Image Depth Estimator
Turn one photograph into a relative depth map. A locally hosted, quantized Depth Anything V2 Small model evaluates the photo in a browser worker. Adjust the display range and save a grayscale PNG at the source dimensions. Bright and dark values compare model scores within this image; they are not measured distances or a calibrated 3D scan.
Key features
- Infer continuous relative depth from a single photo with a real ONNX model
- Choose a 392px or 518px maximum model input edge while retaining image aspect ratio
- Adjust lower and upper score stops without rerunning inference
- Export a grayscale PNG at the original photo dimensions with a clear score legend
- Keep photo pixels on the device and verify the pinned model checksum
How to use
- Choose a nonanimated PNG, JPEG or WebP photo within the listed limits.
- Select fast or detailed model resolution and estimate depth locally.
- Compare the original with the grayscale map and read its relative-score legend.
- Move the lower and upper score controls to reveal a useful display range.
- Download the PNG and avoid interpreting gray levels as physical meters.
Use cases
- Explore rough foreground/background ordering in a travel photo
- Prepare an artistic relative-depth reference for parallax design
- Compare how one model interprets a photo before manual editing
- Share a grayscale depth-style image without sending the original to an API
Frequently asked questions
Does a pixel value tell me its distance in meters?
No. This model produces an uncalibrated relative score for one image. The grayscale and range controls are display transforms; neither provides real distance, camera geometry or a measured 3D surface.
What is the difference from background removal or a texture height map?
Background removal estimates a subject mask; a texture height map interprets brightness as a surface assumption. This tool runs a monocular scene-depth model and produces a continuous relative map.
What does brighter mean?
Brighter pixels have higher model scores after the selected display range. The legend compares values only within this photo. It does not certify a physical near/far boundary, especially where the model makes errors.
Is the photo sent to a server?
No. Photo pixels are decoded locally and passed to a local browser worker. The first run downloads model and WASM files from this site's origin; no third-party model CDN or photo API is used.
Why can the map be wrong?
A single photo is ambiguous. Transparent or reflective surfaces, weak texture, tiny details, unusual perspective and image edits can mislead the model. The exported pixels also interpolate a smaller model output to source size, adding no new detail.
What files and sizes are supported?
Use one nonanimated PNG, JPEG or WebP file no larger than 8 MiB or 2 megapixels, with both sides at least 64px and an aspect ratio no wider than 4:1. The model input is capped at 392px or 518px on its longer edge; browser memory and speed vary.
Privacy
The photo is decoded and processed in your browser and is not posted to a site API. On first use the browser fetches the approximately 27.3 MB model and 13 MB ONNX WASM runtime from this site's own origin. Subsequent availability without a network connection depends on browser caching; cold-start offline use is not guaranteed.
Comments & questions