Hands-On: Removing Backgrounds from 30 Real Images, in the Browser
I ran a batch of product shots, portraits and thumbnails through AIGoMoon's own AI Background Remover — the U²-Net model runs locally via onnxruntime-web, and nothing ever leaves the machine. Here is the workflow that worked.
Cutting out backgrounds used to mean Photoshop, or uploading product photos to a stranger's server. After processing 30 real images with our own AI Background Remover, neither is necessary anymore.
The Setup: Nothing to Install, Nothing Uploaded
Open the tool and drop an image — that is the entire setup. The cutout runs on U²-Net, loaded through onnxruntime-web: the model is about 4.6MB, Apache-2.0 licensed, and cached by the browser after the first visit, so later images process without another download. Files up to 15MB are accepted, and everything happens on-device — the image is never uploaded, which matters when the photos are client products or team portraits. Output is a transparent PNG, or the background can be filled with a solid color in the same step.
Three Workloads, One Workflow
Product shots were the fastest wins: a white ceramic mug on a wooden desk became a clean transparent PNG in one pass, ready for a marketplace listing that requires a pure-white background. Portraits are the stress test — soft hair against a busy background is where cutout models either earn their keep or fall apart; on a plain-wall portrait the edges came out clean enough for a website team page, while a backlit shot taught me the cheaper fix: re-shoot, don't re-run. Better input beats better processing. Thumbnails were the surprise — a book cover photo, cut out and placed on a solid accent color, looked more designed than the original photo ever did.
Where to Help the Model
The batch surfaced three rules. First, contrast is everything: subject edges that clearly differ from the background in brightness or color cut out cleanly, while low-contrast edges (a gray hoodie against a gray wall) need a re-shoot, not a re-run. Second, crop before processing — the model works on what it is given, and a tight crop around the subject hands it more effective resolution. Third, choose the output for the destination: transparent PNG for anything going into a design tool, solid color for social thumbnails, and keep the original file for archiving. Total cost for the 30 images: zero dollars, zero uploads, and about as long as it takes to finish a coffee.