How it works

Frame extraction

Your video plays to an offscreen canvas while every frame is captured at full resolution. No data is sent to any server.

Per-frame AI inference

Each frame is passed to the RMBG-1.4 (Fast) or BiRefNet (Best Quality) segmentation model running in a WebAssembly/WebGPU worker. The model produces a Float32 alpha mask per frame.

Temporal smoothing and encode

Each mask is blended with the previous frame's mask (EMA smoothing) to reduce flicker. Frames are composited and encoded as WebM with alpha or MP4 with a solid/image background.

Import and decode

Your video is decoded in the browser. You can continue from a background-removed WebM or start with any raw footage.

Edit: filters, text, trim

Apply LUT colour presets at adjustable intensity, add text overlays with font and position control, and trim start/end points on the timeline. All edits are non-destructive while the editor is open.

Export

The edited video is re-encoded and downloaded. Output is MP4 or WebM depending on input and browser capabilities.

Model selection

Choose AI mode (Swin2SR — true detail synthesis) or Instant mode (Lanczos + unsharp mask — fast geometric upscale). AI mode produces better detail on photos and faces.

Tile-based processing

For AI mode, the image is divided into overlapping tiles. Each tile runs through the neural network independently and the results are assembled with seamless blending at boundaries.

Export

The upscaled image is exported at 2× or 4× the input resolution (capped at 4096 px per side). PNG output preserves all detail without lossy compression.

Queue upload

Upload up to 100 images. They are queued and processed one at a time using the same AI model as single-image removal.

Sequential processing

Each image runs through the full pipeline: inference → Float32 mask → edge decontamination → straight alpha export. Failed items can be retried individually.

ZIP download

When all images are processed, download a ZIP containing all results in your chosen format (PNG, WebP, or AVIF).

Depth estimation

The MiDaS depth estimation model runs on your image, producing a per-pixel depth map. Areas farther from the camera receive more blur.

Depth-guided blur

A Gaussian blur is applied with intensity proportional to each pixel's estimated depth. The in-focus subject stays sharp while the background softens progressively.

Export

The blurred composite is exported as PNG or JPEG. The blur radius, focal point, and blend amount are adjustable before export.

GIF decode

The animated GIF is decoded into individual frames in the browser, preserving frame timing and metadata.

Per-frame AI removal

Each frame runs through the same AI pipeline as still images. For single-color backgrounds, a color-key threshold path is used. The transparent mode always uses the full AI mask path.

Re-encode and export

Processed frames are re-encoded into a GIF with full transparency preserved. Frame timing from the original is maintained.

Camera stream capture

Your webcam stream is captured at up to 720p. No video is recorded or stored — frames are processed and discarded in real time.

Real-time inference

The AI model processes frames at up to 10 fps. Each frame runs through the RMBG model in a Web Worker with the mask applied to a composited canvas.

Virtual camera output

The processed canvas is available as a virtual video source that can be selected in video call software. No driver installation is required.

LUT application

A Look-Up Table (LUT) maps every input RGB triplet to an output RGB triplet in a single pass over all pixels. The table is applied using WebGL or Canvas 2D depending on browser support.

Intensity blend

The filtered result is blended with the original using linear interpolation at the chosen intensity (0–100%). At 50%, you get a 50/50 mix of filtered and original.

Export

The final image is exported at full input resolution as PNG or JPEG.

Format detection

The browser checks which output formats its MediaRecorder supports. Unsupported options are greyed out automatically.

Real-time transcode

Your video plays to a canvas. A MediaRecorder captures the canvas stream and encodes it in the target format using the browser's native codec.

Export

The encoded file is assembled from chunks and offered as a download in the target container format.

Quality preset selection

Choose from High (4 Mbps), Medium (1.5 Mbps), Low (600 Kbps), or Very Low (200 Kbps) target bitrate.

VP9 re-encode

Your video plays to a canvas captured by a MediaRecorder set to the VP9 codec at the target bitrate. The video plays in real time as it encodes.

Export

The compressed WebM file is offered for download. File size depends on content complexity — talking head footage compresses more than fast-moving action.

Target resolution

Select a preset (4K, 1080p, 720p, 480p, 360p, 240p) or enter custom width and height. Aspect ratio can be maintained or overridden.

Scaled canvas encode

The video plays to a canvas sized at the target resolution. The browser scales each frame using bilinear interpolation as it encodes.

Export

The resized WebM file is offered for download at the specified dimensions.

Aspect ratio calculation

Given your source video dimensions and the target ratio, the tool calculates the minimum canvas size that fits the video and achieves the target ratio.

Padded canvas encode

The video plays centered on a larger canvas filled with the chosen padding color. A MediaRecorder captures the full canvas including the padding.

Export

The padded WebM is offered for download. The video content is preserved at its original resolution — only the canvas is larger.

Rotation transform

The target canvas dimensions are computed by swapping width and height for 90°/270° rotations, or keeping them for 180°.

Per-frame transform

Each frame is drawn to the rotated canvas using a CSS transform (rotate). The canvas captures the correctly oriented frame.

Export

The rotated video is encoded and downloaded as WebM. Convert to MP4 using the Format Converter if needed.

Three sequential encodes

Your video plays through once per format — VP9, VP8, and AVC — for a total of three full passes. Each produces a separate output blob.

Size measurement

Each output blob is measured and displayed with its file size, a size bar relative to source, and percentage difference from the original.

Download any or all

Each format has its own download button. Choose the smallest or the most compatible for your use case.

Frame collection

All frames are captured to offscreen canvases in a first pass — no encoding happens yet.

Motion estimation and smoothing

Block-matching between adjacent frames produces per-frame translation offsets. A sliding-window average smooths the trajectory. Scene cuts are detected and handled independently.

Compensated encode

Each frame is re-rendered with the inverse of its smoothed offset applied, then encoded as a stabilized WebM.

Background removal

Your product photo is processed by the same AI model as the background remover. The subject is extracted with a Float32 alpha mask.

Scene compositing

The extracted product is placed onto the selected lifestyle scene with six blending passes: shadow, ambient occlusion, color match, edge feathering, vignette, and final composite.

Export

The full-resolution composite is exported as PNG with straight alpha. Blending parameters (shadow type, color match %, edge feather) are adjustable.