Research map / Video / Video generation
Identity-Preserving Image-to-Video Generation via Reward-Guided Optimization
Liao Shen, Wentao Jiang, Yiran Zhu, Jiahe Li, Tiezheng Ge, Zhiguo Cao, Bo Zheng
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 IPRO facial reward fine-tuning | Wan2.2 A14B · Small-face eval (600 scenes) | FaceSim: 0.5780 -> 0.6942 | Improved |
| Add IPRO facial reward fine-tuning | In-house I2V 15B · Small-face eval (600 scenes) | FaceSim: 0.4769 -> 0.6960 | Improved |
| Add IPRO facial reward fine-tuning | Wan2.2 5B · Small-face eval (600 scenes) | FaceSim: 0.3788 -> 0.5460 | Improved |
| Add IPRO facial reward fine-tuning | Wan2.2 A14B · VBench-I2V | Dynamic Degree: 19.45 -> 19.17 | Mixed |
| Use facial reward instead of DPO | Wan2.2 A14B · Small-face eval (600 scenes) | FaceSim: 0.6284 -> 0.6942 | Improved |
| Use facial reward instead of GRPO | Wan2.2 A14B · Small-face eval (600 scenes) | FaceSim: 0.6334 -> 0.6942 | Improved |
| Use reward feedback instead of SFT with auxiliary face loss (ablation) | Wan2.2 A14B · Small-face eval (600 scenes) | FaceSim: 0.6392 -> 0.6942 | Improved |
| Use ArcFace instead of CLIP as reward encoder (ablation) | Wan2.2 A14B · Small-face eval (600 scenes) | FaceSim: 0.6099 -> 0.6942 | Improved |
| Backprop through final low-noise steps instead of high-noise steps (ablation) | Wan2.2 A14B · Small-face eval (600 scenes) | FaceSim: 0.6456 -> 0.6942 | Improved |
| Add KL regularization (ablation) | IPRO · Small-face eval (600 scenes) | Reward hacking rate: 58% -> 10% | Improved |
| Add facial scoring mechanism (FSM) (ablation) | IPRO · Small-face eval (600 scenes) | Reward hacking rate: 52% -> 10% | Improved |
| Add KL regularization and FSM (ablation) | IPRO · Small-face eval (600 scenes) | FaceSim: 0.7544 -> 0.6942 (KL), 0.7388 -> 0.6942 (FSM) | Hurt |
Closest papers by meaning
- Identity-Preserving Text-to-Video Generation by Frequency Decomposition (CVPR 2025 · similarity 0.90)
- StableAnimator: High-Quality Identity-Preserving Human Image Animation (CVPR 2025 · similarity 0.89)
- FaithfulFaces: Pose-Faithful Facial Identity Preservation for Text-to-Video Generation (NeurIPS 2026 · similarity 0.88)
- InstructVideo: Instructing Video Diffusion Models with Human Feedback (CVPR 2024 · similarity 0.87)
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · similarity 0.87)
- ARGUS: Stacked Multi-View Identity Mosaic Injection for Subject-Preserving Video Generation (NeurIPS 2026 · 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.
