Faceswap: Developing a platform for practical Deepfakes
summary
Open source FaceSwap (GPL3) is a working end-to-end deepfake pipeline that runs on consumer GPUs, uses Python, Keras, TensorFlow, PlaidML, and supports multi‑GPU training on Windows Mac and Linux. It chains four prep networks—detection, alignment, recognition, segmentation—into a core autoencoder with shared encoder and two decoders to swap identities. Training is hands‑off but takes days to weeks; minimum 4 GB VRAM, recommended ~8 GB, or use Google Colab. Outputs can be auto‑blended or manually composited; several demos show high‑quality, practical results.
AI-written summary · about
chapters (22)
- 0:00Intro
- 0:23Overview
- 2:46How did we get here?
- 4:42Features of Faceswap
- 6:45API Stack
- 8:43Neural Networks
- 10:05The Faceswap Machine Learning Stack
- 10:48Face Detection
- 12:10Face Alignment
- 13:41Face Recognition
- 15:15Face Segmentation
- 17:07Image Reconstruction with Auto-Encoders
- 23:09Final Steps: Compositing & Splicing
- 25:29Getting Started with Faceswap
- 26:11Faceswap Process
- 27:27Gathering the Data
- 28:16Extract
- 30:25Convert
- 31:47Post-processing previews
- 33:46The future of Deeplakes
- 37:09References
- 47:27Conclusion and Q&A
related talks
58mPushing Deepfakes to the Limit-Fake video calls with AI
51mPlaying Offense and Defense with Deepfakes
56mDeep Fakes 2.0 - How Neural Networks are Changing our World
38mDIY Deep Fakes
32mHeart to Heart and Eye to Eye with Deep Fakes
30mReality Under Attack: Just How Serious Is the Deep Fake Threat?
27mIs now the time to solve the deepfake threat?
39mBuilding Highly Scalable Facial Recognition Pipelines