Update README.md
Browse files
README.md
CHANGED
@@ -15,10 +15,10 @@ Based on my personal experience, preparing a LoRA's dataset properly is more imp
|
|
15 |
3. Upscale all low-resolution images to at least 2k resolution. But why? Let's see an example:
|
16 |
We have two identical png images, one with 907x823 pixels, the other with 3624x3288 pixels using simple 4x-upscale from SD WebUI.
|
17 |
<img src="imgs/upscale/comp.png"/>
|
18 |
-
Now let's crop them into roughly 1k resolution (960x832 in this case) using SD WebUI's Auto-size Crop:
|
19 |
<img src="imgs/upscale/cut-comp.png"/>
|
20 |
-
See the differences? The image with 4x-upscale has a sharper edges comparing to the one without any upscale, and hence looks more clean and less blurry when getting cropped into 1k resolution. <br />
|
21 |
-
So what does it translate to LoRA learning? <br />
|
22 |
-
Inside the LoRA training script, we can see a function called [bucket](https://github.com/bmaltais/kohya_ss/wiki/LoRA-training-parameters#enable-buckets). <br />
|
23 |
-
It'll automatically crop your image data into a size that's closed to the training resolution that we specify in training setup.
|
24 |
|
|
|
15 |
3. Upscale all low-resolution images to at least 2k resolution. But why? Let's see an example:
|
16 |
We have two identical png images, one with 907x823 pixels, the other with 3624x3288 pixels using simple 4x-upscale from SD WebUI.
|
17 |
<img src="imgs/upscale/comp.png"/>
|
18 |
+
Now let's crop them into roughly 1k resolution (960x832 in this case) using SD WebUI's Auto-size Crop:
|
19 |
<img src="imgs/upscale/cut-comp.png"/>
|
20 |
+
See the differences? The image with 4x-upscale has a sharper edges comparing to the one without any upscale, and hence looks more clean and less blurry when getting cropped into 1k resolution. <br />
|
21 |
+
So what does it translate to LoRA learning? <br />
|
22 |
+
Inside the LoRA training script, we can see a function called [bucket](https://github.com/bmaltais/kohya_ss/wiki/LoRA-training-parameters#enable-buckets). <br />
|
23 |
+
It'll automatically crop your image data into a size that's closed to the training resolution that we specify in training setup.
|
24 |
|