Research map / Video / Video generation
Taming Camera-Controlled Video Generation with Verifiable Geometry Reward
Zhaoqing Wang, Xiaobo Xia, Zhuolin Bie, Jinlin Liu, Dongdong Yu, Jiawang Bian, Changhu Wang
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 online RL post-training with geometry reward (T2V) | CamVerse SFT · RealEstate10K test (T2V) | translation error: 0.0395 -> 0.0293 | Improved |
| add online RL post-training with geometry reward (T2V) | CamVerse SFT · RealEstate10K test (T2V) | rotation error: 0.6506 -> 0.5140 | Improved |
| add online RL post-training with geometry reward (I2V) | CamVerse SFT · RealEstate10K test (I2V) | rotation error: 0.5613 -> 0.4685 | Improved |
| add online RL post-training with geometry reward (I2V) | CamVerse SFT · RealEstate10K test (I2V) | visual quality: 5.02 -> 4.87 | Hurt |
| add dynamic videos to SFT data (ablation) | CamVerse SFT · RealEstate10K 100-clip subset, image-to-video | rotation error: 0.7306 -> 0.6506 | Improved |
| add shifted timestep scheduler (ablation) | CamVerse SFT · RealEstate10K 100-clip subset, image-to-video | rotation error: 0.6055 -> 0.5613 | Improved |
| add pose normalization (ablation) | CamVerse SFT · RealEstate10K 100-clip subset, image-to-video | rotation error: 0.7135 -> 0.5613 | Improved |
| use relative error for both translation and rotation (ablation) | CamVerse GRPO · RealEstate10K 100-clip subset, image-to-video | rotation error: 0.4537 -> 0.3286 | Improved |
| use dense segment-level reward instead of clip-level (ablation) | CamVerse GRPO · RealEstate10K 100-clip subset, image-to-video | rotation error: 0.3739 -> 0.3286 | Improved |
| use 3 SDE steps instead of 1 (ablation) | CamVerse GRPO · RealEstate10K 100-clip subset, image-to-video | rotation error: 0.4829 -> 0.3286 | Improved |
| use 3 SDE steps instead of 5 (ablation) | CamVerse GRPO · RealEstate10K 100-clip subset, image-to-video | translation error: 0.0278 -> 0.0238 | Improved |
| use balanced 4-4 samples instead of 1-1 (ablation) | CamVerse GRPO · RealEstate10K 100-clip subset, image-to-video | rotation error: 0.4330 -> 0.3286 | Improved |
Closest papers by meaning
- AC3D: Analyzing and Improving 3D Camera Control in Video Diffusion Transformers (CVPR 2025 · similarity 0.89)
- RLGF: Reinforcement Learning with Geometric Feedback for Autonomous Driving Video Generation (NeurIPS 2025 · similarity 0.89)
- Training-free Camera Control for Video Generation (ICLR 2025 · similarity 0.89)
- VD3D: Taming Large Video Diffusion Transformers for 3D Camera Control (ICLR 2025 · similarity 0.88)
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · similarity 0.87)
- Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance (NeurIPS 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.
