How Our Free AI Tools Run Without an API: Background & Watermark Removal in the Browser
Our background remover, passport photo maker and watermark remover use real AI models — but there's no API bill and your photo never leaves your device. Here's how open-source models running in the browser make that possible, and the gotchas we hit.
People reasonably assume that any 'AI' tool sends your file to a server and pays an AI company per use. Ours don't. The background remover, passport photo maker and watermark remover on tools.howautomate.com run small, open-source AI models directly inside your browser. There's no API key, no per-image cost, and the photo you choose never leaves your phone or laptop.
The models
Background removal uses MODNet, a portrait-matting model (Apache-2.0 licence) that outputs how much each pixel belongs to the person. The passport photo maker adds UltraFace, a tiny face detector (MIT licence), so it can find your chin and the top of your head and frame the photo to the official proportions. The watermark remover uses MI-GAN from Picsart AI Research (MIT licence), an inpainting model: you paint over the watermark and it fills that area with what should plausibly be behind it. All three together are about 55 MB — small by AI standards, which is the whole point.
The engine
The models run with ONNX Runtime Web, which executes them using WebAssembly — the same fast, sandboxed code format browsers use for games and video editors. The first time you use a tool, your browser downloads the model once and caches it; after that it works almost instantly, even with a weak connection.
Lesson 1: test quality before you shrink the model
A 'quantised' version of the background model is four times smaller, so we tried it first. It erased half of the person in our test photos. We shipped the full-precision model instead and told users the honest download size. A fast tool that gives bad results isn't saving anyone time.
Lesson 2: keep heavy work off the main thread
Our first watermark remover ran the model directly on the page and froze it for 5–12 seconds — the browser even offered to kill the tab. Moving inference into a Web Worker (a background thread) kept the page responsive, with a progress bar instead of a frozen screen.
Lesson 3: multi-core needs special headers — and they bite
WebAssembly can use several CPU cores, but only on pages the browser considers 'cross-origin isolated', which requires two security headers. We enabled them only on the watermark pages, cutting processing from 4–10 seconds to about 2–4. Then it broke in production: the browser refused to start the worker because the worker's own script file didn't carry the same header — something our local preview server never showed, because it sends no headers at all. The lesson: test security headers on the real host, and add a timeout so a blocked worker shows an error instead of an endless 'loading model…'.
Lesson 4: AI fills look best when you blend the edge
Inpainting a box of sky can leave a faint rectangle where the fill meets the original photo. Fading from the original to the AI fill across a few pixels at the edge made the seam disappear. Small details like this matter more than a bigger model.
What it costs us
Nothing per use. The only cost is bandwidth for the one-time model download from our host, which is why the models are cached aggressively. If you're weighing an AI feature for your own product, ask whether it really needs a cloud API — for a lot of image tasks, a small model on the user's device is cheaper, faster after the first load, and far better for privacy. And use these tools only on photos and documents you own or have permission to edit.
At HowAutomate we build practical AI into business workflows — sometimes with large cloud models, often with smaller ones that run where the data already is.
Frequently Asked Questions
Do HowAutomate's AI tools upload my photo?
No. The background remover, passport photo maker and watermark remover run open-source AI models inside your browser. Your photo stays on your device.
Which AI models do the tools use?
MODNet (Apache-2.0) for removing backgrounds from portraits, UltraFace (MIT) for face detection in the passport photo maker, and MI-GAN (MIT) for filling in watermark areas. They run with ONNX Runtime Web using WebAssembly.
Why is the first use slower?
The browser downloads the model once — roughly 26–28 MB per tool — and caches it. Later uses start almost instantly.
Is there a cost per image?
No. Because the model runs on the visitor's device, there's no AI API bill per image. The only cost to us is hosting bandwidth for the one-time model download.
Amit Singh
Founder, HowAutomate — Data Engineering, AI Automation & Cloud Infrastructure
Amit has 6+ years of experience building data pipelines, AI agents, and automation systems for businesses across India and globally. He founded HowAutomate to make enterprise-grade automation accessible to growing businesses.
Get Weekly Automation Tips
Real scripts, workflows, and AI tips — straight to your inbox.
Want us to implement this for you?
Book a free 30-minute discovery call and we'll map out exactly how to apply this to your business.
More Articles
Building a Marketplace Price Tracker for Amazon, Flipkart, Myntra & Nykaa (What Actually Works in 2026)
Read more
Automating Amazon Vendor Central Direct Fulfillment: Invoices, Labels, Handover & Ship Confirmation
Read more
Exam Photo & Signature Size 2026: IBPS, NEET, JEE Main & SSC (Verified From Official Notices)
Read more