✂️ Annihilate-BG — Free AI Background Remover

A fully free, no-paywall, no-signup background remover that runs entirely in your browser. No images are ever sent to a server. Full resolution output. No watermarks.

Built as a better, free alternative to remove.bg — powered by open-source AI.

---

✨ Features

• 100% Free — no subscriptions, no paywalls, no watermarks, ever
• Full Resolution Output — exported PNG matches the exact pixel dimensions of your input
• Private by Design — all processing happens locally in your browser; your images never leave your device
• Smart Detection — automatically uses instant flood-fill for logos and flat graphics, and the AI model for photos
• Batch Processing — upload multiple images at once with a live queue showing each image's status
• Non-blocking UI — AI inference runs in a Web Worker so the page stays responsive during processing
• Cancel Anytime — stop a batch mid-way; completed images are sorted to the top automatically
• Download All as ZIP — get all processed images in one click
• Multiple Export Formats — save results as PNG, JPG, or WEBP
• Drag & Drop + Paste — drop files, paste from clipboard (Ctrl+V), or use the file picker
• Model Caching — the AI model (~175MB) downloads once and is cached locally; subsequent runs are near-instant
• PWA Support — installable as a home screen app on mobile and desktop
• Up to 100MB per file, up to 20 images per batch (soft warning at 11+)

---

🧠 How It Works

Annihilate-BG uses two processing paths, chosen automatically per image:

Path 1 — Flood Fill (logos, flat graphics, screenshots)

Detects solid-colour backgrounds using corner sampling and removes them with a BFS flood fill. Runs in milliseconds. No model download required.

User uploads image
       ↓
Heuristic analysis on a 120x120 downscale
(corner consistency + background coverage + colour variance)
       ↓
Flood-fill BFS seeded from all 4 corners
       ↓
Background pixels made transparent
       ↓
Full-resolution transparent PNG ready to download

Path 2 — AI Segmentation (photos, complex images)

Uses RMBG-1.4 by BRIA AI running via Transformers.js (ONNX Runtime / WebAssembly). Inference runs in a Web Worker so the UI never freezes.

User uploads image
       ↓
Image transferred to Web Worker as ArrayBuffer
       ↓
RMBG-1.4 runs inference at 1024x1024 internally
       ↓
Mask data transferred back to main thread
       ↓
Mask upscaled to original resolution + composited on Canvas
       ↓
Full-resolution transparent PNG ready to download

Why no resolution is lost

The AI model processes a 1024x1024 version of your image internally, but the resulting mask is upscaled back to your original pixel dimensions before compositing. A 4000x3000px photo stays 4000x3000px in the output.

---

🗂️ Project Structure

/
├── index.html              # Markup only — no inline JS, no inline CSS
├── manifest.json           # PWA manifest (home screen install, theme colour)
├── css/
│   └── styles.css          # All styles
├── js/
│   ├── main.js             # Entry point — UI events, job cards, ZIP, cancel, modal
│   ├── queue.js            # Batch queue — parallel flood-fill, serial AI via worker
│   ├── worker.js           # Web Worker — runs RMBG-1.4 inference off the main thread
│   ├── detect.js           # Image type heuristics (flood vs AI)
│   ├── floodFill.js        # BFS flood-fill algorithm
│   └── aiRemoval.js        # RMBG-1.4 fallback for browsers without module worker support
└── icons/
    ├── icon-192.png        # PWA icon
    └── icon-512.png        # PWA icon (large)

---

🖥️ Running Locally

Since the JS files use ES modules (import/export), they must be served over HTTP — opening index.html by double-clicking won't work.

npx serve .

Then open http://localhost:3000 in your browser.

Safari may have issues with large WASM files and module Web Workers. Chrome, Edge, Firefox, or Vivaldi are recommended.

---

🔢 Batch Limits

Flat graphics (flood-fill) are essentially free and run in parallel regardless of count. The limits exist to protect users on slower or mobile devices from accidentally queueing dozens of AI photo jobs.

---

🛠️ Tech Stack

---

📄 License

The application code in this repository is MIT licensed — free to use, modify, and deploy.

The RMBG-1.4 model is provided by BRIA AI under their own model license. It is free for non-commercial use. Please review their terms before using in a commercial product.

---

🙋 FAQ

Does this work offline? After the first run (which downloads and caches the AI model), photos work fully offline. Logos and flat graphics always work offline since they use no model.

Is my image uploaded anywhere? No. Everything runs in your browser tab. No network requests are made with your image data at any point during processing.

Why is the first load slow for photos? The RMBG-1.4 model is ~175MB and needs to download once. After that it is cached in your browser and future runs are near-instant. Logos and flat graphics are always instant regardless.

What is the difference between Instant mode and AI mode? Instant mode uses flood-fill — it detects the background colour from the image corners and removes it in milliseconds. AI mode uses the RMBG-1.4 neural network and is better for photos with complex backgrounds.

Why does the UI stay responsive during processing? AI inference runs in a Web Worker — a separate thread — so the main thread (which handles the UI) is never blocked. The spinner animates, cards update, and the cancel button stays clickable throughout.

Can I cancel a batch? Yes. Click the Cancel button during processing. The current image finishes, then the queue stops. Completed images move to the top of the queue automatically.

What browsers are supported? Chrome, Edge, Firefox, and Vivaldi. Safari may work but has known issues with large WASM files and module Web Workers — Chrome is recommended for best results.

Can I use this commercially? The app code is MIT. However, the RMBG-1.4 model has a non-commercial clause — check BRIA AI's license before using in a paid product.