Yu-Cheng Chou

Importantly, we open-source the entire curation pipeline: • trajectory reconstruction (SfM / hloc) • geometric verification • motion consistency checks • automatic repair & filtering → You can build your own OpenSafari.

We also built OpenSafari, a dataset designed to break existing models: • in-the-wild FPV drone videos • large-scale camera motion • strong parallax & elevation changes Every trajectory is geometrically verified.

The difference is most visible under aggressive motion: • sharp turns • large parallax • long trajectories Baselines collapse, ours stays consistent.

Our idea: build a world memory. At each frame, the model retrieves 3D-consistent information conditioned on the camera pose. → This enables stable generation under 6-DoF motion → Even for long trajectories in complex outdoor scenes

But in reality, most video generation models today: ❌ Only work on narrow domains (e.g., real estate scenes) ❌ Break as soon as the camera moves ❌ Fail to follow the camera trajectory The core issue? → No persistent world representation

CVPR 2026🎥We built a model that lets you fly through a generated video world. Not just generating frames — but maintaining a consistent 3D world under complex camera motion. Code, ckpt, and even the data pipeline are all open-sourced ↓ #AI #worldmodel #videogen #cvpr #drone

[3/4] Pretty markdown ≠ perception. OSS often formats well yet misses visual evidence or the question intent. Without grounding, style can even mask failure modes. We measure the gap explicitly on MMMU. #LLM #MMMU #VisualAlignment

[2/4] Longer chains, higher gains. Qwen’s deliberate, multi-step reasoning ties evidence→logic→answer, especially in STEM. OSS tends to be shorter, so accuracy lags when problems require real visual grounding. #STEM #Reasoning #LLM

[1/4] We tested InternVL3.5-aligned GPT-OSS on MMMU and found visual alignment gaps: misreads, shallow grounding, and repetition. It’s an early step, not the solution—good scaffolding, weak seeing. #Multimodal #VisionLanguage