Introducing SCoPE, Tencent ARC’s camera-controlled image-to-video model

Prompt

Introducing SCoPE, Tencent ARC’s camera-controlled image-to-video model built on Wan2.2-I2V-A14B. 🎥 Apache 2.0.
🤖https://t.co/yGlJY7Ur4n📄 https://t.co/TGFU6zAczq

🏆 29% lower rotation error and 43% lower FVD at 14B than the strongest baseline, while leading every closed-loop revisit metric.
⚡ Under 0.1% new parameters, with gains at both 5B and 14B while preserving the pretrained RoPE bit-exactly.
🎬 Camera rays become positional coordinates inside attention, turning a first frame, prompt, and camera trajectory into precisely controlled, spatially consistent video.
📦 Self-contained inference weights—no separate Wan2.2 checkpoint required.