Description of the technical principle
CanonSwap's facial animation function is essentially a reverse application of motion information, which needs to break through the problem of static information complementation.
Step-by-step program
- Three-dimensional facial reconstruction: Predicting Depth Information from Photographs to Construct Deformable Meshes Using 3DDFA_V2
- Motion parameter migration: Mapping the FLAME model parameters of the drive video to the source face topology
- Detail Enhancement: Generate pore-level dynamic details via GAN networks (requires -detail_enhance mode to be enabled)
Key Tips
Necessary steps: 1) Prepare a photo of the front face with neutral expression 2) Start with -animation_mode 3) Apply Attention-Enhanced module to the eyes separately. Note that large expressions such as smiles need to ensure that the mouth closure of the photo is ≥70%, otherwise 68 feature points need to be labeled manually.
This answer comes from the articleCanonSwap: A tool for realizing high-fidelity face-swapping in videoThe
May not be reproduced without permission:AI productivity tools " How to realize the conversion of still photos to animated videos with emoticons?
English 
简体中文 
日本語 
Deutsch 
Português do Brasil