[CVPR 2026] MoRe: Disentangling the Dynamic World for Real-Time 4D Reconstruction

TL;DR

MoRe (Motion-aware Reconstruction) is a novel feed-forward Transformer that solves a classic "nightmare" in computer vision: reconstructing a 3D scene from a single video while objects are moving. By introducing an Attention-Forcing strategy and Grouped Causal Attention, MoRe filters out "distracting" motion to estimate camera poses and depth maps accurately in real-time (30 FPS), outperforming previous streaming-based SOTA methods.

The Problem: When the World Moves, SLAM Fails

Most 3D reconstruction systems operate under the "Static World Assumption." When a car drives by or a person walks through the frame, the model gets confused—is the camera moving, or is the object moving? This ambiguity leads to corrupted camera trajectories and "ghosting" artifacts in the 3D geometry.

Existing solutions usually fall into two camps:

1. Optimization-based: High accuracy but painfully slow (non-real-time).
2. Feed-forward Transformers: Fast, but they lack explicit mechanisms to handle dynamic motion, leading to degraded performance in real-world scenarios.

Methodology: The "MoRe" Insight

The core of MoRe lies in its ability to disentangle motion. It doesn't need motion masks at test time; instead, it learns where to look during training.

1. Motion-Aligned Attention-Forcing

The researchers observed that in vanilla Transformers (like VGGT), the "camera tokens" often attend to moving pixels, which poisons the pose estimation. MoRe uses an Attention-Forcing loss during training. Using ground-truth motion masks, it penalizes camera tokens if they attend to moving objects, forcing them to find stability in the static background.

Figure 2: The MoRe pipeline. Note how the Attention-Forcing mechanism aligns attention weights with motion masks during training.

2. Grouped Causal Attention (GCA)

Standard Causal Attention (used in LLMs) treats every pixel token as a sequence. This is bad for images because pixels within the same frame need to "see" each other to understand geometry. GCA allows bidirectional attention within a frame but maintains a "one-way" causal attention across time, enabling efficient KV-caching for streaming video.

Figure 4: Grouped Causal Attention allows spatial reasoning within frames while preserving temporal causality.

Experiments: Superior Accuracy at 30 FPS

MoRe was tested on challenging dynamic datasets like Sintel and Bonn. It consistently outperformed other streaming models such as Stream3R and CUT3R.

One of the most impressive feats is its zero-shot generalization. MoRe was never trained on the Bonn or TUM-dynamics datasets, yet it accurately reconstructed them, proving its robustness.

Figure 6: Qualitative results showing MoRe's ability to maintain clean geometry while other methods warp due to object motion.

Deep Insight & Conclusion

MoRe's success suggests a shift in how we should treat 3D foundation models. Rather than just adding more data, architectural inductive biases (like GCA) and targeted supervision (like Attention-Forcing) are the keys to moving from static 3D to fluid 4D.

Limitations: The model still relies on high-quality motion mask labels during the training phase. If the training masks are noisy, the attention-forcing becomes less effective. Future iterations might solve this using self-supervised flow discrepancy.

Future Outlook: MoRe paves the way for truly real-time AR and robotic navigation in crowded, dynamic environments where traditional SLAM has historically failed.