Most online image tools follow the same pattern:

1. upload an image;
2. wait for a server to process it;
3. download the result.

That model works well. Cloud APIs are convenient, scalable from a developer experience point of view, and often give very consistent results.

But while building an open-source background remover, I kept running into a different question:

How far can image processing move into the browser itself?

Not just the UI. Not just cropping or resizing. Actual AI-powered background removal, export options, batch queues, and image composition — all running locally on the user's device.

This article is a practical comparison between cloud-based image processing and client-side AI image processing, based on the tradeoffs I encountered while building BG-Zero, an open-source browser-based background remover.

This is not a "cloud is bad" article. Cloud APIs are still the right answer for many products. But browser-side processing opens up a very different set of tradeoffs around privacy, cost, UX, and architecture.

The two models

At a high level, image processing tools usually fall into one of two categories.

Cloud API model

The image is uploaded to a backend or third-party API. The server runs the model or calls another service, stores or streams the result, and returns the processed file.

Client-side model

The image stays inside the browser. The model runs with browser technologies such as WebAssembly, WebGPU, Canvas, Web Workers, or JavaScript inference libraries.

The second model sounds attractive, but it is not automatically better. It simply moves complexity from the backend to the browser.

A quick comparison

The important thing is not which column "wins". The important thing is matching the architecture to the product promise.

If the product promise is enterprise automation, centralized file management, and guaranteed throughput, cloud APIs make sense.

If the product promise is privacy-first, local-first, low server cost, and user-controlled processing, browser-side processing becomes much more interesting.

Why try image processing in the browser?

For a background remover, the privacy argument is obvious: images are often personal. A single upload might be a portrait, an ID-style photo, a product image before launch, an internal team asset, a screenshot, a document, a medical or otherwise sensitive image, or creative work owned by a client.

For many casual tools, uploading is fine. But there are plenty of cases where a user may want to remove a background without sending that image to another server.

That was the main reason I started experimenting with BG-Zero as a local-first image tool.

The goal was:

Can the browser do the full workflow?

- load the image;
- run background segmentation;
- generate a transparent result;
- let the user choose export options;
- support formats like PNG, JPEG, WebP;
- handle batch queues;
- avoid uploading the image for processing.

The answer is: yes, but with tradeoffs.

Lesson 1: Model loading is part of the product experience

When you use a cloud API, model loading is mostly invisible to the end user. The server or provider handles warmup, caching, model selection, and hardware acceleration.

In the browser, model loading becomes part of your UX.

The first visit may require downloading model files, WASM assets, tokenizer/config files, or other runtime resources. Depending on the engine, these assets may be cached afterward, but the first load still matters.

For a browser-based AI tool, the loading state cannot be an afterthought.

A bad loading experience looks like this:

Upload image → click button → page appears frozen → user leaves

A better experience needs:

• clear engine loading state;
• progress feedback where possible;
• fallback messaging;
• retry paths;
• honest expectations;
• caching for repeat visits.

In BG-Zero, I ended up treating model loading as a first-class part of the interface rather than a hidden implementation detail.

That is a bigger UX surface than a simple POST /remove-background API call.

Lesson 2: Browser memory is the real limit

Image files are deceptive.

A file might be only a few megabytes on disk, but once decoded into pixels, it can occupy much more memory.

For example, a large image may become:

width × height × 4 bytes

That is just raw RGBA pixel data. Add canvases, intermediate masks, object URLs, model tensors, previews, and export buffers, and memory pressure becomes very real.

This matters even more for batch processing.

A cloud service can queue files and process them with controlled infrastructure. In the browser, you have to be careful not to process too much at once.

For BG-Zero's batch mode, that means:

• a queue instead of parallel processing everything at once;
• per-image status;
• retry for failed items;
• cleanup for preview/result URLs;
• ZIP creation in the browser for completed results.

This is less glamorous than the AI model itself, but it is what makes the tool usable.

