Przemysław Mejna
commited on
Commit
·
c4da60a
1
Parent(s):
efa4c08
Update README.md
Browse files
README.md
CHANGED
|
@@ -9,6 +9,25 @@ This is the reason why I made this embedding.
|
|
| 9 |
I have trained it on 70 images, the version I will be targeting in upcoming weeks will be based on 128 or 256 well-selected and filtered images, and processed
|
| 10 |
through pixelate tool to keep the same pixel size on each of the input data. This should improve the embedding dramatically.
|
| 11 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
### Installation:
|
| 13 |
|
| 14 |
Just drop the embedding files (.pt extension) to your SD embeddings folder (`your-folder/embeddings`).
|
|
@@ -56,7 +75,11 @@ The embeddings give you a great opportunity to change some of your works into pi
|
|
| 56 |
The best way to do it is to follow this process:
|
| 57 |
|
| 58 |
First get your subject. If this is a simple image as:
|
| 59 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 60 |
|
| 61 |
What I did here was to use:
|
| 62 |
**Positive prompt:** "game icon, raven, by pixelart, pixelated" (very important to add pixelated)
|
|
@@ -68,4 +91,30 @@ What I did here was to use:
|
|
| 68 |
**Resolution:** 768x768.
|
| 69 |
|
| 70 |
The result:
|
| 71 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
I have trained it on 70 images, the version I will be targeting in upcoming weeks will be based on 128 or 256 well-selected and filtered images, and processed
|
| 10 |
through pixelate tool to keep the same pixel size on each of the input data. This should improve the embedding dramatically.
|
| 11 |
|
| 12 |
+
**Images:**
|
| 13 |
+
<div style="display: flex; flex-direction: row; flex-wrap: wrap">
|
| 14 |
+
<img src="https://i.imgur.com/lVqWhmx.png">
|
| 15 |
+
</div>
|
| 16 |
+
<div style="display: flex; flex-direction: row; flex-wrap: wrap">
|
| 17 |
+
<img src="https://i.imgur.com/9eOUMiD.png">
|
| 18 |
+
</div>
|
| 19 |
+
<div style="display: flex; flex-direction: row; flex-wrap: wrap">
|
| 20 |
+
<img src="https://i.imgur.com/SJGs2nD.png">
|
| 21 |
+
</div>
|
| 22 |
+
<div style="display: flex; flex-direction: row; flex-wrap: wrap">
|
| 23 |
+
<img src="https://i.imgur.com/eMZvUTY.png">
|
| 24 |
+
</div>
|
| 25 |
+
<div style="display: flex; flex-direction: row; flex-wrap: wrap">
|
| 26 |
+
<img src="https://i.imgur.com/fyGB10h.png">
|
| 27 |
+
</div>
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
|
| 31 |
### Installation:
|
| 32 |
|
| 33 |
Just drop the embedding files (.pt extension) to your SD embeddings folder (`your-folder/embeddings`).
|
|
|
|
| 75 |
The best way to do it is to follow this process:
|
| 76 |
|
| 77 |
First get your subject. If this is a simple image as:
|
| 78 |
+
<div style="display: flex; flex-direction: row; flex-wrap: wrap">
|
| 79 |
+
<img src="https://i.imgur.com/ks4bpSV.png">
|
| 80 |
+
</div>
|
| 81 |
+
|
| 82 |
+
Here it needed just one step!
|
| 83 |
|
| 84 |
What I did here was to use:
|
| 85 |
**Positive prompt:** "game icon, raven, by pixelart, pixelated" (very important to add pixelated)
|
|
|
|
| 91 |
**Resolution:** 768x768.
|
| 92 |
|
| 93 |
The result:
|
| 94 |
+
<div style="display: flex; flex-direction: row; flex-wrap: wrap">
|
| 95 |
+
<img src="https://i.imgur.com/wErverC.png">
|
| 96 |
+
</div>
|
| 97 |
+
|
| 98 |
+
Of course, the result can be better - you can re-roll to infinity or choose better settings or different embedding than recommended pixelart (in some cases you can try others)
|
| 99 |
+
|
| 100 |
+
Pixelating photos/more complex images:
|
| 101 |
+
This is more tricky - but doable. As a baseline use the above settings,
|
| 102 |
+
you can experiment with higher/lower CFG or denoising. To keep likeness I don't recommend you to go over 0.6 denoising.
|
| 103 |
+
Replace the first part of the prompt with a simple description, at the end should be part: "by/in style pixelart, pixelated"
|
| 104 |
+
Probably this will take up to 2-3 rounds. When I like the output - I set it as a base for next iteration. Then I reduce denoising by 0.4 each
|
| 105 |
+
extra round.
|
| 106 |
+
Of course this is rough process - might be different based on images.
|
| 107 |
+
|
| 108 |
+
Here are some examples:
|
| 109 |
+
|
| 110 |
+
Input:
|
| 111 |
+
<div style="display: flex; flex-direction: row; flex-wrap: wrap">
|
| 112 |
+
<img src="https://i.imgur.com/l2dGhgH.png">
|
| 113 |
+
</div>
|
| 114 |
+
|
| 115 |
+
Resuts:
|
| 116 |
+
<div style="display: flex; flex-direction: row; flex-wrap: wrap">
|
| 117 |
+
<img src="https://i.imgur.com/zsrd92w.png">
|
| 118 |
+
</div>
|
| 119 |
+
|
| 120 |
+
I still investigating how to improve on the process.
|