DeepFaceLab Tutorial Series Part 2 – Common Operations and Conceptual Understanding.

1. ​Workspace Directory Structure​

data_dst (Destination Materials)

• Stores target video frames and extracted faces.

data_src (Source Materials)

• Stores source video frames and extracted faces.

model

• Contains trained model files (the core of face-swapping).

data_dst.mp4 (Target Video)

• The original video to receive the face swap.

data_src.mp4 (Source Video)

• The video providing the face to be swapped.

2. What is Face Swapping?

1. Source Video (data_src.mp4)

   • Provides the face to be transferred.

   • Extracted faces are saved in data_src/aligned.

2. Target Video (data_dst.mp4)

   • Provides the body, background, and motion.

   • Extracted frames are saved in data_dst.

3. Output Result

   • The final video replaces the target’s face with the source’s face: Target’s Body + Source’s Face.

 

3. What is a Model?

DeepFaceLab’s workflow boils down to two phases: training a model and applying the model. The model folder houses the AI “brain” that learns to map and blend faces.

3.1 Model Analogy: Stable Diffusion vs. DeepFaceLab

Stable Diffusion Models	DeepFaceLab Models
Generate images in specific styles (e.g., anime, realism).	Specialize in swapping specific faces.
Trained on broad datasets.	Trained on paired source/target faces.

3.2 Training as “Digital Alchemy”

In the DeepFaceLab community, training models is humorously called “Alchemy” . 

• Ingredients: High-quality face datasets.

• Furnace: GPU hardware (NVIDIA RTX recommended).

• Technique: Parameter tuning (like ancient alchemical recipes).

• Time: Days to weeks of training (the “49 days” of digital cultivation).

Note: Beginners rarely achieve perfection on their first try—experimentation is essential.

 

4. Additional Guidelines

1. Workspace Integrity

   • Never rename files/folders in the workspace.

2. Material Recommendations

   • Video > Images: Use video clips for both source and target whenever possible.

   • Example: To swap faces with a pop star, source a video of her or place high-quality images of her in data_src for extraction.

3. Image-Based Workflow

   • If using images instead of video:

     • Place source images directly in data_src.

     • Run the extraction script—all other steps mirror video-based workflows.