Stable Diffusion is a powerful AI model that generates images from text prompts. However, by default, the images generated may not always meet the desired quality or expectations. This is where negative embeddings come into play. Negative embeddings are pre-trained models that help filter out undesirable results from the image generation process, resulting in high-quality and visually appealing images.
In this article, we will explore the top Stable Diffusion negative embeddings available, each with its own unique features and benefits. By using these negative embeddings, you can enhance your image generation process and achieve the desired results more efficiently.
Related:
Stable Diffusion Negative Embedding
versa-simple-neg
This negative embedding model is highly versatile, and it improves the chances of generating normal-looking hands in realistic images. For best results, it is recommended to keep the token usage below 75 and avoid adding other quality specifiers to prompts, such as worst quality or bad quality.
Using versa-simple-neg is simple, and it requires adding the file name of the embedding to the negative prompt section of Stable Diffusion. If any artifacts are observed in the generated images, adjusting the weighting can help to mitigate the issue. Weighting adjustments are made by adding a colon and the desired value (e.g., file_name_of_embedding:1.1 or file_name_of_embedding:0.8).
Versa-simple-neg has been tested on various AI models, including VersaMix, ChilloutMix, and Grapefruit. It is primarily used for anime, but it can also produce semi-realistic images with an artistic touch.
7 Dirty Words negative prompt
One of the stable diffusion negative embeddings available is the “7 Dirty Words” negative prompt. This negative embedding is trained on the output generated from entering George Carlin’s “7 dirty words” in the positive prompt. However, it is important to note that this negative embedding is extremely NSFW, and not in an appealing way, so it should only be used for artistic purposes.
The 7 Dirty Words negative prompt is particularly useful for generating soft and cheerful images with vibrant colors. It was trained mostly on output from various anime models, so it’s likely to work better for anime images. When used in the negative prompt, this negative embedding steers heavily away from producing gross and undesirable images, making it a good choice for those looking to avoid unpleasant results.
It’s important to exercise caution when using this negative embedding and ensure that it is only used in the negative prompt. However, for those who wish to experiment with realistic versions of the negative embedding, it’s possible to do so and see what results are produced.
bad-picture negative embedding
To use the Bad-Picture Negative Embedding with Stable Diffusion, you need to download the file and put it in the appropriate folder. Depending on which version you downloaded, the name of the file will be either bad-picture-chill-75v, bad-picture-chill-32v, or bad-picture-chill-1v. If you want to change the name, you can do so, and it will change the activation word to the new name.
This negative embedding is specifically made for ChilloutMix (version chilloutmix_NiPrunedFp32Fix.safetensors [fc2511737a]). It comes in three versions: 1 vector, 32 vector, and 75 vector versions. The higher the version, the more detail you will be able to extract, but it may also lead to deep-frying of the image. It’s essential to check out the comparison images uploaded as part of the 75 vector version to understand this.
If you want to experiment with the different versions of the negative embedding, you can download the “Training Images” version, which has all the saved .pt versions that were generated during training. Additionally, the “Actual Training Images” version can be downloaded if you want the actual training images.
It’s essential to note that sometimes the original prompt or negative prompt may include a LORA or embedding, and the links to them are provided in the article. However, these are not used in the single image previews, only the original prompts that are borrowed and put into the stacked comparison images.
Danger Zoo
One of the negative embedding options for Stable Diffusion is Danger Zoo. Despite its name, Danger Zoo is an artistic and cinematic negative embedding that can produce amazing images. It is one of the good sets of image enhancers that are worth checking out.
Danger Zoo is compatible with Stable Diffusion versions SD1.5 and SD2.1 and is specifically designed for use in negative prompts. By adding Danger Zoo’s keyword to the negative prompt section of Stable Diffusion, you can eliminate undesirable elements from the generated images and produce better results.
To learn more about Danger Zoo and other stable diffusion negative embedding options, visit the original model pages for full details and examples. Don’t settle for undesirable image results – explore the power of negative embedding with Stable Diffusion.
