About Image Cutout

Image Cutout cuts any object out of a photo when you click on it. A Segment Anything model runs entirely inside your browser: the photo is encoded once, then every click re-runs a small mask decoder in about a tenth of a second. Nothing is uploaded, there is no account and no server-side processing.

How to use it

1. Open a photo (button, drag-and-drop, paste, or a demo). The encoder runs once; the left card shows the timing.
2. Left-click the object to add a positive point. Right-click the background (or switch to Negative mode and click) to add a negative point. The purple mask updates after every click. Two or three points are usually enough; if the mask grabs a neighbour, a negative point on the neighbour separates them.
3. Confirm commits the mask to the result. Repeat for more objects. Undo removes the last point; Reset clears the current points.
4. Zoom: both the work area and the result zoom independently, with the buttons in each panel's corner, the mouse wheel, a trackpad pinch, a two-finger pinch on a touchscreen, or the + / - / 0 keys. Once a photo is larger than its panel you can drag or swipe it to pan, and the middle mouse button always pans; a plain click or tap still places a point. Fit resets a panel to the best fit.
5. Background: the swatches under the result switch between transparent, white, a blurred copy of the photo, colour and gradient presets, any custom colour, or your own uploaded image (cover-fitted to the photo). The preview and the download include it.
6. Download PNG saves the cutout at the photo's resolution with the chosen background (transparent by default); Download JPG appears for opaque backgrounds.

Design canvas — background & collage

After you cut out a subject, press ＋ Design (on the result panel) or the Design button in the header to open the design canvas. There you can:

• Replace the background of the whole design: transparent, solid white, any colour, a two-stop gradient, or your own uploaded image (cover-fit) — the same choices as under the result panel.
• Collage several subjects: press ＋ Design again for each new cutout — they stack as layers. Drag a layer to move it, drag its corner to resize, use Forward/Back to reorder, and Delete (or the Delete key) to remove the selected one.
• Pick a canvas size: 1:1, 4:3, 16:9 or 4:5.
• Export the finished design as PNG (keeps transparency) or JPEG. Everything is composited in your browser; nothing is uploaded.

What runs where

• Models: two Segment Anything image encoders, chosen in the left panel. SlimSAM-77 (uniform; the default, tagged Fast) is a structurally pruned SAM ViT-B whose encoder is in the ViT-Tiny size class (about 9 M parameters). SAM ViT-Base (tagged Quality) is Meta's original ViT-B image encoder (about 86 M parameters): cleaner edges and fewer misses on small or low-contrast objects, at the price of a larger one-time download and slower encoding without WebGPU. Both use the same prompt encoder and mask decoder (the files are byte-identical, so they are downloaded once). Weights come as ONNX from Xenova's exports and are downloaded once from SkillSafe's shared, vetted model registry at models.skillsafe.ai, then held in your browser cache, which every SkillSafe app using the same file shares.
• Runtime: onnxruntime-web 1.27.0. On browsers with WebGPU the fp16 weights run on the GPU (SlimSAM-77 about 21 MB, SAM ViT-Base about 189 MB); elsewhere the int8 weights (about 14 MB and 106 MB) run on single-threaded WebAssembly, where encoding a photo takes a few seconds with SlimSAM-77 and up to a minute with SAM ViT-Base. Every click after that re-runs only the small mask decoder.
• Resolution: the model works at 1024 px on the long side; the mask is resampled to the photo's own size. Photos over 4096 px on the long side are reduced to 4096 px first.
• Network: only this origin's static files and the weight files above. The page's Content-Security-Policy, set by the platform, forbids every other connection.

Limits

• Edges are Segment-Anything quality: clean on solid objects, soft on hair and fur, not a full matting model.
• One click selects one region; thin structures and glass need a negative point or two.
• Very small or very low-contrast objects may need a click near their centre.

Credits & licences

• SlimSAM — Zigeng Chen et al., SlimSAM: 0.1% Data Makes Segment Anything Slim (NeurIPS 2024), Apache License 2.0. ONNX export Xenova/slimsam-77-uniform, served unmodified from the registry. Provenance and hashes: /models/NOTICE.txt; licence text: Apache-2.0; platform attribution page: /models.txt.
• Segment Anything (SAM ViT-Base) — Alexander Kirillov et al., Segment Anything (ICCV 2023), Meta AI, Apache License 2.0. ONNX export Xenova/sam-vit-base, served unmodified from the registry; hashes in /models/NOTICE.txt.
• onnxruntime-web 1.27.0 — Microsoft, MIT licence, vendored unmodified.

The demo photos are CC0 / public-domain images from Wikimedia Commons, resized for this app. No attribution is legally required for CC0 works; sources are credited anyway:

• Two coffee cups — Close-up of espresso machine with two brown coffee cups (CC0).
• Couple — Happy Couple by freestocks.org (CC0).
• Portrait — Brunette woman portrait by Christopher Campbell (CC0).
• Bird — Perched blue and orange bird by Vincent van Zalinge (CC0).

Built by the SkillSafe app fleet. Concept inspired by browser-side SAM demos; implementation original. Back to the tool.