Swap faces in video with a model you can actually fine-tune

What it does GHOST (Generative High-fidelity One Shot Transfer) swaps a single source face onto target images or videos without per-subject retraining. The repository provides inference scripts, pre-trained generator and discriminator weights, and a complete training pipeline built on PyTorch. It handles both one-to-one swaps and multi-face replacements in a single video clip.

The interesting bit Most face-swap projects release a demo and call it a day; this one publishes an IEEE Access paper and still gives you tunable U-Net backbones, configurable AAD ResBlk blocks, and loss-coefficient knobs for fine-tuning. The README also devotes serious real estate to an ethics policy, explicitly banning non-consensual and malicious use—a rarity in a field where misuse dominates the headlines.

Key highlights

• One-shot transfer for both image-to-image and image-to-video domains.
• Supports swapping multiple distinct faces in a single target video.
• Full training code included, with pre-trained weights and architecture options (U-Net backbones, 1–3 AAD ResBlk blocks).
• Local and Google Colab inference notebooks provided.
• Explicit zero-tolerance ethics stance against illicit or non-consensual use.

Caveats

• Installation instructions still reference the old sber-swap repository name, suggesting the documentation has not fully kept pace with the rename.
• The README offers no quantitative benchmarks or speed metrics, so comparing quality against other tools is left to eyeballing the example images.
• Training from scratch requires careful loss scheduling; the authors advise disabling eye-detection loss and schedulers in early epochs.

Verdict Computer-vision researchers and VFX tinkerers who want a reproducible, trainable face-swap baseline will find this valuable. If you are looking for a polished consumer app or guaranteed state-of-the-art results without reading a paper, look elsewhere.