Track, Mask, Cut Out, or Blur Any Object in Video

Yes. AI Video Masking is completely free, requires no signup, and adds no watermark to your exported video.

Upload a video, use the timeline below the preview to find a frame where the object is clearly visible, and click it. Choose your outputs and hit Track Object. By default, masking covers the full video, including frames before your selection.

No. You can select the object on any frame where it is clearly visible. With Full video selected, the tool uses that selection to mask both earlier and later frames. Frames where the model finds no object have no object mask.

Yes. For a later-frame selection, choose From selected frame under Mask coverage. Masking starts on that frame and runs to the end; earlier frames have no object mask. If you select the first frame, masking already covers the full video, so these choices are hidden.

Yes. Instead of drawing a mask frame by frame, click the subject once and SAM2 automatically propagates the mask through the video. You can then export the tracked result as a transparent cutout, blur, black & white matte, or white alpha matte.

Foreground clicks (green dots) tell the model which pixels you want to keep. Background clicks (red dots, added by right-clicking or switching to background mode) tell it which pixels to exclude. Adding a background click near an unwanted region can help refine a noisy selection. The selection preview updates instantly after every click.

The tool accepts MP4, MOV, and WebM files up to 600 MB. Shorter, lower-resolution clips process faster and are more reliable on modest GPUs.

Not currently. Mobile browsers are disabled because full video tracking needs desktop memory and processing power. Use an up-to-date desktop version of Chrome or Edge.

Meta's Segment Anything 2.1 video model — the same architecture behind state-of-the-art video object segmentation. Unlike image-only ports, Fuzz Puppy runs the complete video pipeline including SAM2's memory attention: the model remembers what your object looks like across recent frames, follows it through motion, and recovers it after occlusions. Everything runs locally, using WebGPU when available and CPU compatibility mode if GPU processing is unavailable or fails.

No. Your video never leaves your computer. Tracking and encoding run in your browser.

After the first download (~190 MB), the model is cached in your browser's storage so later visits can reuse the download. Nothing is installed on your computer, and your browser clears the cache automatically if disk space runs low. To remove it manually, use the 'Delete downloaded model' button at the bottom of the tool, or your browser's Clear Site Data option.

The first run downloads the model (about 190 MB), which your browser caches for later use. Masking and export then run locally on your GPU or, in compatibility mode, your CPU. CPU masking can take longer. Shorter clips and lower resolutions reduce processing time.

With Full video selected on a later frame, the tool first masks backward from your selected frame to the beginning. It then returns to your selected frame and masks forward to the end. The frame counter follows this processing order; your exported video plays in its original order. A first-frame selection only needs the forward pass.

Browsers throttle background tabs. Keep the tracker tab visible while it processes so frame extraction and AI inference run at full speed.

No. Every export keeps the full video length and original frame order. If you choose From selected frame, earlier frames have no object mask: cutouts and white alpha mattes are transparent, black & white masks are black, and inverted alpha mattes are white. Object blur leaves earlier frames unchanged; background blur blurs the whole frame.

Four outputs, in any combination from a single tracking run: an object cutout (just the tracked object on transparency, with audio), a blurred MP4 (blur the object or background, with audio), a black & white mask MP4 (for luma/track mattes and compatible AI video models), and white alpha mattes (standard, inverted, or both). Object cutouts and white alpha mattes each offer WebM or MOV inside their output box. WebM is the default when your browser supports transparent export.

To place the object onto other footage, use the object cutout. To hide a face, license plate, or logo — or to fake a shallow depth-of-field — use the blurred MP4. For full control, use the black & white mask as a luma matte / track matte in Premiere Pro, After Effects, DaVinci Resolve, Final Cut Pro, or CapCut. You can also use it as a mask input for compatible AI video models and workflows, including masked generation, inpainting, object replacement, and region-specific effects. The white alpha matte is for stylized glow, outline, and stencil-reveal effects.

Both preserve transparency. WebM uses VP9 compression for smaller files and works in compatible browsers and editors. MOV preserves lossless alpha video for editing, with larger files. You can choose independently for the object cutout and white alpha matte. If your browser cannot export transparent WebM, the tool selects MOV and explains why.

The MOVs use a lossless alpha codec that QuickTime Player previews poorly. They work correctly in editors (CapCut, Premiere Pro, After Effects, DaVinci Resolve, Blender). To make a ProRes 4444 copy for QuickTime playback, run:

The object cutout (WebM or MOV) and the blurred MP4 keep the source audio. The mask and white alpha matte outputs are video-only — they're mattes, and your original clip keeps its audio on another track in your editor.

Start from a frame where the object is clearly visible, even if it appears later in the video. Add object points to refine the selection and background points to exclude nearby distractions before hitting Track Object. Fast motion, heavy blur, and long periods where the object is hidden can make tracking less reliable.

The tool uses CPU compatibility mode if WebGPU is unavailable. If GPU initialization or tracking fails or takes too long, it restarts once on CPU with your selected frame and points preserved. CPU processing can take longer. If that also fails, your selection stays saved for a manual retry. Updating your browser and GPU drivers and enabling hardware acceleration can improve performance.

Yes. Frames where the model finds no object have no object mask: black & white masks are black, and cutouts are transparent. The tracker can pick up the object again when it becomes visible, though long gaps can reduce accuracy. After processing, the tool reports how many frames had no visible object.