Deep Negative V1.x
Deep Negative V1.x is a negative embedding that has been trained specifically to identify and avoid disgusting compositions and color patterns in generated images.
By using Deep Negative V1.x in the negative prompt section of Stable Diffusion, you can greatly reduce the chances of drawing images with faulty human anatomy, offensive color schemes, upside-down spatial structures, and other undesirable elements. The number of vectors per token varies depending on the specific training set used, with options ranging from 2T to 75T.
If you’re looking for the easiest-to-use embedding with the most accurate dataset, 75T is your best bet. However, if you’re working with a well-trained model, 64T may be a better choice, although it may require some tuning. Alternatively, you can opt for 16T or 4T to reduce the chances of drawing bad anatomy, but keep in mind that they may still draw ugly faces or have little effect on light and shadow.
It’s important to note that because Deep Negative V1.x is designed to learn how to create disgusting concepts, it cannot improve picture quality accurately. Instead, it is best used with negative prompts related to worst quality, low quality, logo, text, watermarks, or usernames. It can also be used in conjunction with other similar negative embeddings for even better results.
If you’re interested in seeing the results of Deep Negative V1.x in action, there are several examples and tests available, including drawing buildings, fixing hands, creating portraits with PureErosFace, and fusion body fixes.
Unspeakable Horrors
One such negative embedding is called “Unspeakable Horrors.” It was designed to prevent the generation of disturbing and unsettling images. However, it should be noted that these embeddings may not be effective if your model already excels in composition and anatomy. In such cases, using negative embeddings with fewer vectors might be more suitable.
It’s also worth mentioning that prompt weighting with embeddings can be somewhat flawed, especially on high vector counts. Therefore, it’s important to fine-tune your model carefully and avoid low multipliers for high vector embeddings, especially when using Automatic1111.
It’s essential to remember that these embeddings were created solely for negative prompts and should never be used as positive embeddings. Doing so may lead to the generation of highly disturbing and graphic images, which could cause significant harm. As the creator of these embeddings, I cannot be held responsible for any loss of sanity that may occur due to the misuse of these embeddings.
EasyNegative
EasyNegative is a negative embedding designed specifically for anime images. While you can try using it for realistic images, it’s primarily intended for anime. It’s effortless to use – just place it in the “\stable-diffusion-webui\embeddings” folder and use its corresponding keyword in the negative prompt section of Stable Diffusion. You can adjust the strength of the embedding as desired, although it seems to scale well without any distortions.
If you like EasyNegative, consider giving the repository a like and checking out the other works of HuggingFace. The samples of this embedding are named sample01 – Counterfeit-V2.0.safetensors, sample02 – AbyssOrangeMix2_sfw.safetensors, and sample03 – anything-v4.0-pruned.safetensors. You can compare the strength of these samples using AbyssOrangeMix2_sfw.
According to the author, EasyNegative is trained with Counterfeit, and although it can be used with other models, its effectiveness may not be certain. So, if you’re looking to generate anime images with Stable Diffusion, EasyNegative is an excellent negative embedding to help you avoid undesirable results.
Flonix’s Prompt Embeds
With version 1.5 of Flonix’s Prompt Embeds, you no longer need to use multiple tags to generate high-quality images. These prompt embeds automatically add commonly used tags, resulting in stunning and realistic images. By using Flonix’s Prompt Embeds, you can take your image creation to the next level and unlock its full potential.
Using Flonix’s Prompt Embeds is easy. Simply download the .bin files and place them in the /embeddings folder of your A1111 sd install. Then, use the name of the file in your prompt to activate the prompt embed. For example, if the name of the file is flonixneg.bin, use “flonixneg” in your prompt.
Flonix’s Prompt Embeds have been created for both artistic and realistic images, resulting in some of the cleanest images you’ll see. By using these prompt embeds, you’ll no longer have to worry about undesirable or low-quality images. Download Flonix’s Prompt Embeds now and experience the difference for yourself!
