Image Background Remover
Remove background from images using AI. Works offline in your browser. Free online background eraser, no upload to server required
Removing the background of a portrait used to mean either pricey Photoshop work (lasso, refine edges, magic wand) or uploading to a SaaS like remove.bg. In 2026, U²-Net and BiRefNet neural models run in WebAssembly in your browser, producing accurate background removal locally — including the hard parts (wisps of hair, glass, semi-transparent fabric). This tool handles portrait photos, product images, and arbitrary subject extraction without uploading anything.
How modern background removal actually works
- Deep learning segmentation models (U²-Net, BiRefNet, RMBG-2.0) take an image and output a per-pixel mask indicating "subject" (1) vs "background" (0). The mask is multiplied with the input image's alpha channel.
- Quality depends on model — older U²-Net (2020) gets coarse silhouettes right but loses fine hair detail; BiRefNet (2024) preserves wisps and translucent edges.
- Cost: model size 50-200MB, runs once-per-image in ~2-10 seconds in browser depending on hardware. Reasonable for occasional use; not for batch jobs.
- For batch / production: dedicated services (remove.bg, Photoroom, Clipdrop) use server-side GPUs with newer models — faster and slightly more accurate.
Working example: a product photo
Input
A product photo: ceramic mug on a wooden table, soft shadow
Output
Removal accuracy: Hard edges (ceramic body): 99%+ Handle interior (small hole): ~95% Soft shadow: varies — model decides whether shadow is part of subject Before: 2.1 MB JPEG After: 450 KB PNG with alpha Output: 800×800 transparent PNG ready for catalog use Typical issues: - Subject has same color as background (white mug on white): low contrast → some edge artifacts. - Reflections or glass: model may include or exclude reflective glints. - Hair / fur: depends on model. BiRefNet handles wisps; older models lose them.
For e-commerce, you want shadow REMOVED to drop the product on any background. For portrait photography, you sometimes want the shadow PRESERVED for realism. Modern tools offer "preserve shadow" or "remove shadow" toggles.
What works and what does not
- Portraits with clear subject — 95%+ accuracy. Modern models handle hair, glasses, headphones.
- Products on simple backgrounds — 98%+ accuracy. White-on-white can be tricky.
- Multiple subjects (group photo) — the model treats all people as "foreground". Usually correct; sometimes one person at the edge gets cut.
- Animals — generally works. Long fur and feathers can lose edges.
- Cars, vehicles — works. Reflections in glass and metal can cause artifacts.
- Glass / transparent objects — partial. The model may treat the whole glass as foreground or only the visible outline.
- Backgrounds nearly identical to subject — fails. The model needs distinguishable subject vs background; uniform-color photos defeat it.
- Compressed / blurry images — fails. Low resolution and JPEG artifacts confuse edge detection.
Refining after automatic removal
- Brush touch-up — paint mask with brush to add/remove pixels manually. Most browser tools support this for cleanup.
- Edge refinement — feathering or sharpening the mask edge for blend with new background.
- Color decontamination — fringes of original background color often remain on subject edges. Decontamination algorithms shift edge pixels toward the new background's color.
- Shadow regeneration — if shadow was removed, add a synthetic drop shadow for naturalism when placing on a new background.
When to reach for this tool
- You need product photos with transparent backgrounds for catalog or marketplace listings.
- You are creating profile photos and want a clean white or solid-color background instead of the room behind you.
- You are preparing thumbnails for a video or article and want subject cutouts.
- You are designing slides or posters and need cutouts of people / objects for compositing.
What this tool will not do
- It will not perfectly handle every edge case. Complex transparency (chiffon fabric, water splashes, ice cubes) needs manual refinement.
- It will not replace skilled retouching. For premium photography (magazine covers, art prints), pro retouchers still do better than automatic tools.
- It will not work without a GPU for fast performance. WebAssembly model runs on CPU; expect 5-30 seconds per image on a mid-range laptop. On a phone or Chromebook, expect 30-60+ seconds. For batches, use a hosted service with GPU.
- It will not handle very large images well. Models are designed for 1024-4096px input. Multi-megapixel raw images get downsampled; very high-res output requires upscaling after, which can hide small artifacts.
Everything runs in your browser via WebAssembly. Source images and the masked output stay local. Useful for portrait photos containing PII, internal product shots, or anything you would not upload to a third-party service.
Frequently asked questions
Why does my hair look "haloed" after background removal?
The mask cut hair edges slightly inside the original silhouette, leaving fine background pixels behind. Use color decontamination (most refinement tools) to shift edge pixels toward the new background. Or use a newer model (BiRefNet, RMBG-2.0) that handles fine detail better.
Is my data uploaded anywhere?
For browser-based tools using WebAssembly models: no. The model and image stay in your tab. For SaaS services (remove.bg, Photoroom): yes, your image is uploaded to their servers. Check the service's privacy policy if uploading sensitive content (employee headshots, internal product designs).
How does this compare to Photoshop's Select Subject?
Photoshop's Select Subject (Sensei AI) is comparable quality to current free models. For typical subjects, results are similar. Photoshop offers more refinement tools (Refine Edge, Quick Mask) for manual fixes. Free tools win on price and speed; Photoshop wins on integration with the rest of your photo-editing workflow.
Can I remove backgrounds from videos?
Frame-by-frame removal is slow but works. Real-time video background removal (for video calls) uses different (faster, less accurate) models — see Zoom virtual background, Google Meet blur. For video production, use dedicated tools (Runway, Descript) with temporal coherence to avoid jittery masks.
Why is the model so big (200MB)?
Modern segmentation models have 30-100M parameters. Quantized versions (INT8) are smaller; full precision (FP16/FP32) is larger and slightly more accurate. The trade-off: smaller models = faster + lower-quality; larger models = slower + better edges. Browser tools usually ship quantized for download speed.
Can I use the output commercially?
Yes — you own the input image and the resulting mask. Some browser tools using upstream models (open-source) make output free for any use. Verify the specific tool's terms if you are publishing for paid commercial work.
Related tools
Edit images with crop, resize, rotate, filters, text. Draw shapes and annotations. Free online photo editor in browser, no upload required
Convert images between PNG, JPEG, and WebP formats. Batch conversion with quality control. Free online image format converter
Compress and resize images online without losing quality. Reduce JPEG, PNG, WebP file size. Free image optimization tool with adjustable quality
Generate favicons in all sizes (16x16, 32x32, 180x180) from any image. Create ICO and PNG favicons. Free online favicon maker for websites
Last updated · E-Utils editorial team