---
title: "Depth Estimation - Create an AI Depth Map"
description: "Estimate relative scene depth from a photo with on-device AI. Create and download a private color depth map with WebGPU or CPU fallback."
url: https://candoya.com/tools/depth-estimation/
locale: en
type: browser-tool
---
# Free Depth Estimation Tool
Category: [Social & Creator](https://candoya.com/tools/social/)
## What is monocular depth estimation?
Monocular depth estimation uses one ordinary image to predict which parts of a scene appear nearer or farther away. This browser tool turns that prediction into a color depth map on your device. It shows relative scene structure, not calibrated distance in metres, feet, or any physical unit.
**The tool runs Depth Anything V2 Small through Transformers.js in a dedicated browser worker.** It tries WebGPU with half-precision weights first and switches to quantized WebAssembly inference on the CPU if graphics acceleration is unavailable or fails. The image itself is never sent to a CanDoYa analysis API.
Choose Inferno for a perceptually ordered map, Turbo for strong visual separation, or grayscale for workflows that expect a single brightness channel. Inversion changes only the display and download. It does not rerun the model or change the underlying relative-depth values.
## How to use it
1. **Choose a photo** - Drop in one JPEG, PNG, or WebP image up to 12 MB and 16 megapixels, or load the built-in scene example.
2. **Estimate the depth** - Click Estimate depth. The first run downloads the public model, then a worker performs inference with WebGPU or the CPU fallback.
3. **Style and save the map** - Compare the result, switch between Inferno, Turbo, and grayscale without rerunning AI, optionally invert near and far colors, then download a PNG.
## Who it's for
- **3D artists and motion designers** creating an initial depth mask for parallax, fog, blur, or compositing experiments.
- **Photographers** visualizing how a monocular model separates foreground subjects, middle distance, and background scenery.
- **Computer vision students** exploring dense depth prediction without installing Python, CUDA, or a local model environment.
- **Prototype developers** testing whether relative depth could support segmentation, scene understanding, or accessibility ideas before integrating a model.
## Expert note
Monocular depth has an unknown scale and shift: two pixels can be ranked as nearer and farther, but their numeric difference is not a reliable physical distance. Reflective surfaces, transparent objects, mirrors, thin structures, unusual camera views, and scenes outside the model's training distribution can produce convincing-looking but incorrect geometry.
## Frequently asked questions
### Is my image uploaded to a server?
No. Your image is decoded and passed to a worker inside your browser for inference and rendering. The worker fetches public Transformers.js runtime and model files from external hosts, but those requests do not include the image. CanDoYa does not receive or store the source or depth map.
### Is the depth estimation tool free?
Yes. There is no account, payment, credit allowance, watermark, or download limit imposed by the tool. Your own device performs the computation. The first run downloads public runtime and model files, so it can use significant data and processing power.
### What image limits apply?
The tool accepts one JPEG, PNG, or WebP file up to 12 MB. Each side must be between 32 and 4096 pixels, and the complete image cannot exceed 16 megapixels. These limits reduce browser memory spikes while still supporting detailed photos.
### Does the result show distance in metres or feet?
No. The map represents relative, scale-ambiguous depth inferred from a single view. It can indicate that one region appears nearer than another, but it cannot recover a trustworthy physical distance without camera calibration, known reference dimensions, stereo views, or another source of metric scale.
### How accurate is AI depth estimation?
Depth Anything V2 can recover useful scene structure from many everyday photos, but the result is still a prediction. Mirrors, glass, water, textureless walls, fine edges, extreme perspectives, and unfamiliar scenes can be wrong. Inspect the map before using it for editing, measurement, navigation, or safety decisions.
### What do the colors mean?
Colors encode normalized model output across this image, from the smallest response to the largest. Inferno and Turbo make changes easier to see, while grayscale is convenient for masks. Invert swaps the color direction. Because values are normalized per image, colors should not be compared as fixed distances across different photos.
### Does depth estimation work without WebGPU?
Yes, when the browser supports WebAssembly and module workers. The worker checks for a usable WebGPU adapter first. If that path is unavailable or inference fails, it retries with quantized model weights on the CPU. The fallback stays on-device but is usually slower.
### Can I use it offline?
A later run may work after the browser caches the runtime and model, but offline availability is not guaranteed. Browsers can evict cached files, private sessions may discard them, and cleared site storage can require another download. Load and test the model before relying on it without a connection.
## Related tools
- [Object Detection](https://candoya.com/tools/object-detection/)
- [Pose Detection](https://candoya.com/tools/pose-detection/)
- [Image Captioner](https://candoya.com/tools/image-captioner/)
- [Image Upscaler](https://candoya.com/tools/image-upscaler/)
- [Background Remover](https://candoya.com/tools/background-remover/)
---
*Markdown edition of , published for AI agents and other automated readers. Index of key pages: .*