Research map / Video / Video generation
CoDMD: Copula-aware Distribution Matching Distillation for Fast Video Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CoDMD, a copula-aware regularizer added to distribution matching distillation that matches teacher and student pairwise cosine-similarity matrices across batch and frames, distilling Wan-2.1-T2V 50-step teachers into 4-step students.
Problem
Few-step DMD for video diffusion degrades under small NFE budgets, causing layout instability, oversaturation, and broken motion, since its coordinate-wise, reverse-KL objective leaves cross-sample and cross-frame dependence unconstrained.
Approach
Reuse the teacher and fake score predictions already computed in each DMD step to build pairwise cosine-similarity matrices across batch elements and across frames, and add a KL loss that aligns student relation matrices with a target shifted by the teacher-fake relational difference, with no extra networks, data, or trajectories.
Key results
- On Wan2.1-T2V-1.3B with 4 steps, CoDMD reaches VBench total 84.46 vs DMD 83.38 and rCM 82.81.
- On Wan2.1-T2V-14B with 4 steps, CoDMD reaches VBench total 84.87 vs DMD 83.81 and rCM 84.05.
- In a 25-person blind user study, CoDMD is preferred over DMD in 81.8% of comparisons and over rCM in 74.4%.
Surprising findings
- Direct relational matching between student and teacher similarity matrices (without the score-difference term) gives only 83.8 total, barely above DMD's 83.4.
- Adding the frame-level loss alone raises dynamic degree to 93.3 but lowers semantic score by 2.0 relative to DMD.
Limitations
- Sensitivity to relational weight: performance drops at lambda=1.0.
- The first-order gradient derivation holds only when the linearization is moderate, which is verified only empirically.
- Evaluation relies mainly on VBench and a relatively small user study; sCM and AVD were reproduced by the authors rather than using official weights.
Experimental setup
Datasets: mixed corpus of public video text descriptions
Benchmarks: VBench
Metrics: VBench total score quality score semantic score dynamic degree VBench per-attribute scores user win rate
Baselines: DMD rCM sCM DCM AVD
Builds on: Distribution Matching Distillation DMD2 Relational Knowledge Distillation Sklar's theorem
Closest papers by meaning
- Continuous-Time Distribution Matching for Few-Step Diffusion Distillation (NeurIPS 2026 · similarity 0.89)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.88)
- Motion Consistency Model: Accelerating Video Diffusion with Disentangled Motion-Appearance Distillation (NeurIPS 2024 · similarity 0.87)
- Emergent Temporal Correspondences from Video Diffusion Transformers (NeurIPS 2025 · similarity 0.87)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.86)
- Improved Distribution Matching Distillation for Fast Image Synthesis (NeurIPS 2024 · similarity 0.86)
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.