This article is a popular science and advice piece about AI image upscaling, published on June 27, 2023, with an estimated reading time of 2 minutes. The article's core argument is that in an AI workflow, upscaling an image may seem like an extra step, but it can have a significant impact on the final result.
Purpose: The article aims to explain why images should be upscaled, especially when using an image-to-image (img2img) workflow.
Main capabilities/principle: The article explains how a diffusion model works—the model first adds noise to an image until it is completely unrecognizable, then tries to restore the image from the noise. It learns noise patterns through a large number of reference images and many attempts. The key point is that when noise is added to an image, the clearer the image's details are, the more accurately the AI can restore it. Therefore, the resolution of the reference image plays an important role in the img2img workflow.
Usage/examples: The author gives an example: with a base resolution of 896×1344, using resolutions of 200×300 and 3584×5376 respectively through ControlNET (canny model) in Stable Diffusion to show the difference in details read by the AI. Under the same prompt and the same seed, the "Air Djinn" results generated from the low-resolution reference image and the upscaled reference image are different.
Tool options: The article lists several upscaling tools:
- ImgUpscaler: free, browser-based, 5 credits per day.
- TopazLabs: professional-grade workflow, many features, may be expensive for some people (affiliate link).
- Upscale Media: another free alternative, not many options but usable. In addition, upscaling can also be done directly in Stable Diffusion, but the higher the resolution, the greater the pressure on the GPU.
Applicable scenarios: Suitable for users who use Midjourney or Stable Diffusion for image-to-image creation, especially in scenarios where the quality of the reference image needs to be improved to improve the final generated result.
Limitations: The article points out that upscaling takes very little time; when upscaling directly in Stable Diffusion, high resolution places high demands on the GPU. TopazLabs may be expensive. Other features, prices, platforms, effects, etc. not mentioned are not provided for now.
Additional information: At the end, the article mentions that the author has updated the Midjourney style templates (free download, voluntary paid support), adding new styles such as Watercolor Manga, Magic University, Mythology (v5 only), and mentions techniques from their Brainstorming guide.
