RVM — cut the background out of a video, on your own PC

Give it a video, get back the person with the background gone.
No green screen. No upload. No subscription. No watermark. It runs on your computer.

uv tool install git+https://github.com/metiu1/rvm-web
rvm

RVM is a small app you install with one command. It opens in your browser like a website, but there is no website: everything runs on your own machine, and your footage never leaves it.

You pick a video, press start, and it removes the background — the same trick a green screen gives you, except you do not need the green screen. What comes out is real transparency (an actual alpha channel), so you can drop the clip straight into Premiere, DaVinci Resolve, After Effects or OBS and put whatever you like behind it.

You can also just cut out a photo, upscale a blurry video to 4×, or trim, crop, speed up and compress a clip — the tools you keep opening a different website for, all in one place, offline.

• Creators, streamers and editors who want the "no green screen" cutout without paying a monthly fee per export.
• Anyone whose footage cannot be uploaded — client work, NDAs, faces of people who did not consent to a stranger's server, anything you would rather keep on your disk.
• People with long or high-resolution video, which is exactly what free online tools refuse to process.

Web background removers cap the resolution, stamp a watermark on the result, put you in a queue behind everyone else, charge by the minute, and keep a copy of your video on their servers.

This one has no cap, no watermark, no queue, no account and no server. Your graphics card does the work, your disk keeps the file. If you do not have a GPU it still runs, just slower.

Sidebar navigation with six tools:

• Video background removal (RobustVideoMatting) — MobileNetV3 (fast) or ResNet50 (accurate); export as MP4 composition, alpha mask, or PNG sequence with real RGBA transparency
• Image background removal (rembg / U2Net) — upload a photo, get a transparent PNG, with before/after preview
• Image upscaling — 2× / 4×, Lanczos (no extra deps) or Real-ESRGAN (AI)
• Video upscaling — 2× / 4×, Lanczos or Real-ESRGAN, original audio preserved (muxed back via bundled ffmpeg)
• Image editing (Pillow) — crop, resize, rotate/flip, brightness/contrast/saturation/sharpness, filters (B&W, sepia, invert, blur, sharpen, auto-enhance), format convert + quality (PNG/JPG/WEBP), live before/after preview
• Video editing (bundled ffmpeg) — trim, crop, resize, change speed, compress, format convert, export GIF, extract audio, mute, reverse — with live progress

Plus:

• GPU accelerated (CUDA) with automatic CPU fallback
• Upload files directly from the browser (saved to ~/Downloads/rvm-uploads/)
• ffmpeg ships bundled via imageio-ffmpeg — no system install required
• Installable in one command via uv

Requires uv.

uv tool install git+https://github.com/metiu1/rvm-web
rvm

A browser window opens automatically at http://localhost:7860. (The old rvm-web command still works as an alias.)

On Windows, PyPI serves CPU-only torch by default. To get GPU acceleration, tell uv to auto-detect your CUDA driver and pull matching wheels:

uv tool install --reinstall --torch-backend=auto git+https://github.com/metiu1/rvm-web

Or set it once for every future install (persists across terminals):

# PowerShell (Windows) — set once
[Environment]::SetEnvironmentVariable('UV_TORCH_BACKEND','auto','User')

# bash/zsh (Linux/macOS) — add to your shell profile
export UV_TORCH_BACKEND=auto

Verify CUDA is active — the upscaling log shows Modello caricato su cuda:0, or check directly:

uv tool run --from rvm python -c "import torch; print(torch.cuda.is_available())"

Real-ESRGAN AI upscaling is built-in — no extra packages needed. Model weights download automatically on first use to ~/.cache/rvm_web/. imageio-ffmpeg (bundled ffmpeg, used to keep audio when upscaling video) is also installed automatically.

The image section uses rembg with the U2Net model.
On first use, the model (~170 MB) is downloaded automatically to ~/.u2net/.
Output is a PNG with full alpha transparency.

Both image and video upscaling support 2× and 4× with two methods:

• Lanczos — classic high-quality resampling, no extra dependencies, fast.
• Real-ESRGAN — AI super-resolution, built-in (no extra packages). Model weights download automatically on first use to ~/.cache/rvm_web/.

For video upscaling the original audio track is preserved: the video is upscaled frame-by-frame, then the audio is muxed back with ffmpeg (imageio-ffmpeg, bundled). If no audio track exists the video is written directly.

RVM's inference_utils.py passes frame rate as a string to av.add_stream().
Newer av versions require a Fraction. The fix is applied automatically on first run.

• RobustVideoMatting by Peter Lin et al.
• Web UI built with FastAPI

MIT