Beyond Distribution Matching

Self-Supervised Representation Forcing for Few-Step Video Generation

NeurIPS 2026

1Shanghai Jiao Tong University, 2Hunyuan, Tencent, 3Nanyang Technological University

*Work done during an internship at Hunyuan, Tencent. †Corresponding author.

Comparison with 5 Methods
Each column corresponds to one method, and each row corresponds to one example.
Wan
CausVid
SF
CausalForcing
Ours
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
Comparison with 3 Methods
Each column corresponds to one method, and each row corresponds to one example.
HunyuanVideo
DMD2
Ours
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕
☕