How we check every AI-generated room render before a user sees it
Image models are very good at restyling a room and quietly bad at leaving the room alone. Ask for a Japandi living room and you get beautiful furniture, and sometimes also: a window that moved, shrank, or appeared over a kitchen sink where the photo shows a solid wall; a closet door that came back as a curtain, a sliding screen, or plain wall; a sofa parked across the only archway; a new doorway…
Artificial intelligence-generated room renders undergo a rigorous three-step verification process before they are displayed to users. This process involves three separate gates that each perform a specific type of check on the rendered image.
The first gate is a vision-model judge that compares the original photo with the AI-generated render and answers a set of fixed yes/no questions. This gate is designed to catch obvious errors such as doors morphing into curtains or windows appearing in places where they shouldn't be. The judge uses a structured JSON output to answer the questions, ensuring that it cannot provide any explanatory prose.
There are 16 such checks for indoor rooms and 9 for outdoor scenes, covering everything from door and window preservation to wall uniformity and the absence of faces in artwork.
The second gate measures the camera's viewpoint to ensure that the render was captured from the same perspective as the original photo. This is done through a combination of visual inspection and mathematical analysis, looking for shifts in the ceiling line and depth profiles that might indicate a change in viewpoint. This gate was previously prone to false positives, but recent improvements have greatly reduced the error rate.
The third gate performs a crop-level presence check, examining individual elements such as doors, windows, and stairs in full resolution. This is important because a door or window can be very thin in the image and might not be captured by the earlier gates. If any of these elements are missing, moved, or otherwise altered, the render will fail this gate.
If a render fails any of these gates, it is automatically retried, and the process continues until all checks pass. If no acceptable render can be produced, the system will display the best attempt as a draft, clearly marked as such. This multi-layered approach helps ensure that the renders displayed to users are accurate, realistic, and free of the kinds of errors that can mislead or confuse potential customers.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
