Skip to content
Documentation

Getting good results

Last updated

The two things you control that move output quality the most: the photo you upload and the instruction you write.

Most quality problems are input problems

When a render disappoints, the instinct is to rewrite the prompt. It is usually the photograph. The model reads the room's geometry, light direction, and surfaces out of the image you gave it, so a dark, tightly cropped, or heavily angled shot removes information no wording can put back. A wide, well-lit photo taken with your back to the longest wall, with both the floor and the ceiling line in frame, prevents more problems than any phrasing will fix. That is why photographing a room comes before writing a good prompt in this section, and not the other way round.

When to stop tuning and regenerate

Generation is not deterministic, and that trips people up more than any other property of the tool. The same photo and the same instruction produce a different image on every run, which means a mediocre result is often just a mediocre roll rather than evidence your prompt is wrong. The habit that follows is to generate two or three and pick, rather than rewriting after every attempt. Why two runs give different results covers this properly, including how to tell ordinary variation apart from a result that is consistently wrong. When a render fails outright, or comes back with a warped floor or a smeared edge, that is a different problem and has its own page.