FoodDesert commited on
Commit
f82c848
1 Parent(s): 4a59959

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +5 -20
README.md CHANGED
@@ -22,11 +22,11 @@ depend on manually curated lists of tags describing features people do not want
22
  * Manually compiled lists will inevitably be incomplete.
23
  * Models might not always understand the tags well due to a dearth of training images labeled with these tags.
24
  * It can only capture named concepts. If there exist unnamed yet visually unappealing concepts that just make an image look wrong,
25
- * but for reasons that cannot be succinctly explained, they will not be captured by a list of tags.
26
  <br>
27
 
28
  To address these problems, boring_e621 employs textual inversion on a set of images automatically extracted from the art site
29
- e621.net, a rich resource of millions of hand-labeled artworks, each of which is both hand-labeled topically and rated
30
  according to its quality. E621.net allows users to express their approval of an artwork by either up-voting it, or marking it as a favorite.
31
  Boring_e621 was specifically trained artworks automatically selected from the site according to the criteria
32
  that no user has ever Favorited or Up-Voted them. boring_e621 thus learned to produce low-quality images, so when it is
@@ -41,26 +41,11 @@ used in the negative prompt of a stable diffusion image generator, the model avo
41
 
42
  # Evaluation
43
 
 
44
 
45
- I extracted the tags from three e621 images and used them to construct a set of test prompts.
46
- * one prompt was constructed from an image with a high number of favorites.
47
- * one prompt was constructed from an image with a moderate number of favorites.
48
- * one prompt was constructed from an image with 0 favorites.
49
- <br>
50
-
51
- I then generated test images from each of these prompts, each time using a different negative embedding as the negative prompt. Particularly, I tried:
52
- * [EasyNegative](https://huggingface.co/datasets/gsdf/EasyNegative)
53
- * [Bad Artist](https://huggingface.co/nick-x-hacker/bad-artist)
54
- * [Bad Prompt](https://huggingface.co/datasets/Nerfgun3/bad_prompt)
55
- * [boring_e621](this)
56
- <br>
57
-
58
- Finally, I qualitatively evaluated the attractiveness and interestingness of the resulting images, though I will let you draw your own conclusions from the output below.
59
- <br>
60
-
61
- ## Results
62
 
63
- ![Negative Embedding Comparison](https://i.imgur.com/d7R4gGi.jpg)
64
 
65
 
66
  ## Other Models
 
22
  * Manually compiled lists will inevitably be incomplete.
23
  * Models might not always understand the tags well due to a dearth of training images labeled with these tags.
24
  * It can only capture named concepts. If there exist unnamed yet visually unappealing concepts that just make an image look wrong,
25
+ but for reasons that cannot be succinctly explained, they will not be captured by a list of tags.
26
  <br>
27
 
28
  To address these problems, boring_e621 employs textual inversion on a set of images automatically extracted from the art site
29
+ e621.net, a rich resource of millions of hand-labeled artworks, each of which is both human-labeled topically and rated
30
  according to its quality. E621.net allows users to express their approval of an artwork by either up-voting it, or marking it as a favorite.
31
  Boring_e621 was specifically trained artworks automatically selected from the site according to the criteria
32
  that no user has ever Favorited or Up-Voted them. boring_e621 thus learned to produce low-quality images, so when it is
 
41
 
42
  # Evaluation
43
 
44
+ To qualitatively evaluate how well boring_e621 has learned to improve image quality, we apply it to 4 simple sample prompts using the base Stable Diffusion 1.5 model.
45
 
46
+ [boring_e621 and boring_e621_v4 Performance on Simple Prompts](tmpoqs1d_vv.png)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
47
 
48
+ As we can see, putting these embeddings in the negative prompt yields a more delicious burger, a more vibrant and detailed landscape, a prettier pharoah, and a more 3-d-looking aquarium.
49
 
50
 
51
  ## Other Models