelerate launch train_controlnet.py \
17
+ --pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5" \
18
+ --train_batch_size=4 \
19
+ --gradient_accumulation_steps=8 \
20
+ --proportion_empty_prompts=0.5
21
+ --mixed_precision="fp16" \
22
+ --learning_rate=1e-5 \
23
+ --enable_xformers_memory_efficient_attention \
24
+ --use_8bit_adam \
25
+ --set_grads_to_none \
26
+ --seed=0 \
27
+ --num_train_epochs=2
28
+ ```
29
+
30
+ # Image dataset
31
+
32
+ * laion2B-en aesthetics>=6.5 dataset
33
+ * --min_image_size 512 --max_aspect_ratio 2 --resize_mode="center_crop" --image_size 512
34
+ * Cleaned with `fastdup` default settings
35
+ * Data augmented with right-left flipped images
36
+ * Resulting in 214244 images
37
+ * Set whole channel to 0 by alternating between R-G-B channels
38
+
eval.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d334a8a368fefc4d34105f621d6e440ca717f30bb51eaecd1b8a3ab9711e2890
3
+ size 25911220