AI background removal uses deep learning to automatically separate a subject from its background, outputting a transparent PNG or a flat JPG/WebP. Upload an image (PNG, JPG, or WebP, up to 10 MB), choose a model, and review results with the built-in before/after slider. The RemBG 1.4 model includes Alpha Matting, which refines the transition zone at pixel level for tricky edges like hair and fur.
Output Format: When PNG Matters and When It Doesn't
- PNG: Full alpha channel preserved — required for compositing onto a new background, design tools, or printing
- JPG / WebP: Transparent areas become white; useful when you only need an opaque image and want a smaller file
For JPG and WebP the quality slider (20–99, default 95) controls compression. Keep quality at 85 or above when clean edges matter — lower values introduce compression noise that shows most clearly along the cutout boundary.
When Alpha Matting Changes the Result
Alpha Matting is exclusive to RemBG 1.4. It analyzes the gradual transition zone at the subject's edge rather than making a hard cut — which makes a visible difference in these cases:
- Portrait hair, especially flyaways and fine strands
- Animal fur and feathers
- Thin fabric and lace
- Subjects with slightly blurred edges
Three parameters control how the edge is judged:
- Foreground threshold (default 240): higher values classify more pixels as subject — if the cutout is missing edge detail, lower this slightly
- Background threshold (default 10): higher values remove background more aggressively — if faint background fringe remains, raise this value
- Erode size (default 10): controls the width of the edge-smoothing zone — too high causes fine strands to thin or break
Adjust one parameter at a time and compare results before touching another.
Getting Cleaner Cutouts
The subject should occupy at least 60% of the frame and the background should be solid or low-texture. When the subject is partially blocked by another object the model infers the boundary from what is visible — the more occluded, the more likely gaps appear at the edge. For multi-subject scenes where you only need one, crop to that subject before uploading.
What Produces Unreliable Results
- Subject and background in very similar colors (off-white subject on beige) — the model relies on contrast; near-zero contrast produces boundary errors
- Glass, smoke, water — semi-transparent materials typically cannot have their transparency accurately preserved
- Multiple overlapping subjects — the model picks a "primary" subject and may choose the wrong one