Technology

Technology stack

This browser-based AI background remover is built for transparent, technical operation and privacy-first local image processing. Technology on nobg.eu is framed for operators who need durable process notes, not a slogan page. Use the sections below as a working brief before you standardize the editor in a team workflow.

Core stack

Next.js for app routing and static generation.
TypeScript for strict typing and maintainability.
ONNX Runtime Web for client-side inference.
WebGPU acceleration where available.
Fallback runtime behavior for broader browser support.

Definition: client-side inference

Client-side inference means model execution happens on the user device in browser runtime, not by sending the image to a remote inference API for core processing.

Supported image workflow

Typical inputs include PNG, JPG, and WebP. Transparent outputs are exported as PNG or WebP depending on user settings.

Application layer

The public product is a Next.js application with TypeScript. Routing, metadata, and localized educational pages are first-class—not an afterthought bolted onto a single canvas demo. That structure lets us ship trust pages (About, Privacy, Editorial standards) beside the editor without maintaining a disconnected marketing microsite.

On-device inference stack

Segmentation uses ONNX Runtime Web with a WebGPU acceleration path when the browser exposes a compatible adapter, plus WASM/CPU fallback when it does not. The privacy posture stays the same across paths: inference for the core edit is designed to run in the page session after model assets load, not as a mandatory upload to a cutout API.

Assets, caching, and cold starts

Model and runtime assets download over HTTPS like other modern web apps. Cold starts feel slower; warm caches feel snappier. Document that difference for stakeholders who otherwise demand a single vendor millisecond claim. Local speed still scales with megapixels and concurrent tabs.

Quality tooling around the mask

Technology only matters if operators can inspect results. The editor emphasizes preview and export discipline—transparent PNG/WebP masters, channel flatten as a deliberate second step. We avoid synthetic benchmark tables that invent lab conditions your warehouse laptop will never see.

Security and dependency hygiene

Client bundles, dependency updates, and consent tooling are maintained as part of shipping a free ad-supported web app. Report vulnerabilities through Contact. Do not confuse third-party marketing tags (when consented) with the cutout inference path described on How processing works.

What this stack deliberately excludes

We do not require accounts for basic edits, do not present a cloud project library of your photos, and do not pretend desktop GPU magic works identically on every phone. Roadmap items that would break the local-edit promise get filtered. That product constraint is a technology choice.