Research map / Video / Video generation
MAD: Motion Appearance Decoupling for efficient Driving World Models
Ahmad Rahimi, Valentin Gerard, Eloi Zablocki, Matthieu Cord, Alexandre Alahi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| decouple motion (pose video) and appearance stages instead of end-to-end LoRA fine-tuning of LTX (ablation) | MAD-LTX 2B · OpenDV human preference | preference count: 29 vs 16 | Improved |
| decouple motion and appearance instead of end-to-end fine-tuning (ablation) | MAD-LTX 13B · OpenDV human preference | preference count: 33 vs 15 | Improved |
| decouple motion and appearance instead of end-to-end fine-tuning (ablation) | MAD-LTX 13B · OpenDV open-loop planning | minADE6 (m): 5.83 -> 3.64 | Improved |
| decouple motion and appearance instead of end-to-end fine-tuning (ablation) | MAD-LTX 13B · OpenDV open-loop planning | APD6 (cm): 63.06 -> 101.45 | Improved |
| standard end-to-end fine-tuning of base LTX | LTX 13B · OpenDV open-loop planning | minADE6 (m): 4.14 -> 5.83 | Hurt |
| standard end-to-end fine-tuning of base LTX | LTX 2B · OpenDV open-loop planning | minADE6 (m): 5.42 -> 5.28 | Improved |
| add targeted latent noise injection to appearance synthesizer pose conditioning (ablation) | MAD-LTX 2B · OpenDV human preference | human preference for full model (%): 62% | Improved |
| use pose skeleton representation instead of panoptic segmentation (ablation) | MAD-LTX 2B · OpenDV human preference | human preference for pose model (%): 74% | Improved |
| use pose skeleton representation instead of HDMap (ablation) | MAD-LTX 2B · OpenDV human preference | human preference for pose model (%): 78% | Improved |
| adapt SVD with MAD instead of VISTA-style full fine-tuning | MAD-SVD · adaptation cost | GPU hours: 25,000 -> 1500 | Improved |
Closest papers by meaning
- DrivingGen: A Comprehensive Benchmark for Generative Video World Models in Autonomous Driving (ICLR 2026 · similarity 0.90)
- Motion Forcing: Decoupling Ego and Object Motion via Sparse Inputs for Structured Video Generation (NeurIPS 2026 · similarity 0.89)
- MaskGWM: A Generalizable Driving World Model with Video Mask Reconstruction (CVPR 2025 · similarity 0.88)
- DualAD: Disentangling the Dynamic and Static World for End-to-End Driving (CVPR 2024 · similarity 0.87)
- DriveDreamer4D: World Models Are Effective Data Machines for 4D Driving Scene Representation (CVPR 2025 · similarity 0.87)
- Grounding Driving VLA via Inverse Kinematics (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
