Rembg

Daniel Gatis's cutout tool hides a catalogue of segmentation models behind one interface: call it from a shell, import it in Python, run it as an HTTP server that also serves a Gradio page, or start the published Docker image. The CLI processes single images, whole folders (with a watch mode for new files), and raw RGB24 frames piped from FFmpeg for video. Models download on first use: the default BRIA RMBG-2.0 checkpoint is around a gigabyte, slower but sharp, while U2-Net variants, ISNet (including an anime version), BiRefNet checkpoints and Segment Anything with point prompts trade speed against edge quality; custom ONNX files and a paid cloud backend are options too. Hair and other soft edges can be cleaned with colour decontamination, alpha matting or a ViTMatte network. Install with pip using a cpu, gpu (CUDA) or rocm extra on Python 3.11 to 3.13. The code is MIT, but each model keeps its own license, and BRIA requires a commercial contract for its weights, so businesses should choose BiRefNet (MIT) or U2-Net (Apache-2.0). Handy for e-commerce photo batches, thumbnail pipelines and compositing prep.