Newest papers

CausalAlign: Causal Alignment of Video Generation Models for Consistent Ego-Motion

Kai Zhu, Yintianrun Zhang, Shuwei Jiang, Ruolin Zuo, Yixin Chen, Duomin Wang, Miao Liu, Yin Li, Daquan Zhou, Yiwu Zhong · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.29915 · cited by 0 works (OpenAlex) · open-access PDF

Indonesian

Model generasi video berbasis difusi (VGMs) dapat menghasilkan video yang dapat diandalkan secara visual, namun gerakan ego kamera mereka sering mengalir melalui cakrawala yang panjang. Kamera ego-motion menentukan bagaimana sudut pandang berubah antara bingkai dan penting untuk mempertahankan geometri dan trajektori yang konsisten. Untuk mengatasi masalah ini, kami mengusulkan CausalAlign, kerangka kerja penyesuaian kausal dengan dua komponen. Pertama, kami mendiagnosis bagaimana gerakan ego yang akurat diwakili, digunakan, dan disimpan di dalam VGM menggunakan prob linear dan intervensi kausal, mengungkapkan pola naik dan turun yang konsisten selama denoise. Kedua, kami memperkuat keadaan ego-motion di VGM dengan menyesuaikan representasi pada lapisan yang diidentifikasi ke model dasar 3D melalui ruang delta-token. Ruang ini mengkodekan perubahan

Machine translation (Meta M2M100); the English original follows.

Abstract

Diffusion-based video generation models (VGMs) can produce visually plausible videos, yet their camera ego-motion often drifts over long horizons. Camera ego-motion determines how the viewpoint changes between frames and is essential for maintaining consistent geometry and trajectories. To address this problem, we propose CausalAlign, a causal alignment framework with two components. First, we diagnose how precise ego motion is represented, used, and retained within VGMs using linear probes and causal interventions, revealing a rise-and-fall pattern during denoising. Second, we strengthen ego-motion state in VGM by aligning the representations at the identified layers to a 3D foundation model through delta-token space. This space encodes the changes between consecutive frames, allowing the alignment to focus on ego-motion dynamics rather than static scene content. With Wan2.2 fine-tuned on DL3DV and RealEstate10K, CausalAlign can largely improve motion accuracy, geometric consistency, and temporal smoothness, while preserving visual quality. For example, it improves geometric consistency by more than 20% relative to the base model. Moreover, with the fine-tuned VGM, CausalAlign improves the success rate of world action models (WAMs) by 10 points on mobile manipulation tasks of RoboCasa365 benchmark. Overall, CausalAlign improves ego-motion consistency in diffusion-based VGMs and yields noticeable gains in embodied models.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer vision, Artificial intelligence, Consistency (knowledge bases), Computer science, Motion (physics), Focus (optics)

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange