Telegram: Improving the quality of AI images, and reducing AI artifacts
New to writing AI prompts? Check our blog for articles on how to create prompts and navigate the most common problems.
Navigating around artifacts
Getting the perfect image can sometimes be a challenge. One common problem are artifacts, a pixelated nuisance that distorts a spot of an otherwise normal image. It sometimes comes in the form of glitter, blue blocks, noise, or black cutouts. Learn to spot different kinds of artifacts and you can solve those problems right away.
#1 – Distortion and unnatural elements
Disfigured hands? Squinting left eye? These kinds of distortions are very common.These are the early days of generative imaging, and will likely go away as our systems become smarter with more training. For now, we offer some manual solutions.
In many cases, this can be remedied by a stronger prompt. Read up on Negative prompting to resolve this. Reinforcing words like symmetrical, the exact number of fingers, and other such techniques can work wonders.
#2 – Not enough reference data
First, consider if the model you are using is likely to have the person studied in its catalog. In the case of a brand new celebrity, for example, there isn’t enough reference information about that person, so the results will be poor unless we train a model with that person’s face first. Learn about Training
If that doesn’t do the trick, considering finding a concept/model that is more likely to have the subject matter. For example, a model that isn’t trained on anime is likely to produce distorted anime faces, whereas a model like Ghibli, or Waifu will likely get it right. Learn how to work with Concepts
However, if it’s a well photographed person, like a president or a national hero, the problem might just be lowering the guidance, which we cover below.
#3 – Blue Blocks / Crystallization / “Terminator” Look
A complete failure of diffusion results in a broken image. This usually pertains to the model not having enough reference material to fill in the blanks, such as using a very stylized model. The good news is, in many cases it can be solved quickly by lowering the Guidance
Remix commands that end in blue blocks may also be the problem. If the background has too much going on, and you’re trying to target an isolated subject, it may be difficult to parse out what to work with. In such a case, it’s better to begin with a well-lit, very isolated image first before remixing.