hjLabs AI Playground
Starting the on-device runtime…
This page runs Depth Anything V2 on your GPU to predict a depth value for every pixel of the photo you drop in, then lifts those pixels into 3D: each pixel becomes a point, its colour comes from the photograph and its distance from the camera comes from the depth map. The result is rendered with WebGL, so you can orbit it, zoom into it and export it. The whole chain — the model download, the depth inference and the rendering — happens inside your browser tab, so the photo is never uploaded anywhere.
Be clear about what this is, because the phrase "image to 3D" is used for several very different things. What you get here is a 2.5D projection from a single viewpoint, not a watertight mesh and not photogrammetry. One photograph contains no information about the back of the subject or about anything hidden behind it, so those regions are simply absent — orbit far enough round and you see the stretched edges and the empty space where occluded geometry would be. Photogrammetry solves that by triangulating dozens of overlapping photographs; novel-view synthesis (NeRF, Gaussian splatting) solves it by fitting a radiance field to many views. Both need far more input and far more compute than a browser tab, and neither is what this tool does.
What a single-view depth projection is genuinely good for: parallax and "3D photo" animation, fake depth-of-field and fog, depth-driven compositing and relighting, quick pre-visualisation, and getting a feel for how a scene is laid out in space. It is also the fastest way to see monocular depth estimation working, since the point cloud makes errors in the depth map obvious in a way a grayscale image never does.
image to 3dphoto to 3d online freedepth map to point cloud2d to 3d converter onlinemonocular depth estimation3d photo effect generatordepth anything v2 3dsingle image 3d reconstructionwebgl point cloud viewerfree image to 3d no signup
No. It is a 2.5D projection from one viewpoint. A single photo has no data about occluded or back-facing surfaces, so the point cloud has holes and stretched edges where those would be. Photogrammetry needs many overlapping photographs and is not something a browser tab can do.
You can export the point cloud and the depth map. There is no watertight mesh to export, because the underlying data is a depth image rather than a closed surface — meshing it would invent geometry the photograph never contained.
Depth Anything V2 Small, the same ONNX model behind the depth estimation page, in a roughly 100 MB download. It predicts relative inverse depth, so the scale is arbitrary — near is near and far is far, but nothing is measured in metres.
A browser with WebGPU for the depth model (Chrome or Edge 113+, Safari 26+) and WebGL for the viewer. Point counts scale with image resolution, so a very large photo is downsampled before projection to keep the orbit smooth.
No. The model weights are downloaded from Hugging Face to your browser once, and every inference after that runs on your own GPU through WebGPU. Your text, images and audio are never sent anywhere — hjLabs.in has no inference server and no way to see your input.