Research map / Video / Video generation
Causality in Video Diffusers is Separable from Denoising
Xingjian Bai, Guande He, Zhengqi Li, Eli Shechtman, Xun Huang, Zongze Wu
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 |
|---|---|---|---|
| decouple causal encoder from frame-wise diffusion decoder (SCD-M vs Causal DiT-M) | causal DiT · TECO-Minecraft | FVD / Sec per frame: FVD 38.7 -> 37.6; 2.4 -> 0.52 s/frame | Improved |
| decouple causal encoder from decoder (SCD-B vs Causal DiT-B) | causal DiT · UCF-101 | FVD / Sec per frame: FVD 187.6 -> 174.7; 3.9 -> 1.1 s/frame | Improved |
| increase causal encoder depth (ablation) | SCD-M · TECO-Minecraft | FVD: 37.6 -> 36.1 | Improved |
| increase causal encoder depth (ablation) | SCD-B · UCF-101 | FVD: 174.7 -> 171.1 | Improved |
| increase diffusion decoder depth (ablation) | SCD-M · TECO-Minecraft | FVD / Sec per frame: FVD 37.6 -> 34.9; 0.52 -> 1.6 s/frame | Mixed |
| increase diffusion decoder depth (ablation) | SCD-B · UCF-101 | FVD / Sec per frame: FVD 174.7 -> 158.7; 1.1 -> 2.8 s/frame | Mixed |
| replace Self Forcing with SCD fine-tuned from Wan2.1 1.3B | Self Forcing · VBench | throughput (FPS): 8.9 -> 11.1 FPS; total score 84.26 -> 84.03 | Mixed |
| skip 15 middle layers across later denoising steps then finetune (ablation) | AR Wan2.1 1.3B · qualitative videos | visual fidelity | No clear effect |
| switch last 5 layers to frame-diagonal attention with 5K-step finetune (ablation) | AR Wan2.1 1.3B · qualitative videos | visual fidelity | No clear effect |
| use frame-wise concatenation instead of channel-wise concatenation of context and noisy frame (ablation) | SCD decoder · SCD pretraining ablation | quality | Improved |
| inject Gaussian noise into causal context (ablation) | SCD · SCD pretraining ablation | quality | Improved |
Closest papers by meaning
- Video-Mirai: Autoregressive Video Diffusion Models Need Foresight (NeurIPS 2026 · similarity 0.90)
- TRIP: Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models (CVPR 2024 · similarity 0.89)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.88)
- Emergent Temporal Correspondences from Video Diffusion Transformers (NeurIPS 2025 · similarity 0.88)
- Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution (NeurIPS 2025 · similarity 0.88)
- Denoising Task Routing for Diffusion Models (ICLR 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.
