Monday, April 24, 2023

Negative Embeddings - ULTRA QUALITY Trick for A1111 [VIDEO SUMMARY]

The video "Negative Embeddings - ULTRA QUALITY Trick for A1111" is published by Olivio Sarikas and demonstrates two tricks to enhance the quality of AI-generated images. The first trick involves using negative embeddings, while the second trick focuses on upscaling and refining images. The creator shares practical examples and provides step-by-step guidance for using these techniques.


Negative embeddings are trained on undesirable image features, and they can be applied to the art style, prompt, or other aspects of an image to improve its quality. To use a negative embedding, simply download it into the appropriate folder and use its name (minus the .PT extension) as a negative prompt. The weight applied to the negative embedding can be adjusted as needed, with the creator recommending a value of 0.8 for many cases. The examples presented in the video show that using different negative embeddings can significantly change the appearance of an image while still improving its overall quality.

The second trick shared by the creator involves upscaling and refining images. To do this, first render an image at a lower resolution, then double its resolution using an image-to-image tool. The creator suggests using the same seed and setting the denoise strength to 0.25 to maintain similarity while allowing the AI to introduce new details. After the initial rendering is complete, copy the image and open it in a photo editing software like Affinity Photo or Photoshop.

In the photo editing software, apply an unsharpen mask with a radius of one pixel and a strength of 1 pixel at 0% to enhance the image's texture and sharpness. The creator demonstrates this by showing the difference in quality between the original and sharpened images. Save the sharpened image as a PNG file to avoid JPEG artifacts, then load it back into the AI tool. It is important to disable face restoration, as it can blur the face, but keep the other settings the same.

A comparison between the original upscaled image and the new, AI-rendered sharpened version reveals that the latter has clearer and more detailed features. This is particularly noticeable in the fabric and background details. However, be cautious of oversharpening, which can create bright edges and other artifacts. If this occurs, either reduce the sharpening or use the eraser tool in the photo editing software to remove the problematic areas.

In conclusion, the video "Negative Embeddings - ULTRA QUALITY Trick for A1111" by Olivio Sarikas presents two valuable techniques for enhancing the quality of AI-generated images. By using negative embeddings and a combination of upscaling and refining, users can achieve improved image quality and detail.