OmniCam from Tencent Hunyuan generates camera trajectories from panoramas and text
Zhenyang Liu · hf · 2026-10-21
OmniCam (Fudan University, Shanghai Innovation Institute, Tencent Hunyuan, Zhejiang University) is an autoregressive model that generates camera pose sequences from panoramas and text instructions.
Key points:
- Geometry-grounded pose token learning and target-aware conditioning improve trajectory accuracy and reduce collisions
- Accompanied by the OmniCaT dataset with 267.7K trajectories covering orbit, reconstruction, and wandering tasks
- Downstream applications: camera-controlled video generation, camera-guided 3D reconstruction, and camera-guided active perception in robotics
More from Multimodal
- Same-seed image edit model comparison: Flux 3 Klein alone nails pose and jacket side — Dunc4n1d4h0 · 2026-10-21
- Fully AI-made short film 'The Life of the Gorgon': Grok writes, Kling 4.0 animates — nima_owji · 2026-10-21
- Longtime Suno user roasts its vocal shimmer artifact: 'makes me puke' — lfschiavo · 2026-10-21
- Reddit user shares AI-made Death Note short film for TheLoopHole Project — neptune765 · 2026-10-21
- Hands-on with Fibo 2.0 image editing: solid 360° turns, but object occlusion fails — Big_Parsnip_9053 · 2026-10-21
- Minimax H3 Character Replacement: Ref2va 8-Step LoRA Transfers Facial Emotion Acting — mexxmillion · 2026-10-21