Technical

Chrome 94+, Edge 94+, and Opera 80+ provide the best experience with WebGPU acceleration. Firefox 90+ and Safari 16.4+ work via WebAssembly (slower inference, same output quality). iOS Safari 16+ is supported; older versions show a compatibility notice.

No. Each model downloads once on first use and is cached in your browser (IndexedDB / Cache Storage). After that the tool loads instantly and works offline. Different tools use different models, so the first run of a new tool may download its model.

The AI models (background removal, image upscaling) cache in your browser after first use. To reclaim the space, use the "Clear session" button in the app, or clear site data in your browser settings (Settings → Privacy → Site Data). The models simply re-download the next time you use a tool that needs them — clearing never breaks functionality.

Yes. The classical utilities (filters, crop, resize, convert, compress) run instantly on any device, including phones — no download. Background removal and image upscaling use ML models that run on most laptops with 4 GB+ RAM; they are faster with any GPU but fall back to the CPU (WebAssembly) automatically with identical output.

Yes. Background removal offers Fast (~0.5s/frame) and Best Quality, which re-runs BiRefNet for maximum accuracy on hair, fur, and complex edges (~1.5s/frame). Image upscaling offers an instant WebGL Lanczos mode and a higher-quality AI mode (Swin2SR).

Most image tools finish in 1–3 seconds once the model is loaded. Fast background removal is ~0.5s/frame and Best Quality (BiRefNet) ~1.5s/frame. Video and upscaling scale with length/resolution and show a live progress bar with ETA. The first run of any tool also includes a one-time model download.

Yes. PNG, WebP, and AVIF exports carry a true straight-alpha channel, so cutouts open transparent in Photoshop, Figma, and similar tools. For video, transparent exports use WebM (VP9, yuva420p) which After Effects and Premiere import with their alpha intact; MP4 has no alpha channel, so use WebM when you need transparency.

No. WebGPU (modern Chrome/Edge) makes things faster, but every tool falls back to WebAssembly on the CPU automatically with identical output — just slower. The larger models (Best Quality background removal, AI upscaling) benefit most from a GPU.

That was the AI model for that tool (sizes are shown on each tool page). It is cached after the first run, so subsequent uses are instant and offline. Classical tools that need no model show "Runs instantly — no download".

The tool uses Transformers.js to run ONNX-format neural networks directly in your browser. On supported hardware, inference runs on your GPU via WebGPU. On others, it falls back to multi-threaded or single-threaded WebAssembly. No cloud API is used.

The AI model weights need to download the first time (~80 MB for Fast, ~180 MB for Best Quality). After that, they are cached in your browser and subsequent runs are fast. A progress bar shows the download status.

Images up to 4096 × 4096 pixels are processed at full resolution. Larger images are automatically downscaled for inference, then the full-resolution mask is upscaled back — so your export is always at your original dimensions.

Upload all images at once by selecting multiple files or dragging a group. They enter the queue and process sequentially. When done, click "Export all" to download a ZIP file containing all results.

PNG, JPG/JPEG, WebP, AVIF, and HEIC (iPhone photos). Animated images (animated GIF, animated WebP) are supported but only the first frame is processed.

No. Video processing (background removal, filters, upscaling, stabilization) all run in your browser. Your video never leaves your device. We have no server infrastructure capable of receiving video files.

Input: MP4 (H.264/H.265), WebM (VP8/VP9), and MOV. Chrome and Edge support all three; Firefox and Safari have limited MOV support. Output: all video tools export WebM (VP9). The Video Editor can also export MP4.

For a 30-second clip at 720p with the Fast model, expect 2–5 minutes on a WebGPU-capable browser (Chrome). Best Quality model is 2–4× slower. Longer clips, higher resolution, and WASM fallback all increase processing time significantly.

Two models: Fast (RMBG-1.4, ~80 MB) is optimized for speed with good general accuracy. Best Quality (BiRefNet, ~180 MB) produces finer edge detail, especially for hair, fur, and complex outlines. Both run entirely in your browser via ONNX Runtime.

Yes. The image editor, video editor, and staging tool all share the same layers panel — drop multiple files for instant overlay layers, reorder, lock, duplicate, change opacity, and pick from 12 blend modes (multiply, screen, overlay, soft-light, and so on). Keyboard shortcuts: Ctrl+J duplicates, Ctrl+Shift+] / [ moves up / down, Delete removes the selected layer.

The image editor, video editor, and staging tool each have a 3D mode powered by Three.js. You can switch the active layer to a flat 3D plane (with light, material, and background controls) or a depth-displaced relief mesh, then capture a 4-second 360° orbit recording as WebM. The 3D bundle is dynamic-imported, so it only loads when you open it.

Yes. The editor exports a .nss-project file that bundles the original image, the current mask, every filter / background / layer, and the operation history. Reopen on any device and continue exactly where you left off — original layer IDs and ordering are preserved.

Yes. The AI panel handles natural-language commands like "remove background, place on marble studio, upscale 2x, export PNG" and chains them automatically. Recipe goals such as "make this professional for e-commerce" expand into multi-step plans. The assistant ships with a rule-based parser; you can optionally enable an in-browser LLM (Qwen 0.5B, Llama 1B, or Phi-3 mini) — see the AI panel's settings.

Start or join a session from the editor; edits stream peer-to-peer over WebRTC. Only operation descriptions ("apply filter cinematic", "remove background") cross the wire — pixel data never leaves your device. Our signalling endpoint at /api/collab-signal only exchanges WebRTC offer / answer, never media.

Yes. The Tuning Profiles panel takes 5–10 reference images of your product category (glassware, jewellery, hardware, etc.) and derives ΔE, erosion radius, and guided-filter overrides applied in the inference worker. No model retraining; everything happens in-browser.