Driving · 2026-10-22

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

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

CausalAlign diagnoses where ego-motion information is represented, used, and forgotten inside diffusion-based video generation models, then preserves it by aligning VGM representations to 3D delta-token supervision. Using Wan2.2 fine-tuned on DL3DV and RealEstate10K, the authors report improved motion accuracy, geometric consistency, and temporal smoothness, while maintaining visual quality. They also report gains for world action models on mobile manipulation.

Why it matters

Diffusion-based VGMs can produce plausible visuals but drift in camera ego-motion over long horizons, hurting geometric consistency and downstream control. By targeting ego-motion “geometric state forgetting” with causal diagnosis and delta-token alignment, the method aims to improve internal ego-motion retention rather than only optimizing output-level preferences. The reported RoboCasa365 improvements suggest relevance to embodied world action models.

Method

Causal diagnosis: use linear probes to measure ego-motion decodability and causal interventions to measure how hidden-state changes affect generated camera trajectories, revealing a rise-and-fall pattern across denoising steps and Transformer layers.
Delta-token construction: compute delta tokens encoding changes between consecutive frames, using a pretrained 3D foundation model as geometric supervision in a delta-token space.
CausalAlign training: fine-tune Wan2.2 with LoRA updates at causally selected layers/timesteps, supervising internal ego-motion states via delta-token alignment while keeping other components frozen.

Abstract (from arXiv)

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.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-21 · Indexed by Pub-AI: 2026-10-22 UTC

preprint