rembg

Remove image backgrounds in R. An R port of the Python rembg package by Daniel Gatis.

It runs pre-trained segmentation models (U-2-Net, ISNet, BiRefNet, BRIA RMBG) through the ONNX Runtime to predict a foreground alpha matte, then composites a cutout with a transparent (or solid-colour) background. Inference goes through onnxr (a cpp11 binding to ONNX Runtime, no Python). Images are decoded with jpeg/png and everything else is base-R array math, so the dependency footprint is small.

Install

# 1. the ONNX Runtime binding + the runtime library (one-time, ~35 MB CPU build)
install.packages("onnxr")
onnxr::onnx_install()

# 2. this package
# install.packages("rembg")   # once on CRAN

Use

library(rembg)

# simplest: path in, RGBA array out (default model "u2net")
cutout <- rembg("photo.jpg")
png::writePNG(cutout, "cutout.png")

# reuse a model across many images
sess <- new_session("isnet-general-use")
rembg("a.jpg", session = sess, out = "a.png")
rembg("b.jpg", session = sess, out = "b.png")

# just the mask, or a solid background
mask  <- rembg("photo.jpg", only_mask = TRUE)
white <- rembg("photo.jpg", bgcolor = c(255, 255, 255, 255))

# soft hair/fur edges via alpha matting (slower)
soft  <- rembg("photo.jpg", alpha_matting = TRUE)

input accepts a file path, a raw vector of PNG/JPEG bytes, or a numeric [h,w,c] array. Output is an [h,w,4] array in [0,1] by default, raw PNG bytes with output = "raw", and a written PNG when out is a path.

Models

rembg_models()

u2net, u2netp, u2net_human_seg, silueta, isnet-general-use, isnet-anime, birefnet-general, birefnet-general-lite, birefnet-portrait, birefnet-dis, birefnet-hrsod, birefnet-cod, birefnet-massive, bria-rmbg, u2net_cloth_seg, sam.

u2net_cloth_seg segments clothing rather than the salient subject, returning one mask per garment class. Pick a cloth_category ("upper", "lower", "full") or omit it for all three stacked vertically:

sess <- new_session("u2net_cloth_seg")
top  <- rembg("person.jpg", session = sess, cloth_category = "upper")

sam (Segment Anything) is click-to-segment: give it point prompt(s) and it segments the object you point at. Points are (x, y) pixel coordinates; labels are 1 (foreground) or 0 (background), defaulting to foreground:

sess <- new_session("sam")
obj  <- rembg("scene.jpg", session = sess, points = c(400, 165))

Bring your own model: the u2net_custom / dis_custom / ben_custom presets run a local .onnx through a fixed preprocessing profile, or pass size/mean/std directly:

sess <- new_session("u2net_custom", model_path = "~/.u2net/my_model.onnx")
# or fully explicit:
sess <- new_session(model_path = "my.onnx", size = 1024,
                    mean = c(0.5, 0.5, 0.5), std = c(1, 1, 1))

Models download on first use into a per-user cache (tools::R_user_dir("rembg", "cache")). The first download asks you to confirm the location; in non-interactive use, opt in with options(rembg.download = TRUE) or REMBG_DOWNLOAD=1. Set U2NET_HOME to override the location — e.g. to ~/.u2net to share the Python rembg cache.

Notes

• Needs a GPU? new_session(model, backend = "cuda") (requires a CUDA runtime build via onnxr::onnx_install(cuda = TRUE)).
• Fidelity: masks match Python rembg closely (≈99.9% pixel agreement) but are not bit-identical. We resize with bilinear where PIL uses LANCZOS; the difference shows only at soft edges (hair, fur).
• JPEG EXIF orientation is honoured, so sideways-stored phone photos are corrected on read (like PIL’s exif_transpose).
• alpha_matting = TRUE refines edges with closed-form matting (a port of pymatting); the alpha matte matches Python to ≈0.9999 correlation. It’s slower and needs the Matrix package.
• This is a complete port of upstream rembg’s models and features.

License

MIT. Original Python rembg © Daniel Gatis (MIT).