CausalAlign aligns video generation to 3D dynamics, boosting geometric consistency over 20%
Kai Zhu · hf · 2026-10-27
Researchers from Peking University, NVIDIA, Tsinghua and others propose CausalAlign to fix long-horizon camera ego-motion drift in diffusion video generators.
Method and results:
- Linear probes and causal interventions reveal ego motion is represented with a rise-and-fall pattern during denoising
- Aligning internal representations at the identified layers to a 3D foundation model via a delta-token space (encoding changes between consecutive frames, focusing on dynamics rather than static scene content)
- Fine-tuned on Wan2.2 with DL3D and RealEstate10K, geometric consistency improves by more than 20% relative to the base model, with better motion accuracy and temporal smoothness without sacrificing visual quality
- As a world action model (WAM), the fine-tuned generator improves success rate by 10 points on RoboCasa365 mobile manipulation tasks
Bridges video generation quality with embodied applications needing consistent camera motion.
More from Multimodal
- Claude Code Plus MiniMax Video Model Generates Fast Combat Anime Clips — Inevitable-Bus-8654 · 2026-10-27
- Turn product photos into interactive 3D-style displays with AI video frames, no 3D modeling — aigclink · 2026-10-27
- Full walkthrough: interactive product displays from photos using GPT-6 Astra and AI video as UI — MaziyarPanahi · 2026-10-27
- MiniMax H3 image-to-video produces heavy artifacts, worse with photos — Radiant-Photograph46 · 2026-10-27
- AI-generated sitcom-style video with a sloth goes viral for its uncanny realism — heresalexandria · 2026-10-27
- AI virtual artist MALFOID releases MV teaser for new single — staszh · 2026-10-27