About the looking at viewer tag
looking at viewer is a composition tag that NovelAI Diffusion can use. On Danbooru it has 3,315,722 posts, and its best-rated still image scores 2,722. The samples above put the tag at the front of a fixed base prompt with seed 1234, so the differences between tags come from the tag alone.
How each model draws looking at viewer
Follows composition and expression instructions closely, so the effect shows in a tidy form.
Reacts to the tag more strongly, so its effect is easier to see.
WAI illustrious SDXL v17, a local model. Shows how the same tag reads outside NovelAI.
Writing looking at viewer in a prompt
Put the subject first, then this tag, then quality tags.
1girl, solo, looking at viewer, best quality, masterpiecewlop, 1girl, looking at viewerThe negative prompt and the full settings are on the terms page.
How these samples were made
Every image uses 832×1216, seed 1234, 23 steps and the k_euler_ancestral sampler. The base prompt and negative prompt stay the same for every tag, and only the tag under test changes. Images flagged by the NSFW check are regenerated with a new seed or hidden.
From the Danbooru wiki
Show the wiki text ▾Hide ▴
A character looking directly at the viewer, i.e. looking at the camera or through the fourth wall.
Don't confuse with eye contact, which is for characters making eye contact with each other.
Don't confuse with facing viewer, which should be used when the eyes are closed, covered (e.g. with a blindfold), or looking at another direction.
Examples
Non-examples
- : facing viewer with closed eyes
- : facing viewer with blindfold
See also
FAQ
Does NovelAI recognize looking at viewer?
Compare the samples on this page. Every tag on this site shares the same seed and base prompt, so any change you see comes from looking at viewer.
How do I write looking at viewer in a prompt?
Write "looking at viewer" near the front of the prompt and separate it from other tags with commas.
Should I use V5 or V4.5?
V4.5 shows the tag's effect more strongly. V5 gives steadier images with fewer artifacts.