DEMOS & QUALITATIVE COMPARISON
Audio-driven video results
Qualitative results cover multilingual self-driven reconstruction, redrawing of MiniMax H3 generated videos, and cross-audio driving under large poses, occlusion, rapid head motion, and unconstrained real-world conditions. Video is more informative than isolated frames for evaluating synchronization, oral detail, and temporal stability.
METHOD
Method
TBDub builds on the mask-free, reference-conditioned video editor of X-Dub/Wan. The complete source video and motion context constrain identity and spatiotemporal structure, while frame-aligned speech controls target articulation. Production-domain post-training yields a 30-step Teacher, which task-aware DMD2 compresses into a two-step Student.
The generative backbone is inherited from X-Dub/Wan: the complete reference video and motion context constrain identity and spatiotemporal structure, while frame-aligned audio controls speech-related target dynamics.
Production adaptation and two-step distillation
TBDub uses two consecutive stages: task-adaptive post-training produces a 30-step Teacher, and conditional DMD2 compresses it into a two-step Student.
QUANTITATIVE RESULTS
Evaluation
We compare the 30-step Teacher, two-step EMA Student, and representative methods through objective metrics and subjective MOS on 38 randomly sampled TalkVid clips.
INFERENCE EFFICIENCY
Efficiency
The two-step Student sharply reduces DiT computation while largely retaining the Teacher's visual quality, identity consistency, and audiovisual synchronization.
CITATION
BibTeX
Citation information for the TBDub technical report.