Research map / Video / Video generation
Reward Lightning: Fast Video Generation via Homologous Preference Distillation
Jiaxiang Cheng, bing ma, Xuhua Ren, Kai Yu, Peng Zhang, Tianxiang Zheng, Qinglin Lu
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 few-step augmentation to LRM training data (ablation) | LRM · VideoGen-RewardBench (LRM ablation) | accuracy: 68.47 -> 72.24 | Improved |
| add real-synthetic pairs to LRM training data (ablation) | LRM · VideoGen-RewardBench (LRM ablation) | accuracy: 65.12 -> 72.24 | Improved |
| use time-step projection head instead of self-attention head (ablation) | LRM · VideoGen-RewardBench (LRM ablation) | accuracy: 64.38 -> 72.24 | Improved |
| use time-step projection head instead of trainable query head (ablation) | LRM · VideoGen-RewardBench (LRM ablation) | accuracy: 69.85 -> 72.24 | Improved |
| add reward regularization (rho=0.95) (ablation) | LRM · VideoGen-RewardBench (LRM ablation) | accuracy: 62.15 -> 72.24 | Improved |
| initialize discriminator with LRM instead of random (ablation) | HPD · VBench-I2V (4-NFE) | I2V Score: 88.26 -> 95.34 | Improved |
| use adaptive preference weight instead of lambda=0 (ablation) | HPD · VBench-I2V (4-NFE) | I2V Score: 90.87 -> 95.34 | Improved |
| share backbone between reward and distillation (homology) (ablation) | HPD · Gradient conflict (4-NFE I2V) | cosine similarity: -0.14 -> 0.41 | Improved |
| add LRM to rCM | rCM · VBench-I2V (4-NFE) | I2V Score: 93.45 -> 94.32 | Improved |
| add LRM to DMD2 | DMD2 · VBench-I2V (4-NFE) | I2V Score: 91.92 -> 92.71 | Improved |
| add LRM to HPD (ablation) | HPD · VBench-I2V (4-NFE) | FVD: 167.14 -> 162.87 | Improved |
| use BTT loss instead of BT loss (ablation) | LRM · VideoGen-RewardBench | accuracy w/o Ties: 71.89 -> 72.24 | Improved |
Closest papers by meaning
- Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search (NeurIPS 2025 · similarity 0.89)
- Improving Video Generation with Human Feedback (NeurIPS 2025 · similarity 0.88)
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · similarity 0.88)
- VideoDPO: Omni-Preference Alignment for Video Diffusion Generation (CVPR 2025 · similarity 0.88)
- Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization (NeurIPS 2025 · similarity 0.88)
- InstructVideo: Instructing Video Diffusion Models with Human Feedback (CVPR 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.
