Optimization of Identity-Coherence via Direct Preference Learning
Replace online sampling reinforcement learning with direct preference optimization on offline video pairs to maximize identity coherence.
Source paper
Xu Guo, Fulong Ye, Xinghui Li, Pengqi Tu, Pengze Zhang, Qichao Sun, Songtao Zhao, Xiangwang Hou, Qian He
Generated by AI. The paper’s authors did not write these.
Replace online sampling reinforcement learning with direct preference optimization on offline video pairs to maximize identity coherence.
Train a layered diffusion model to predict foreground, background, and alpha mattes simultaneously, replacing heuristic video composition pipelines.
A method that prunes diffusion transformer tokens outside the face mask region during video generation to accelerate face-swapping inference without sacrificing visual fidelity.
The paper introduces DreamID-V, a video face-swapping framework that addresses the lack of temporal consistency and paired training data in existing methods. The authors bridge the gap between image and video domains by creating a synthetic data pipeline (SyncID-Pipe) that generates high-fidelity video-face pairs and employing a Diffusion Transformer (DiT) aimed at preserving identity and motion continuity.
The core methodology utilizes 'SyncID-Pipe,' a pipeline that pre-trains an Identity-Anchored Video Synthesizer (IVS) to generate video clips from start/end frames and pose sequences, creating 'Bidirectional ID Quadruplets' for explicit supervision. The model architecture is a Diffusion Transformer (DiT) with Modality-Aware Conditioning to separate identity, pose, and context, trained using a Synthetic-to-Real curriculum and fine-tuned with Identity-Coherence Reinforcement Learning (IRL) to minimize flickering.