Lesson 3: Privacy-first features need precise wording

One subtle UX problem: users often associate "sign in" with "upload".

For a heavier feature like batch processing, sign-in may be useful for abuse prevention or access control. But if the image processing still runs locally, the product needs to communicate that clearly.

Bad wording:

Sign in to process your images

This can sound like images are being sent somewhere.

Better wording:

Sign in to access batch mode. Your images are still processed locally in your browser.

For privacy-first tools, the implementation is not enough. The copy must explain the boundary:

• What is local?
• What is sent to the server?
• What is not uploaded?
• Why is sign-in required, if it is required?
• Are analytics collecting image content? They should not.

In my case, I found that privacy is both a technical feature and a communication problem.

Lesson 4: Export is not just a download button

Background removal sounds like a single operation, but the output workflow can branch quickly.

After removing a background, users may want:

• transparent PNG;
• WebP with transparency;
• JPEG with a white background;
• product photo on a square canvas;
• ID-style photo with a blue, white, red, gray, or custom background;
• a manual refinement step;
• batch ZIP export.

That means export logic becomes part of the product architecture.

For example:

• PNG can preserve transparency.
• JPEG cannot preserve transparency, so it needs a background fill.
• WebP can support transparency, but platform support and user expectations vary.
• Product photos may need padding, centering, ratio control, and optional shadow.
• ID-style photos need careful wording because official requirements vary.

A simple "download result" button eventually becomes a composition pipeline: the app starts from the foreground mask or processed result, optionally fills a background, composites everything on a canvas, applies the selected format and quality settings, exports a Blob, and finally generates a filename for download.

This is one of the places where browser-side processing is actually pleasant: Canvas, Blob, object URLs, and local downloads give you a lot of flexibility without involving a server.

But it also means you need to test edge cases carefully.

Lesson 5: Multiple engines are useful, but they complicate the UI

In cloud products, model selection is often hidden. The provider may choose the model automatically or expose a simple quality/speed option.

In a browser-based tool, multiple engines can be useful because browser capabilities vary. One user may have WebGPU available on a fast desktop machine, while another may only have WASM on a memory-constrained phone or a browser with different support for advanced APIs.

BG-Zero currently experiments with multiple background removal engines, including libraries such as @imgly/background-removal, @huggingface/transformers, and @bunnio/rembg-web.

For developer tools, exposing engine choices can be acceptable. For consumer tools, too many choices can feel confusing. My current preference is to provide a sensible default, expose engine selection only for advanced users, explain the tradeoffs briefly, and avoid making people understand ML internals before they can remove a background.

What I built as an experiment

BG-Zero is my attempt to explore the local-first side of this tradeoff.

The project is an open-source background remover built with Next.js, React, TypeScript, WebAssembly/WebGPU-based inference libraries, Canvas export flows, and browser-side image handling.

The main idea is simple:

Remove image backgrounds without uploading the image for processing.

Some of the workflows include:

• automatic background removal;
• manual cleanup for difficult images;
• transparent PNG export;
• background color replacement;
• WebP/JPEG export options;
• product photo white-background composition;
• ID-style photo background drafts;
• batch processing with local queues;
• ZIP export for completed batch results.

The project is not meant to prove that every image tool should be browser-only. It is more of a practical exploration of what modern browsers can handle today.

Final thoughts

Client-side AI image processing is not a universal replacement for cloud APIs.

It is slower on some devices, harder to test across browsers, and more complicated on the frontend. Batch processing needs careful memory management. Model loading becomes part of the product experience. Error handling cannot be hidden behind a backend job queue.

But the upside is real:

• images can stay local;
• server costs can stay low;
• tools can work more like privacy-first utilities;
• users can get more control over what happens to their files;
• open-source projects can offer useful AI features without running expensive inference infrastructure.

For image tools, I think this architectural choice will become more common.

Not because the browser is always the best place to run AI, but because for some workflows, the browser is the most honest place to run it.