Daily Paper Briefing

Abstract Overview

MVTrack4Gen is a motion-aware training framework for camera-conditioning-only novel-view video diffusion models. The paper analyzes 3D attention in these models and finds that certain intermediate layers encode both temporal correspondences within a view and cross-view correspondences between reference and target videos, with failures in these correspondences linked to motion and geometric inconsistencies. Based on this observation, the method adds an auxiliary multi-view tracking head and a correspondence loss that directly supervises attention using multi-view point tracks. The framework is applied to ReCamMaster and Redirector, aiming to improve motion fidelity and cross-view geometric consistency without relying on explicit 3D reconstruction at inference time.

Novelty

The paper’s distinctive contribution is to treat emergent correspondences in specific diffusion attention layers as a source of geometric supervision for 4D video generation. It combines attention-level correspondence supervision with an auxiliary multi-view point-tracking objective, using shared DiT features to strengthen both temporal motion cues and cross-view alignment.

Results

Across DAVIS and iPhone benchmarks, MVTrack4Gen consistently improves the two camera-conditioning-only backbones it is attached to, especially on geometric consistency metrics. On DAVIS, the method reports the lowest MEt3R and dynamic-region MEt3R among compared methods, along with improved translation error and camera-motion consistency; on iPhone, it improves PSNR, LPIPS, and MEt3R over the camera-conditioning-only baselines. Ablations indicate that the tracking module and correspondence loss are complementary, with their combination giving the best overall trade-off across quality, geometry, and camera accuracy.

Key Points

1. The method is motivated by an analysis showing that specific intermediate 3D attention layers in video diffusion models contain strong temporal and cross-view correspondence cues.
2. MVTrack4Gen adds two supervision signals—a multi-view tracking head and a correspondence loss on attention maps—to improve motion following and geometric consistency.
3. The framework improves ReCamMaster and Redirector without explicit 3D reconstruction at inference, with the strongest gains reported on geometric consistency and dynamic-object alignment.

References

• arXiv: https://arxiv.org/abs/2606.26087v1
• Fugu-MT: https://fugumt.com/fugumt/paper_check/2606.26087v1
• Hugging Face Papers: https://huggingface.co/papers/2606.26087