Research map / Video / Video generation
DualCamCtrl: Dual-Branch Diffusion Model for Geometry-Aware Camera-Controlled Video Generation
Hongfei Zhang, Kanghao Chen, Zixin Zhang, Harold Haodong Chen, Yuanhuiyi Lyu, Kun Zhou, Yuqi Zhang, Shuai Yang, Yingcong Chen
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 depth branch guidance (ablation) | DualCamCtrl · RealEstate10K (I2V, ablation) | FVD / FID: 96.3 -> 80.4 FVD; 66.1 -> 49.9 FID | Improved |
| use two-stage training (decoupled then fusion) instead of single-stage (ablation) | DualCamCtrl · RealEstate10K (I2V, ablation) | FVD / FID: 96.7 -> 80.4 FVD; 58.3 -> 49.9 FID | Improved |
| use dual-branch instead of Naive-Prop depth propagation (ablation) | DualCamCtrl · RealEstate10K (I2V, ablation) | FVD: 98.6 -> 80.4 | Improved |
| use dual-branch instead of Plucker+D depth embedding (ablation) | DualCamCtrl · RealEstate10K (I2V, ablation) | FVD: 100.8 -> 80.4 | Improved |
| use dual-branch instead of single-branch RGB-depth (ablation) | DualCamCtrl · RealEstate10K (I2V, ablation) | FVD: 125.1 -> 80.4 | Improved |
| use 3D fusion with frame gating instead of ControlNet-style 1D fusion (ablation) | DualCamCtrl · RealEstate10K (I2V, ablation) | FVD: 112.3 -> 80.4 | Improved |
| use 3D spatiotemporal fusion instead of 2D spatial fusion (ablation) | DualCamCtrl · RealEstate10K (I2V, ablation) | FVD: 98.4 -> 85.4 | Improved |
| add frame-level gating to 3D fusion (ablation) | DualCamCtrl · RealEstate10K (I2V, ablation) | FVD / FID: 85.4 -> 80.4 FVD; 52.7 -> 49.9 FID | Improved |
| use SIGMA mutual alignment (RGB 1-5, depth 6-15) instead of one-way depth-only alignment (1-5) (ablation) | DualCamCtrl · RealEstate10K (I2V, ablation) | FVD / rotation error: 91.4 -> 80.4 FVD; 1.76 -> 1.25 rotation error | Improved |
| use SIGMA instead of geometry-guided alignment (RGB 6-15, depth 1-5) (ablation) | DualCamCtrl · RealEstate10K (I2V, ablation) | FVD / rotation error: 84.6 -> 80.4 FVD; 1.55 -> 1.25 rotation error | Improved |
| widen depth injection range from layers 1-5 to 1-15 in one-way alignment (ablation) | DualCamCtrl · RealEstate10K (I2V, ablation) | FVD: 91.4 -> 83.2 | Improved |
| allocate more early denoising steps (ablation) | DualCamCtrl · RealEstate10K | FVD / rotation error | Improved |
Closest papers by meaning
- Pantheon360: Taming Digital Twin Generation via 3D-Aware 360deg Video Diffusion (CVPR 2026 · similarity 0.89)
- RealCam-I2V: Real-World Image-to-Video Generation with Interactive Complex Camera Control (ICCV 2025 · similarity 0.88)
- UniScene-MoTion: Unified Scene & Motion-aware Diffusion Transition Framework (AAAI 2026 · similarity 0.88)
- Taming Video Models for 3D and 4D Generation via Zero-Shot Camera Control (CVPR 2026 · similarity 0.88)
- IllumiCraft: Unified Geometry and Illumination Diffusion for Controllable Video Generation (NeurIPS 2025 · similarity 0.87)
- Training-free Camera Control for Video Generation (ICLR 2025 · similarity 0.86)
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.
