LaMa Inpainting in Your Browser

LaMa (Large Mask inpainting) is a neural network published by Suvorov et al. at WACV 2022, “Resolution-robust Large Mask Inpainting with Fourier Convolutions”. It became the default open-source inpainting model because it handles the cases where classic algorithms fall apart: big holes, irregular textures, and structures that cross the masked region.

• Fourier convolutions (FFC). Regular convolutions only see a small neighborhood, so structure fades inside large holes. LaMa mixes local convolution with a global Fourier transform, giving every pixel awareness of scene-wide patterns — periodic textures like fences, bricks and blinds survive the hole.
• Trained with large masks. During training, huge random masks force the network to invent coherent structure instead of just smearing nearby pixels.
• High-receptive-field perceptual loss. The training objective rewards matching textures at scene scale, not just pixel-by-pixel similarity.

The result is a generative model that stays close to your photo: it borrows aggressively from the surroundings and synthesizes texture only where there is nothing to borrow. It is far less prone to inventing phantom objects than smaller GAN models — but it is still generative, so check the filled area before shipping the result.

The LaMa weights are exported to ONNX and executed by onnxruntime-web, which compiles the network to WebAssembly running in a Web Worker on your device. The tool above bundles the packing, padding and compositing: your image is scaled to at most 512 px, padded to the model's fixed 512 × 512 input, filled, and cropped back.

Because the model is large, the first run downloads about 208 MB once and the browser caches it — the following runs work offline. Unlike online generators, there is no upload step, no queue, no per-image cost and no watermark. If you want the zero-download alternative, the site's default inpainting tool fills small holes instantly with patch matching.

Both tools run locally and keep your photo private; they differ in what they fill well. The fundamentals behind both are covered in our guides to image inpainting and the inpainting algorithm.

• Choose LaMa when the hole is hard. Large regions, complex or irregular textures, and structures crossing the hole — cables, railings, brick courses, a person standing in front of foliage. LaMa's global context reconstructs what patch matching cannot.
• Choose patch matching when the hole is easy. Small marks backed by simple or repetitive surfaces — walls, grass, water, sky, bokeh — fill instantly with no download, and the result is deterministic: the fill only reuses real pixels from your photo.
• When in doubt, try both. They share the same interface, so the fastest workflow is a patch-matching pass first, and LaMa for whatever is left.

Where LaMa sits in the wider research landscape — from classic patch methods to transformers and diffusion — is mapped in our complete guide to inpainting techniques.

• Cover the whole object plus a small margin — including the shadow it casts. LaMa is good at extending structure through slightly larger masks.
• The working image is capped at 512 px on the long edge (the model's native resolution). Fine detail beyond that scale is beyond the model's reach, so results look best at this working size.
• The first run needs a one-time ~208 MB model download plus a few seconds of loading; after that the session stays warm and repeat runs skip straight to inference. Desktop browsers are recommended — the model needs a few hundred MB of memory.