The proposed framework, termed stable video face restoration (SVFR), effectively combines video BFR, inpainting, and colorization tasks to enhance temporal coherence and restoration quality.
Our approach introduces a learnable task embedding to boost task identification, along with Unified Latent Regularization (ULR) for shared feature representation among different subtasks.
Facial prior learning and self-referred refinement are implemented as auxiliary strategies in our framework to enhance both training and inference restoration outcomes.
The novel SVFR framework leverages the generative and motion priors of Stable Video Diffusion (SVD), establishing a new paradigm for generalized video face restoration.
Collection
[
|
...
]