MeRoPE encodes metric camera translation as rotary phases without exploding attention logits

MeRoPE: Metric Rotary Position Embedding for Camera-Controlled Video Generation

Zhijian Qiao, Xinjiang Wang, Jiajie Chen, Haoming Huang, Meng Li, Chih-Chung Chou, Jing Wang, Shaojie Shen

cs.CV, cs.RO

2026-09-01

MeRoPE maps metric camera displacement into rotary phases so attention logits stay bounded. CamMC is 2.32 on nuScenes vs UCPE 2.45; PanShot TransErr 12.66 vs UCPE 17.91.

What problem this solves

Camera-controlled video generation writes extrinsics and per-token viewing rays into attention. Homogeneous encodings such as GTA, PRoPE, and UCPE pack rotation and translation into a non-orthogonal 4×4 block. Translation then enters the score as an unnormalized inner product. Stretch the physical baseline and both logits and feature norms grow without bound; distant frames steal attention from nearby covisible ones.

Three properties cannot hold together: full metric relative pose, strict per-token factorization, and norm-preserving action. An appendix theorem says any continuous finite-dimensional unitary encoding that factorizes per token is blind to metric translation. Something has to give. This paper drops strict per-token factorization and groups by query camera.

Method

MeRoPE splits relative pose into two orthogonal operators and adds a disparity prior.

A minimum-rotation (MinRot) frame carries the optical axis onto each calibrated unit ray. UCPE's cross-product frame degenerates as vertical FoV approaches 180°; MinRot is singular only looking backward. Relative rotation between query and key rays is a repeated 3×3 orthogonal block, so feature norms stay exactly one.

Metric translation is not a homogeneous add. The 3D displacement in the query frame becomes multi-frequency rotary phases on x, y, and z, with log-spaced wavelengths. Each 2D rotation is orthogonal: a longer baseline spins the phase faster and does not scale the logit.

Static-scene correspondence uses spherical disparity anchors. As depth on a key ray runs from infinity to zero, directions sweep a great-circle arc from the infinite-depth ray to the epipole. Bounded angular fractions along that arc become extra MinRot blocks from the query ray. Attention is not masked to pixel epipolar lines, and no dataset-specific metric depth bins are required. Degenerate cases (coincident centers, epipole parallel to the infinite ray) collapse every anchor to the infinite-depth direction.

The full operator is the orthogonal direct sum of disparity, rotation, translation, and the backbone's native spatiotemporal RoPE. Integration follows UCPE: a parallel camera-attention branch with a zero-initialized residual into a frozen video DiT.

Results

All methods share Wan2.2 TI2V-5B. Training mixes nuScenes driving clips with World Engine synthetic trajectories (347.7k sequences) for 20k steps on 64 H20 GPUs. Poses are recovered from generated frames with VGGT-Ω.

Methodrot°↓tr%↓CamMC↓FID↓
No camera PE5.575.947.9621.06
URoPE1.662.672.5319.98
UCPE1.392.952.4520.20
MeRoPE1.372.842.3219.82

CamMC is the headline metric because it scores rotation and translation together. MeRoPE is lowest, and FVD 134.15 is the best point estimate in the table. URoPE still wins tr% and AUC; dropping metric translation can even raise AUC, which ignores translation magnitude.

On PanShot (pinhole through fisheye), with near-depth normalization off so raw metric displacements are kept, MeRoPE records TransErr 12.66 and CamMC 15.56 against UCPE's 17.91 / 20.13. UCPE remains slightly better on rotation (4.40 vs 4.82). Query-camera grouping costs 12.47 ms per CamSA block versus 5.03 ms for UCPE; full-model training throughput drops only 3.3%.

Why it matters

Driving and simulation need commanded trajectories in meters, not after a global scale wipe. MeRoPE shows metric translation can live in attention if it is a rotary phase rather than a homogeneous add. The comparison is a matched-backbone swap, so the gain is the encoding. It is an adapter on a pretrained DiT. Translation control moves more than rotation, and the gap widens on large baselines and odd optics.

Limitations

Grouping by query camera relaxes per-token factorization. Isolated CamSA memory is about 3× UCPE; a fused kernel is future work. MinRot is singular looking backward. URoPE still leads nuScenes tr% and AUC, so disparity anchors do not dominate every directional score. Recovered poses come from VGGT-Ω, tying generation quality to that extractor. Training includes 244.1k synthetic clips; real long-tail motion remains thin.

Terms

Source

What people are saying

Related papers

All paper explainers