LaMa
Resolution-robust large-mask inpainting using fast Fourier convolutions; the classic open baseline for removing big unwanted regions.
About
LaMa (WACV 2022, arXiv 2109.07161) is the large-mask inpainting model from Samsung AI Center Moscow and EPFL that became the standard open baseline for filling in big missing regions. Its core idea is fast Fourier convolutions, which give the network an image-wide receptive field so it can complete periodic structures and coherent backgrounds across holes that defeat conventional CNN inpainters. It was trained at only 256x256 yet generalizes surprisingly well to higher resolutions at lower parameter and compute cost than competing baselines.
The official code lives at github.com/saic-mdal/lama under Apache-2.0, with a popular community mirror (advimman/lama) carrying over 10k stars. It powers countless watermark removers, object erasers, and photo-restoration tools, including ONNX-exported builds that run on CPU. Development on the original repo has largely stalled since 2022, but the model remains the go-to lightweight inpainting backbone and a fixture in commercial eraser products.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (arXiv 2109.07161): state-of-the-art across inpainting datasets at publication, strong on periodic structures; widely adopted as a comparison baseline
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: paper arXiv 2109.07161 (Samsung AI Center Moscow, EPFL), official repo saic-mdal/lama with Apache-2.0, community mirror advimman/lama at 10k+ stars, third-party docs confirming Apache-2.0 model license. Strengths: seminal open inpainting work, genuinely permissive license, huge downstream adoption, runs on CPU. Docked for: (1) original repo largely unmaintained since 2022; (2) benchmark numbers are author-reported from 2021 and predate diffusion inpainters; (3) no official hosted API. License: Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 4.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · WACV 2022, arXiv 2109.07161 (Samsung AI Center Moscow, EPFL)
- · Official repo saic-mdal/lama; community mirror advimman/lama at 10k+ stars
- · Original repo largely unmaintained since 2022; widely embedded in eraser products
- · Trained at 256x256, generalizes to higher resolutions