Research map / Video / Video generation
Dual-Expert Consistency Model for Efficient and High-Quality Video Generation
Zhengyao Lv, Chenyang Si, Tianlin Pan, Zhaoxi Chen, Kwan-Yee K. Wong, Yu Qiao, Ziwei Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add optimization decoupling into semantic and detail experts (ablation) | consistency distillation of HunyuanVideo · VBench ablation (HunyuanVideo, 4 steps) | VBench total: 80.30 -> 83.08 | Improved |
| make dual expert parameter-efficient (time embeddings + LoRA) instead of two full experts (ablation) | DCM · VBench ablation (HunyuanVideo, 4 steps) | VBench total: 83.08 -> 83.03 | No clear effect |
| add Temporal Coherence loss to semantic expert (ablation) | DCM (OD+PE) · VBench ablation (HunyuanVideo, 4 steps) | VBench total: 83.03 -> 83.42 | Improved |
| add Temporal Coherence loss when GAN loss already used (ablation) | DCM (OD+PE+GF) · VBench ablation (HunyuanVideo, 4 steps) | VBench total: 83.71 -> 83.80 | Improved |
| add GAN and Feature Matching loss to detail expert (ablation) | DCM (OD+PE) · VBench ablation (HunyuanVideo, 4 steps) | VBench total: 83.03 -> 83.71 | Improved |
| add GAN and Feature Matching loss when TC already used (ablation) | DCM (OD+PE+TC) · VBench ablation (HunyuanVideo, 4 steps) | VBench total: 83.42 -> 83.80 | Improved |
| add full DCM (all components) over vanilla consistency distillation (ablation) | HunyuanVideo 4 steps · VBench ablation (HunyuanVideo, 4 steps) | VBench semantic: 78.36 -> 78.62 | Improved |
| move switch step kappa away from 37 (28, 35, 39, 46) (ablation) | DCM · HunyuanVideo VBench | video quality | Hurt |
| use DCM instead of LCM at 4 steps | HunyuanVideo · VBench (4 steps) | VBench total: 80.33 -> 83.83 | Improved |
| use DCM instead of PCM at 4 steps | HunyuanVideo · VBench (4 steps) | VBench total: 80.93 -> 83.83 | Improved |
| use DCM instead of PCM at 8 steps | CogVideoX-2B · VBench (8 steps) | VBench total: 79.70 -> 80.26 | Improved |
| distill 50-step HunyuanVideo to 4 steps with DCM | HunyuanVideo · VBench | VBench total / latency: 83.87 -> 83.83; latency 1504.5 -> 121.52 | No clear effect |
Closest papers by meaning
- Motion Consistency Model: Accelerating Video Diffusion with Disentangled Motion-Appearance Distillation (NeurIPS 2024 · similarity 0.91)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.90)
- Towards Precise Scaling Laws for Video Diffusion Transformers (CVPR 2025 · similarity 0.89)
- Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE (NeurIPS 2026 · similarity 0.88)
- One-Step Diffusion for Detail-Rich and Temporally Consistent Video Super-Resolution (NeurIPS 2025 · similarity 0.88)
- Phased Consistency Models (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
