An AI depth reader guesses what is near and far in a picture, the kind of depth map ControlNet and 3D tools are steered by. Show it only part of the picture and it guesses again. Drag the window: left, how it read that patch while seeing everything; right, how it reads it through the window alone. Scan maps where it is unsure and tells you how far to trust a depth map of this picture.
What you are looking at. Bright = near in the depth panels. A depth reader only knows relative depth, so the window's reading is first rescaled to best match the whole-picture reading; what is left over is the change of mind. The percentage is the share of the layout that does not survive that rescale (1 − r²). The steadiness score averages it over every window of a scan, giving more weight to windows with more depth structure (flat sky or bare wall says little and would add noise). Everything runs in this browser; nothing is uploaded.
What it means, and does not. Most photographs barely waver; a mirror is the exception, because it puts a second space inside the picture. Densely etched prints (Piranesi's prisons, Hogarth) waver a lot. In the full study, MiDaS, the annotator ControlNet ships, wavered more on dense hatching than Depth-Anything did; the browser build of MiDaS did not reproduce the study's MiDaS (Beer Street, centre window: 19% here, 48% in Python), so it is not offered here. In tests, that wavering did not visibly damage tiled image generation when each tile was anchored to the picture (img2img), so treat it as the reader's uncertainty, not as guaranteed harm. It matters most where depth is used on its own: 3D or parallax effects, relighting, text-to-image ControlNet without the picture.
The write-up: Where depth changes its mind. The instrument it grew from: Depth Holonomy.
From the depth-holonomy studies. Model: Depth-Anything-V2-Small via transformers.js 3.5.1 (checked against the Python model: centre-window values within a few points).
© 2026 Russell Parrish / A.rtist I.nfluencer / Parallax Metrology. All rights reserved. The photographs are by Russell Parrish. No part of this system, visual material, or accompanying documents may be reproduced, distributed, or transmitted in any form or by any means, including AI training datasets, without explicit written permission from the creator. More info: www.artistinfluencer.com and www.parallaxmetrology.com