MeRoPE: norm-preserving metric rotary encoding improves camera-controlled video generation
zhenjun_zhao · x · 2026-09-05
A new paper targets scale-dependent failure modes in geometry-aware positional encodings for camera-controlled video generation, where homogeneous projective encodings make attention logits and feature norms grow unbounded with physical translation baselines. MeRoPE (Metric Rotary Position Embedding) encodes relative orientations between calibrated viewing rays with orthogonal rotation blocks, maps metric displacements into multi-frequency rotary phases, and adds a disparity-anchored epipolar prior. It preserves feature norms, bounds pre-softmax logits regardless of scale, and stays invariant to global rigid transforms, achieving stronger camera control on nuScenes and PanShot. Code will be released.
More from Research
- DeepMind's 100-agent simulated conference descends into cheaters, converts and whistleblowers in 27 minutes — The Decoder · 2026-09-05
- HIM Arena opens robot sports challenges: 15 community sims on MuJoCo and Unitree G1 — ericjang11 · 2026-09-05
- Full recipe: running Qwen3.8 27B on AMD Strix Halo with patched ROCm llama.cpp — ilintar · 2026-09-05
- A 5M-parameter adapter injects cross-architecture semantics into MiniMax H3 — NoMouse9610 · 2026-09-05
- Do AI solvers already do Peircian abduction? Researchers debate the induction boundary on ARC-AGI — balazskegl · 2026-09-05
- Michael Levin preprint: membrane voltage and connexin expression jointly drive tumor growth — drmichaellevin · 2026-09-05