Reproduce “VividFace: High-Quality and Efficient One-Step Diffusion For Video Face Enhancement” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shulian Zhang, Yong Guo, Long Peng, Ziyang Wang, Ye Chen, Wenbo Li, Xiao Zhang, Yulun Zhang and 1 more
VividFace proposes an efficient one-step diffusion framework for video face enhancement, delivering high-quality facial detail recovery and robust temporal consistency. It leverages a novel MLLM-curated dataset and face-focused training to achieve a 12x inference speedup compared to multi-step methods while demonstrating superior perceptual quality and identity preservation on various benchmarks.
Key insights
- One-step flow matching can efficiently adapt powerful pre-trained video generation models for high-fidelity video restoration tasks by predicting a direct linear flow trajectory.
- Optimizing for perceptually critical facial regions across both latent and pixel spaces, combined with stochastic face-focused supervision, effectively enhances detail recovery and temporal consistency.
- Multimodal Large Language Models (MLLMs) can systematically assess and filter high-quality data, offering a scalable method for curating superior task-specific datasets that improve model performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.