Research map / Video / Video generation
SALT: Self-Consistent Distribution Matching with Cache-Aware Training for Few-Step Video Generation
Xingtong Ge, Yi ZHANG, Yushi Huang, Dailan He, Xiahong Wang, Bingqi Ma, Guanglu Song, Yu Liu, Jun 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 self-consistency loss on top of denser 8-point grid DMD (ablation) | DMD-8 · Wan2.1 T2V 1.3B VBench (4 NFE) | Total Score: 82.54 -> 83.36 | Improved |
| train DMD on denser 8-point grid instead of 4-point grid (ablation) | DMD-4 · Wan2.1 T2V 1.3B VBench (4 NFE) | Total Score: 82.78 -> 82.54 | Hurt |
| add mixed-step rollout when using SC loss (ablation) | Causal Forcing + L_SC · VBench AR (Causal Forcing, 4 NFE) | Total Score: 83.83 -> 85.02 | Improved |
| add naive SC loss to official baseline without mixed-step (ablation) | Causal Forcing · VBench AR (Causal Forcing, 4 NFE) | Total Score: 84.62 -> 83.83 | Hurt |
| add cache-conditioned reference alignment (ablation) | Causal Forcing + mixed-step + L_SC · VBench AR (Causal Forcing, 4 NFE) | Total Score: 85.02 -> 85.08 | Improved |
| add cache-conditioned reference alignment (ablation) | Causal Forcing + mixed-step + L_SC · VBench AR (Causal Forcing, 2 NFE) | Semantic Score: 80.65 -> 81.49 | Improved |
| add SC-DMD regularization to DMD (ablation) | DMD · VBench-I2V Wan2.1 14B (4 NFE) | Temporal Flickering: 95.21 -> 97.41 | Improved |
| add SC-DMD regularization to DMD | DMD · VBench-I2V Wan2.1 14B (4 NFE) | Dynamic Degree: 58.46 -> 52.85 | Hurt |
| add Salt training to Causal Forcing | Causal Forcing · VBench-Long 30-second | Dynamic Degree: 76.95 -> 54.02 | Hurt |
| add Salt training to Causal Forcing | Causal Forcing · VBench 5-second (4 NFE) | Total Score: 84.62 -> 85.08 | Improved |
| add Salt training to LongLive | LongLive · VBench 5-second (4 NFE) | Total Score: 84.40 -> 84.93 | Improved |
| hybrid DMD with shortcut flow matching loss on mismatched grid (ablation) | DMD · Wan2.1 T2V 1.3B VBench (4 NFE) | Total Score: 82.78 -> 75.06 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Causal Forcing — Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation (ICML 2026) · better on 2 of 2 shared results
- LongLive — LongLive: Real-time Interactive Long Video Generation (ICLR 2026) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MagCache: Fast Video Generation with Magnitude-Aware Cache (NeurIPS 2025 · similarity 0.87)
- CoDMD: Copula-aware Distribution Matching Distillation for Fast Video Generation (NeurIPS 2026 · similarity 0.87)
- Continuous-Time Distribution Matching for Few-Step Diffusion Distillation (NeurIPS 2026 · similarity 0.87)
- Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion (NeurIPS 2025 · similarity 0.86)
- Pyramidal Flow Matching for Efficient Video Generative Modeling (ICLR 2025 · similarity 0.86)
- AdaState: Self-Evolving Anchors for Streaming Video Generation (NeurIPS 2026 · 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.
