MAD: Motion Appearance Decoupling for efficient Driving World Models
TLDR
Proposes MAD, a motion-appearance decoupling framework to efficiently adapt video diffusion models into controllable driving world models.
Reasoning
Strengths include novel decoupling approach, high efficiency (matching SOTA with <6% compute), and comprehensive controls. Weaknesses: abstract lacks discussion of limitations or failure cases.
Read-first score
Read-first score 61.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 53.
Field roles
Rank sensitivity
Stability: volatile; rank range: 413.
Keyword Scores
Deep Analysis
Innovations
- Decoupling motion learning from appearance synthesis in a two-stage reasoning-rendering paradigm
- Adapting generalist video diffusion models to driving world models with minimal supervision by first predicting structured motion in skeletonized form
- Efficient adaptation that matches prior SOTA with less than 6% of their compute
- MAD-LTX model outperforms all open-source competitors and supports text, ego, and object controls
Methodology
The proposed framework adapts generalist video diffusion models (e.g., SVD, LTX) into controllable driving world models via a two-stage process. First, the model is adapted to predict structured motion in a simplified skeletonized form of agents and scene elements, focusing learning on physical and social plausibility. Then, the same backbone is reused to synthesize realistic RGB videos conditioned on these motion sequences, effectively 'dressing' the motion with texture and lighting.
Key Results
Adapting SVD, the method matches prior state-of-the-art models with less than 6% of their compute. Scaling to LTX, MAD-LTX outperforms all open-source competitors and supports a comprehensive suite of text, ego, and object controls.