This briefly describes how the upscaler works and its parameters. You can change them in the settings section
• Scale factor is the image scaling coefficient. This means if you requested a 500x1000 image, with a Scale Factor value set to 2, we will get a 1000x2000 image. We simply multiply the original image resolution by the Scale Factor value. Usually, everyone sets it to 2 as the maximum allowable value, but some users asked for the ability to reduce this value, so I added that option.
• Denoising strength - This parameter determines the image redrawing coefficient. Essentially, upscaling an image is the process of overlaying another, higher-quality image onto the original. The higher this parameter, the more the neural network will "make up" details. The lower the value, the more it will try to preserve the details of the original image, but the sharpness of the image will be lower. In some cases, with a high value, the image can even be ruined or develop artifacts. I would recommend a value between 0.2 and 0.3.
• CFG - works the same as with generation. Since upscaling also implies generating a new image, this parameter applies to this process as well.
• Positive, Negative - To understand the meaning of this setting, I'll try to explain how the upscaling algorithm works. First, the image is broken down into small segments, after which each segment is processed and upscaled separately. The script does not see the entire image, so there's no point in giving it your original prompt. But since processing segments is a generation process, it also has a prompt and a negative prompt. Due to the segmentation, this setting will only be useful for specifying quality tags. But there is one exception where these parameters are not applied, which I will write about next.
• Upscaler Script - the main engine of the image upscaling process. To understand which ones will be useful for you specifically, you need to test them all. I tested many different options and liked all versions of ESRGAN4x. They work well with Illustrious models, but perform poorly with realistic models or realistic images. I also noticed that sometimes the final image changes its color reproduction.
Therefore, another option was added - Resize and fill. This is the only script that does not divide the image into segments but processes it as a whole. In this mode, the prompt and negative prompt fields do not work; instead, it takes your prompt from the original file. It works well on all models and hardly changes the colors in the final image, but sometimes gives a less sharp result than ESRGAN4x.