Watermark Remover models
Inpainting models used by the Watermark Remover on rave-tools.com: a free tool that removes a logo or watermark from your own video or image, entirely in the browser. Nothing here runs on a server; the browser downloads one of these files once, caches it, and runs it on the device. The user's video or image never leaves their machine.
This repository is a mirror, so the tool keeps working regardless of what happens to the upstream uploads. Both files are unchanged, byte-for-byte copies of the upstream ONNX files listed below.
| File | Tier in the tool | Size | SHA-256 |
|---|---|---|---|
migan_pipeline_v2.onnx |
Fast (default) | 28,079,181 B | 6f1f3530a1a2324b19752018ce756088b07973cda8d7d890034ace5c8a48c40b |
lama_fp32.onnx |
Best (opt-in) | 208,044,816 B | 1faef5301d78db7dda502fe59966957ec4b79dd64e16f03ed96913c7a4eb68d6 |
migan_pipeline_v2.onnx: MI-GAN
MI-GAN: A Simple Baseline for Image Inpainting on Mobile Devices (ICCV 2023) by Andranik Sargsyan, Shant Navasardyan, Xingqian Xu and Humphrey Shi, Picsart AI Research.
- Original repository: Picsart-AI-Research/MI-GAN
- Copied from: andraniksargsyan/migan (the "pipeline v2" ONNX export, which takes uint8 RGB plus a mask at any resolution and does its own resizing and paste-back)
- License: MIT, Copyright (c) 2024 Picsart AI Research (PAIR). The authors publish the weights under MIT in a dedicated LICENSE-WEIGHTS file. Full text: LICENSE-MIGAN.
lama_fp32.onnx: LaMa (big-lama)
Resolution-robust Large Mask Inpainting with Fourier Convolutions (WACV 2022) by Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park and Victor Lempitsky, Samsung Research.
- Original repository: advimman/lama
- ONNX export by Carve: Carve/LaMa-ONNX (fixed 512x512 input, opset 17; exporter code in Carve-Photos/lama)
- License: Apache License 2.0, Copyright 2021 Samsung Research. Full text: LICENSE-LAMA.
- Note on the weights: the upstream repository is licensed Apache 2.0 but has no separate license file for the pre-trained weights. Its maintainers have answered license questions by pointing to Apache 2.0 (#132, #230), and Carve publishes the ONNX export under Apache 2.0. Requests for an explicit weights license are open upstream (#365, #368). If the authors publish different terms, this mirror will follow them.
License
See LICENSE.md: MIT for MI-GAN, Apache 2.0 for LaMa. All credit for the models goes to their authors; rave-tools.com only hosts this copy.