Create README.md

README.md

@@ -0,0 +1,56 @@
+
+# boring_e621
+
+This embedding attempts to capture what it means for an image to be uninteresting.  It was trained as a negative embedding for [ye18](https://huggingface.co/Doubleyobro/yiffy-e18) using e621 style tags as prompts during training
+though it may also work in other SD 1.4-based models. 
+If you're using the [Automatic1111 Stable Diffusion WebUI](https://github.com/AUTOMATIC1111/stable-diffusion-webui), place the .pt file in 
+stable-diffusion-webui\embeddings and add "by boring_e621" to your negative prompt for more interesting outputs.
+<br>
+
+## Model Description
+
+
+The motivation for this is that negative embeddings like [Bad Prompt](https://huggingface.co/datasets/Nerfgun3/bad_prompt), 
+whose training is described [here](https://www.reddit.com/r/StableDiffusion/comments/yy2i5a/i_created_a_negative_embedding_textual_inversion/) 
+depend on manually curated lists of tags describing features people do not want their images to have, such as "deformed hands".  Some problems with this approach are:
+* Manually compiled lists will inevitably be incomplete. 
+* Models might not always understand the tags well due to a dearth of training images labeled with these tags.
+* It can only capture named concepts.  If there exist unnamed yet visually unappealing concepts that just make an image look wrong, but for reasons that cannot be succinctly explained, they will not be captured by a list of tags.
+<br>
+
+To address these problems, boring_e621 employs textual inversion directly on a set of 100 of the most boring images on the furry message board e621.net.  
+These images were automatically collected by querying e621 for images that no user has ever Favorited or Up Voted. 
+Almost all images on the site have at least one up vote or favorite, so the images that no one favorited are some of the very most uninspiring images on the website. 
+Thus, you should be able to use the resulting embedding as a negative prompt to make your generated images Not Boring.
+<br>
+
+# Bias, Risks, and Limitations
+* Using this as a negative embedding often sacrifices some fidelity to the prompt.  For example, characters in the image may disappear or change eye/skin/fur color.
+* Using this as a negative embedding may introduce sexual content into images that would not otherwise have it.  I suspect this is because people do not find sexual content boring, and thus it may be absent in the embedding's training data.  
+* Unlike other negative embeddings, this is not intended to fix problems like extra limbs or deformed hands.  It can be used alongside other negative embeddings to fix deformities.
+<br>
+
+# Evaluation
+
+
+I extracted the tags from three e621 images and used them to construct prompts in the style used by [ye18](https://huggingface.co/Doubleyobro/yiffy-e18).
+* one prompt was constructed from an image with a high number of favorites.
+* one prompt was constructed from an image with a moderate number of favorites.
+* one prompt was constructed from an image with 0 favorites.
+<br>
+
+I then generated test images from each of these prompts, each time using a different negative embedding as the negative prompt.  Particularly, I tried:
+* [EasyNegative](https://huggingface.co/datasets/gsdf/EasyNegative)
+* [Bad Artist](https://huggingface.co/nick-x-hacker/bad-artist)
+* [Bad Prompt](https://huggingface.co/datasets/Nerfgun3/bad_prompt)
+* [e621_boring](this)
+<br>
+  
+Finally, I qualitatively evaluated the attractiveness and interestingness of the resulting images, though I will let you draw your own conclusions from the output.
+<br>
+
+## Results
+
+![Negative Embedding Comparison](https://i.imgur.com/d7R4gGi.jpg)
+
+