Remove Video Backgrounds in Your Browser — No Software, No Upload
How NSS Background Remover applies frame-by-frame AI to remove backgrounds from MP4, WebM, and MOV clips, entirely in your browser with no data upload.
Removing a background from a single photo takes a few seconds. Doing the same to every frame of a video clip — while keeping the result smooth and flicker-free — is a different problem. Here is how the NSS video background remover works, what to expect, and tips for getting the best output.
Why video is harder than images
A 30-second clip at 30fps contains 900 individual frames. Running background removal on each one independently produces a strobing effect: the mask edges jump slightly frame to frame because the model never sees two frames at once. Small differences in lighting, motion blur, or hair position can cause noticeable flicker.
The solution is temporal smoothing — an exponential moving average applied to consecutive masks. Each frame's mask is blended with the previous frame's result, which dampens per-frame noise while preserving real motion:
smoothed = 0.75 × current_mask + 0.25 × previous_mask
This alone eliminates most visible flickering without blurring edges.
The processing pipeline
- File decoding — your video is decoded frame by frame using an HTML5 video element. No data is sent to any server at any point.
- Inference — each frame is passed through the RMBG-1.4 (Fast) or BiRefNet (Best Quality) segmentation model, running as a WebAssembly/WebGPU worker.
- Temporal smoothing — the mask is blended with the previous frame's result.
- Compositing — the smoothed mask is applied to the original frame. Depending on your settings, this produces either a transparent frame or a frame composited onto a solid colour, image, or blurred background.
- Encoding — frames are written to a WebM container (VP9 + alpha channel) or an MP4 container.
All of this happens locally in your browser. Your video never touches a server.
Fast vs Best Quality
| Fast (RMBG-1.4) | Best Quality (BiRefNet) | |
|---|---|---|
| Model size | ~80 MB | ~180 MB |
| Speed | ~0.5s per frame | ~1.5s per frame |
| Edge quality | Good | Excellent |
| Hair/fur | Good | Very good |
| 30s clip at 720p 30fps | ~14 min | ~45 min |
For most social media content — talking heads, product demos, short clips — Fast mode produces results good enough to publish. Switch to Best Quality for footage with complex hair, fine fabric, or semi-transparent elements.
Output formats
WebM with alpha channel — preserves true transparency. The background is stored as an alpha channel in the VP9 stream. Use this if you plan to composite the video in editing software or on a web page. Note: WebM alpha is supported in Chromium-based browsers and Premiere Pro / DaVinci Resolve. Safari requires a conversion step.
MP4 with background — composites the selected background directly into each frame. Produces a standard H.264 MP4 that plays everywhere. Use this for final delivery.
Tips for best results
Start with short clips. The tool works on any length, but a 10–30 second clip is a practical starting point. You can always trim in your phone camera app before dropping the file.
Shoot on a plain background. Even a solid-colour wall or a bedsheet dramatically improves AI accuracy. You don't need a green screen — any consistent background the model hasn't seen before helps.
Use Fast mode first. It's 3× faster and often good enough. If you see choppy edges around hair or complex subjects, re-run with Best Quality.
Lower the resolution if speed is critical. 720p processes roughly 4× faster than 1080p. If you need a quick result, export from your phone at 720p before uploading.
Export as WebM for editing. If you are going to composite the result in video editing software, export as WebM with alpha. This preserves all the flexibility — you can add any background in post.
What doesn't work well (yet)
- Very fast motion — motion blur at the subject boundary confuses the segmentation model. Slower-moving subjects produce cleaner results.
- Transparent or reflective subjects — glass, reflective jewellery, and semi-transparent fabric are difficult for any segmentation model. The morphological close pass helps fill interior gaps, but edges near highly transparent material may require manual touch-up.
- Clips longer than 2 minutes — the tool handles longer clips, but processing time scales linearly. A 2-minute 720p Fast-mode clip takes roughly 30 minutes.
Upcoming improvements include frame deduplication (skipping identical frames to save time), WebGPU inference (3–5× faster on supported hardware), and a per-frame brush editor in the video editor.