CameraVLM: 9B model cracks camera-motion understanding where most VLMs flip coins
tencent · hf · 2026-10-16
A HKUST-GZ, Tencent and LIX paper shows most VLMs perform no better than a coin flip at telling whether the camera moved left or right. CameraVLM pairs motion-state tokens (off-screen camera movement) with flow tokens (on-screen dynamics) under state–language alignment training. At only 9B parameters it beats all open-source and camera-motion-specific models across 3 benchmarks and 20+ baselines, matching or surpassing GPT-6 Astra and humans in a 48-participant study. Project page: 3dagentworld.github.io/CameraVLM
More from Multimodal
- SeedAudio 1.5 hits CapCut PC: one prompt generates dialogue, SFX and music — LearnWithBishal · 2026-10-16
- DLSS 5 Assetto Corsa demo stuns sim racers: 'better than reality' — CSProfKGD · 2026-10-16
- Video recast with fal agent: swap a character via one character sheet — adamho · 2026-10-16
- Dev open-sources Krea 2 pixel-art LoRA for 16x16 game icons with transparent backgrounds — Angaisb_ · 2026-10-16
- Opus 5.5 + Midjourney: 'models have taste, you just point them right' — yuwen_lu_ · 2026-10-16
- Gemini 4 Argon hits AssemblyAI's LLM Gateway, tops voice bench at 99.5 across 50+ models — AssemblyAI · 2026-10-16