ClearCut v2.0: Building the Creative Studio of the Future

It has been 18 months since we launched the first version of ClearCut. When we started, we had a simple but radical conviction: image editing tools should not require users to upload their files to third-party servers. Every image you process through a cloud-based service is a file that could potentially be stored, analyzed, or used to train third-party models. For a designer working under NDA, a photographer with sensitive client images, or simply a person who values their privacy, that is simply unacceptable.

With ClearCut v2.0, we not only fulfill that original promise—we take it to a level that, a few years ago, would have seemed like science fiction: complete Artificial Intelligence models running inside your browser, using your own device's hardware.

The Technical Challenge: AI in the Browser—How Is It Possible?

To understand what we've achieved, it's important to grasp the technical problem that needed solving.

The Traditional Obstacle

Modern Machine Learning models are computationally intensive. A professional-quality image segmentation model can have tens of millions of parameters and require massive matrix operations for each inference. Historically, the web browser was completely inadequate for this task: slow, without access to specialized hardware, and with severe memory restrictions.

The solution until now was obvious but costly in terms of privacy: send the image to a server with specialized GPUs, process it there, and return the result. Millions of images traveling through the internet every day.

The Revolution: WebAssembly + WebGPU

WebAssembly (Wasm) is a technology that allows running low-level code (compiled from C++, Rust, or Go) directly in the browser, with execution speed close to native code. Frameworks like ONNX Runtime Web and TensorFlow.js have leveraged Wasm to run ML models on the client side.

WebGPU is the next leap. Approved as a web standard in 2023 and with widespread support since 2024, WebGPU provides low-level access to the device's GPU from the browser. This means the same hardware acceleration previously only possible in native applications is now available on a web page.

The combination is powerful: our AI models, compiled to WebAssembly and executed through WebGPU, reach inference speeds that compete with cloud-based solutions, but with zero network latency and complete privacy.

Model Optimization for the Edge

We can't simply take a 500MB model designed for server-grade GPUs and run it in a browser. The adaptation process is complex engineering that includes:

Model Quantization: Reducing the precision of model weights from 32-bit float (FP32) to 8-bit integer (INT8) or even 4-bit. Some precision is lost, but the model goes from 200MB to 50MB with minimal and imperceptible quality degradation.
Pruning: Removing the less important neural connections from the model, reducing its complexity without a noticeable impact on result quality.
Knowledge Distillation: Training a smaller model (the "student") to mimic the behavior of a larger model (the "teacher"), transferring knowledge in a compressed form.

The result of this process in ClearCut v2.0 are models weighing between 15MB and 80MB. They are downloaded once, cached in the browser, and subsequently work completely offline.

ClearCut v2.0's New Flagship Tools

Magic Eraser: Object Removal with Background Reconstruction

The most anticipated feature from our community. "Object removal" or "content-aware fill" has existed in desktop tools like Photoshop for years, but implementing it in the browser in a private, cost-free manner is a different challenge entirely.

How it works:

Area selection: The user paints over the object they want to remove using our selection brush. The system automatically suggests the object's boundaries.
Contextual analysis: The AI model analyzes the content surrounding the selected area: textures, patterns, colors, and lighting.
Generative reconstruction: Using inpainting techniques based on compressed diffusion models, the model "guesses" and synthesizes what should be behind the eliminated object, creating a visually coherent reconstruction.

Real-world use cases:

Removing unwanted people from tourism photos
Removing objects from product photos (tags, wrinkles, reflections)
Cleaning backgrounds in real estate photos (visible outlets, cables)
Correcting session errors without needing to reshoot

What can't it do? The magic eraser has its limits. Reconstructing large areas with complex, non-repetitive textures (such as a crowd of people) remains a challenge. The best results are obtained on relatively small areas over uniform or periodic texture backgrounds.

SVG Optimizer: From Chaos to Clean Code

SVG files exported from vector design tools (Figma, Illustrator, Sketch) are often code disasters. They contain:

Unnecessary editor metadata (hidden layers, unused font definitions)
Unsimplified transforms (transform="translate(100, 50) rotate(45)" instead of direct coordinates)
Redundant paths with duplicate control points
Auto-generated IDs that provide no semantic value
Empty groups and elements outside the viewport

All of this translates into SVG files that weigh 5 or 10 times more than necessary, degrading web performance.

Our optimizer applies more than 40 distinct transformations:

Removes unnecessary metadata and comments
Converts primitive shapes (rect, circle) to their most efficient representation
Simplifies paths using the Ramer-Douglas-Peucker algorithm with configurable tolerance
Combines transforms into direct coordinate values
Applies ID compression (replacing them with shorter identifiers)
Removes attributes with values that match SVG standard defaults

Typical result: File size reduction of 50-70%, with a visual difference imperceptible to the human eye.

EXIF Metadata Cleaner: Your Privacy, Restored

When you photograph something with your smartphone, the image doesn't only contain the pixels you see. Invisibly embedded in the file is a block of data called EXIF (Exchangeable Image File Format) that can reveal:

GPS Coordinates: Exact latitude and longitude where the photo was taken, with meter-level precision.
Device Make and Model: iPhone 16 Pro, Samsung Galaxy S25, etc.
Exact Date and Time: Down to the second.
Camera Settings: Aperture, shutter speed, ISO.
Editing Software: What tool you used to process the image.
Owner or Artist Name: If configured on the device.

This information is a treasure for malicious actors. If you post a photo taken in your home on social media, someone with the right technical knowledge can extract your exact address in seconds.

Our EXIF cleaning tool offers:

Full clean: Removes all metadata with one click.
Selective clean: Choose exactly which fields to remove. For example, you may want to keep copyright information but remove GPS location.
Metadata preview: Before cleaning, inspect exactly what information your image contains.
Batch processing: Clean dozens of images simultaneously.

ClearCut v2.0 Performance: The Numbers

Metric	v1.0	v2.0	Improvement
Initial load time	3.2s	1.9s	-40%
Processing time (512px)	1800ms	320ms	-82%
Processing time (2048px)	N/A	890ms	New
Peak RAM usage	480MB	210MB	-56%
Lighthouse Performance score	72	96	+33%

The Privacy by Design Principle

Before closing, we want to talk about something that goes beyond technical features. Privacy is not a feature of ClearCut; it is its fundamental architecture.

The concept of Privacy by Design, coined by Dr. Ann Cavoukian in the 1990s and today elevated to a legal requirement in multiple jurisdictions (GDPR in Europe, CCPA in California), establishes that privacy protection must be incorporated into the system design from the very beginning, not added as a patch at the end.

In ClearCut v2.0, this means:

Zero image telemetry: We never know what images you process.
Data-free business model: Our business model does not depend on monetizing user data.
Open source: Our code is publicly auditable. You don't have to take our word for it.
Offline functionality: Once the models are loaded, the tool works without an internet connection. Your work cannot be intercepted in transit because there is no transit.

ClearCut v2.0 is not just a software update. It is the realization of a vision: that the world's most powerful creative tools should not cost the privacy of those who use them.