Research map / Video / Video generation
AAD-1: Asymmetric Adversarial Distillation for One-Step Autoregressive Video Generation
Haobo Li, Yanhong Zeng, Yunhong Lu, Jiapeng Zhu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Yujun Shen, Zhipeng 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 warmup before adversarial refinement (ablation) | AAD-1 · VBench-I2V | Aesthetic Quality: 53.63 -> 58.64 | Improved |
| add DMD warmup before adversarial refinement (ablation) | AAD-1 · VBench-I2V | Imaging Quality: 62.81 -> 69.37 | Improved |
| use bidirectional instead of causal discriminator backbone (video-wise logit) (ablation) | AAD-1 discriminator · Drift Score (20s) | Drift Score: 7.10 -> 4.02 | Improved |
| use video-wise instead of frame-wise logits on causal backbone (ablation) | AAD-1 discriminator (causal) · VBench Dynamics | Dynamic Degree: 1.08 -> 42.07 | Improved |
| use video-wise instead of frame-wise logits on bidirectional backbone (ablation) | AAD-1 discriminator (bidirectional) · Drift Score (20s) | Drift Score: 4.38 -> 4.02 | Improved |
| use video-wise instead of frame-wise logits on bidirectional backbone (ablation) | AAD-1 discriminator (bidirectional) · VBench Dynamics | Dynamic Degree: 39.04 -> 39.29 | No clear effect |
| Stage-III adversarial training over Stage-II DMD model | AAD-1 · VBench-I2V | Subject Consistency: 92.14 -> 94.34 | Improved |
| Stage-III adversarial training over Stage-II DMD model | AAD-1 · VBench-I2V | Dynamic Degree: 50.30 -> 41.46 | Hurt |
| Stage-III adversarial training over Stage-II DMD model | AAD-1 · VBench-I2V | I2V Subject: 96.56 -> 98.65 | Improved |
| raise sampling budget from 1 NFE to 2 NFE | AAD-1 · VBench-I2V (480p, 5s) | Dynamic Degree: 41.46 -> 50.04 | Improved |
| extend generation from 5s to 20s | AAD-1 (1 NFE, 480p) · VBench-I2V | Subject Consistency: 94.34 -> 84.31 | Hurt |
| use bidirectional instead of causal discriminator in Stage III training (ablation) | AAD-1 · Training cost | training time: 65 hours -> 49 hours | Improved |
Closest papers by meaning
- Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation (NeurIPS 2025 · similarity 0.91)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.89)
- Autoregressive Video Generation without Vector Quantization (ICLR 2025 · similarity 0.88)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.88)
- Diffusion Adversarial Post-Training for One-Step Video Generation (ICML 2025 · similarity 0.87)
- Taming Teacher Forcing for Masked Autoregressive Video Generation (CVPR 2025 · similarity 0.87)
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.
