>>>The Role of Negative Prompts
In Section 3-1, I explained how to write a prompt, but AUTOMATIC1111 has an additional feature called negative prompt. While a prompt is composed based on the 'required elements' that you want to create, a negative prompt is composed based on the 'elements that should not be created'. For example, if you want to output a boy but keep getting a girl, add Prompt girl to the negative prompt to remove the girl element. The writing method is the same as the prompt.
>>>Learn how to use negative prompts
The words entered in the negative prompt are usually determined by referring to the generated image results. If features that are not relevant to the learning process are output with the prompt connected, it can effectively prevent the appearance of those features.
In addition, just as there was a quality prompt among the prompts, you can specify a prompt to display the quality in the negative prompt. For example, three phrases are often used: Prompt worst quality, low quality, normal quality.
Prompt masterpiece, high quality, 1girl, witch, blue robe, hat, long black hair, sitting, smile, looking at viewer, full body, flower garden, blue sky, castle,fantasy, vivid color, noon, sunny, Negative Prompt worst quality, low quality, normal quality |
It's a good idea to start with a small number of negative prompts, check the generated images, and modify them as needed.
>>>Use embedding
Including the negative prompts explained so far, there is a function in the WebUI that allows you to load a completed learning file that contains all the concepts you want to exclude in the image generation process by entering a keyword in the negative prompt. This is called embedding.
Embedding is a way to express words in numerical form so that machines can understand them in the natural language processing process. It refers to the process of inserting vectors that exist as many as words or words into low-dimensional vectors so that the correlation between words or sentences can be analyzed. In Stable Diffusion, embedding refers to a learning model file that uses a method called Textual Inversion, and it is stated that it is a narrow concept.
Embedding files, like models, are distributed on sites such as Hugging Face and Civitai, and can be downloaded and used. Depending on the purpose, there are embedding files that are suitable for each purpose, and for example, there are files specialized in accurately generating finger structures.
negativeXL
https://civitai.com/api/download/models/97691?type=Negative&format=Other
easynegative(SD 1.5 series)
gsdf/EasyNegative · Datasets at Hugging Face
If you are using Colab,
upload the downloaded file to
sd > stable-diffusion-webui > embeddings in Google Drive.
If you are in a local environment,
upload the downloaded file to
root{stable diffusion directory}\stable-diffusion-webui\embeddings
After restarting the WebUI, enter the emeddings you want to use for the negative prompt. Or, place the cursor in the blank space of the negative prompt, and in the Textual inversion tab, click to select the embeddings you want to use, and they will be added automatically.
>no embeddings
