hacksider/Deep-Live-Cam

Deep-Live-Cam is a desktop face-swap tool for images, video files and live camera feeds. You pick one source face photo, point it at a target image, clip or webcam, and it renders the swap, aiming at real-time output for the live mode. It suits creators doing character animation, memes, live performance or movie-style previews — and the same capability carries obvious impersonation and fraud risk.

How it works

The program runs a GUI launched with run.py. In image/video mode you select a source face and a target, press Start, and output is written to a folder named after the target. In webcam mode you select a source face, press Live, wait roughly 10–30 seconds for the preview, then capture the window with a tool such as OBS to stream; changing faces means selecting a new source image.

Distinguishing controls matter in practice: Mouth Mask keeps the original mouth so lip movement stays accurate, and Face Mapping / many-faces lets different faces be assigned to multiple people in frame at once. Project README

Setup paths

Prebuilt builds come from the project's official site, offered per platform (Windows, Mac Silicon, CPU, NVIDIA, AMD) as a paid "Ultimate" edition with extra features and priority support, advertised as zero manual installation. The README warns that only that site and the GitHub repo are official sources. Quickstart

Manual installation is explicitly described as requiring technical skills and not for beginners. It needs Python (3.11–3.14 supported, 3.14 recommended), pip, git, ffmpeg and the two ONNX model files placed in a models folder; on Windows, Visual Studio 2022 runtimes. macOS additionally needs the tkinter package for the GUI, and Apple Silicon uses the CoreML execution provider from the standard onnxruntime wheel — uninstall the older onnxruntime-silicon fork first. GPU acceleration is optional and provider-specific: CUDA for NVIDIA, CoreML for Apple, DirectML for Windows, OpenVINO for Intel, with pinned version pairings. Without a GPU it falls back to CPU, which the README notes is slower.

Constraints and licensing

The README documents a built-in check that blocks inappropriate media such as nudity, graphic content or sensitive material; it also states the maintainers may shut down the project or add watermarks if legally required. Users are asked to obtain consent for a real person's face and to label output as a deepfake. README

The code is AGPL-3.0, but the bundled insightface model is credited as non-commercial research use only, so commercial deployment needs separate licensing attention. The README also lists the command-line interface as unmaintained; the documented workflow is the GUI.

HotRepo's take

A single daily-board observation on 2026-10-31 showed 97,461 stars, +44 that day — too little history to read a trend. Discussion exists but dates mostly to 2024 and concerns earlier versions; a 2024 HN comment questioned the ethics of the demo footage itself HN thread.