Research map / Image generation / Fast diffusion sampling
Large Scale Diffusion Distillation via Score-Regularized Continuous-Time Consistency
Kaiwen Zheng, Yuji Wang, Qianli Ma, Huayu Chen, Jintao Zhang, Yogesh Balaji, Jianfei Chen, Ming-Yu Liu, Jun Zhu, Qinsheng Zhang
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 DMD score-distillation regularizer to sCM (rCM vs pure sCM) (ablation) | sCM · T2I/T2V distillation | visual quality | Improved |
| rCM versus DMD2 on image quality | rCM (Cosmos-Predict2 2B, 4 steps) · GenEval (Cosmos-Predict2 T2I) | GenEval overall: 0.80 -> 0.81 | No clear effect |
| rCM versus DMD2 on video quality | rCM (Wan2.1 1.3B, 4 steps) · VBench total score (Wan2.1 T2V, 832x480x81) | VBench total score: 84.56 -> 84.43 | No clear effect |
| rCM versus DMD2 on diversity | rCM (Wan2.1) · Video diversity | sample diversity | Improved |
| scale rCM student from 0.6B to 14B | rCM · GenEval (Cosmos-Predict2 T2I) | GenEval overall: 0.79 -> 0.83 | Improved |
| reduce sampling steps from 4 to 1 | rCM (Cosmos-Predict2 14B) · GenEval (Cosmos-Predict2 T2I) | GenEval overall: 0.83 -> 0.82 | Hurt |
| reduce sampling steps from 4 to 1 | rCM (Wan2.1 14B) · VBench total score (Wan2.1 T2V, 832x480x81) | VBench total score: 84.92 -> 83.02 | Hurt |
| reduce sampling steps from 4 to 2 | rCM (Wan2.1 14B) · VBench total score (Wan2.1 T2V, 832x480x81) | VBench total score: 84.92 -> 85.05 | No clear effect |
| distill with rCM versus teacher on video | rCM (Wan2.1 14B) · VBench total score (Wan2.1 T2V, 832x480x81) | VBench total score: 83.58 -> 84.92 | Improved |
| increase DMD weight lambda (ablation) | rCM (Wan2.1 1.3B) · Video quality vs diversity | quality and diversity: larger lambda: better quality, less diversity | Mixed |
| use continuous-time consistency trajectory model (sCTM/MeanFlow) instead of sCM (ablation) | sCM · T2I distillation | quality and diversity | Hurt |
| use SiD instead of DMD as score regularizer (ablation) | rCM · T2I/T2V | quality | No clear effect |
Closest papers by meaning
- SCott: Accelerating Diffusion Models with Stochastic Consistency Distillation (AAAI 2025 · similarity 0.90)
- Continuous-Time Distribution Matching for Few-Step Diffusion Distillation (NeurIPS 2026 · similarity 0.89)
- Simplifying, Stabilizing and Scaling Continuous-time Consistency Models (ICLR 2025 · similarity 0.89)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.89)
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.89)
- Diffusion Model is Effectively Its Own Teacher (CVPR 2025 · similarity 0.89)
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.
