This is a download site for the negative embedding boring_e621. Its original huggingface repository, which goes into more detail about how it was trained and what it does, can be found here.
Briefly, This embedding attempts to capture what it means for an image to be uninteresting. So if you're using the Automatic1111 Stable Diffusion WebUI, you can place this file in stable-diffusion-webui\embeddings and add the trigger word to your NEGATIVE prompt, and your results should become more interesting, both stylistically and topically.
The title image for this model illustrates the quality improvements it can make to four simple prompts in the base SD 1.5 model. In the first column, for example, boring_e621_v4 makes the hamburger more photorealistic, places it on a cocobolo table, freshens the vegetables, and adds a slice of cheese and an order of fries.
As this embedding is known for its ability to improve furry art specifically, the second image illustrates the improvements it yields on four simple prompts on YE-18, a model appearing in many furry mix models.
Which version should I get?
boring_e621_v4 (newest and works best with most non-furry models)
boring_e621_fluffyrock_v4 (May work best on models with Fluffyrock in their ancestry, which includes most popular furry models)
boring_e621 (works well with most models, but is less intense than v4)
Each version has two downloads, a pt file and a safetensors file. They should behave identically, but some programs might only be able to load one of the formats. So if the first one you download doesn't work with your program, try the other.
Each of the Suggested Resource models below has specifically been reported to produce great results with one of these embeddings.
Description
Trained and works on Fluffyrock.
No need to prepend "by" to this version's name to invoke it. Just type "boring_e621_fluffyrock_v4".
FAQ
Comments (21)
The fluffyrock version seems to be broken, images generated with same seed/settings are identical with or without it. Regular 4.0 version working fine.
I re-downloaded boring_e621_fluffyrock_v4 from civitai and used it to try to reproduce your problem, but boring_e621_fluffyrock_v4 still changed my output. Can you give me some generation details (including the prompt used and model version) about the failure? Additionally, if you are using the Automatic1111 interface, can you confirm that the embedding is showing up in the generation details window post-generation, as I've circled in red in this image: https://huggingface.co/FoodDesert/boring_e621/blob/main/boringeval.png ?
@FoodDesert thanks for going to so much effort! I'll re-download the file too and do more testing as you asked
@FoodDesert It does seem to be working now, I should have made a plot before or checked if it was the generation details when I had the issue. I overwrote the file before testing so maybe a bad download? Anyway working now thank you!
@degenerated Great! I am glad that it is working now.
Love this one <3
I'm using the non furry version. It does make things more interesting. How does this embedding do this?
Thanks for your comment. All the versions of this embedding are trained on images that no one liked from an art website. So by putting the embedding in the negative prompt, you are telling your model to NOT make your image look like images that people don't like. It is described in more detail at the huggingface repository here https://huggingface.co/FoodDesert/boring_e621
@FoodDesert I read it and its very clever. I have a doubt. Did you make sure to exclude new posts so that they can have 0 upvotes cos they are very new?
@monkeyman007 I did exclude from the dataset any image that was not at least a couple months old for exactly that reason. I did not choose the cutoff date in any really formal way, though. I made the original version in February, and I think my arbitrary cutoff date for image inclusion then was December 31 2022. I am not sure, though. I did not record all the choices I made at that level of detail.
@FoodDesert That seems to work fine!
just so im getting this right the negative prompt need to be (Boring e621), for whichever version im using ? or it needs to be different
No. For v1.0, "by boring_e621" should be in your negative prompt. For v4.0, it should be "boring_e621_v4" . For fluffyrock v4.0, it should be "boring_e621_fluffyrock_v4". I could not set the trigger words for v1.0 on this site correctly because civitai tries to change the filename if I do that.
@FoodDesert
I could not set the trigger words for v1.0 on this site correctly because civitai tries to change the filename if I do that.
This is because the trigger is the filename. Unless you're using some kind of custom syntax or programmatic workflow (idk how diffusers does it), interfaces use the filename to trigger the embedding (a1111 just uses the filename alone, ComfyUI uses something like embedding:filename to prevent accidentally usage of embeddings). The filename in the prompt is replaced with the contents of the file (I have no sources for this, but I'm pretty sure this is why embeddings are also called textual inversions). It's easier to think of an embedding as shorthand for a lengthy prompt addition; unlike loras, TIs don't modify the model.
I have seen embeddings that recommend using something extra with it for a better effect that may be unwanted, and embeddings that either go along with a negative embedding or recommend adding something to negative prompt (because it's not possible to have the same embedding apply to both positive and negative prompts). I've also seen I think a single embedding that recommends adding an extra trigger because it was mistrained but worked properly with the extra tokens (you can adjust this by adding the extra tokens into the embedding, but that's not generally a simple thing to do)
@civitai997 boring_e621 (version 1) is an example of one of the embeddings you mention that should work better if invoked with extra words that are not part of the embedding's name.
When I made version 1 of boring_e621, I prepended "by " to the embedding name during training to imitate the syntax the model it was trained on uses for artists. The idea is that the embedding should resemble an artist name to the model so that it more readily affects style. So the correct way to use version 1 is by typing "by boring_e621" (in automatic1111), even though the word "by" is not part of the embedding. I wanted to include the extra word "by " in the "trigger word" so that when people press the copy button and put it in their prompt, it would invoke the embedding in the recommended way with "by " prepended, but was thwarted by the renaming.
Version 4 does not have this issue because I noticed when searching online for prompts using boring_e621 that lots of people leave the "by" off, making it less effective. For that reason, I trained later versions so that the first vector in the embedding resembles the word "by" to the model, thus getting the same effect without relying on the user to put extra words in their prompt. So I do not have "trigger word" problems with it.
It would be easy to hack the vector for "by" to the beginning of version 1 of the embedding, but that would result in inconsistent versions of it floating around on the internet.
Out of curiosity, why is fluffyrock v4.0 not available on the civitai generator? Sorry if this has already been asked before, I can't find any info
Oh wow, I did not even realize that. My best guess is that I marked "Other" as its base model, so civitai has no way of knowing automatically that it will work with SD 1.5 models. I'll see if I can switch it to SD 1.5.
Thank you for pointing this out. I do not think on-site generation was a thing when I set it up, and I do not pay a lot of attention to Civitai. It seems to be resolved now.
pleasing name for it
thank you for uploading safetensors! this looks awesome, im excited to try it out.
Why is it called, ''Boring''? To me, ''boring can be an insult. Any images I make using this resource, it's the complete polar opposite of boring and my pictures are NOT boring
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.