Research map / Image generation / Autoregressive and latent generation
Decoupled Diffusion Sparks Adaptive Scene Generation
Yunsong Zhou, Naisheng Ye, William Ljungbergh, Tianyu Li, Jiazhi Yang, Zetong Yang, Hongzi Zhu, Christoffer Petersson, Hongyang Li
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 |
|---|---|---|---|
| add noise-masking training (ablation) | Diffusion Policy baseline · nuPlan conditioned generation | ADE: 7.53 -> 3.42 | Improved |
| add positional embedding of physical and denoising time (ablation) | Nexus with noise masking · nuPlan conditioned generation | ADE: 3.42 -> 1.44 | Improved |
| add Nexus-Data safety-critical training data (ablation) | Nexus · nuPlan conditioned generation | ADE: 1.44 -> 1.32 | Improved |
| add classifier guidance (ablation) | Nexus · nuPlan conditioned generation | collision rate R_col (%): 1.92 -> 1.47 | Improved |
| add classifier guidance (ablation) | Nexus · nuPlan conditioned generation | instability M_k: 4.87 -> 4.02 | Improved |
| use pyramidal or trapezoidal scheduling instead of full-sequence (ablation) | Nexus · nuPlan scheduling comparison | reaction time (sec): 4.96 -> 0.16 | Improved |
| use trapezoidal scheduling instead of full-sequence (ablation) | Nexus · nuPlan conditioned generation | ADE: 1.28 -> 1.39 | Hurt |
| update histories with interaction feedback in trapezoidal scheduling (ablation) | Nexus trapezoidal · nuPlan conditioned generation | ADE: 1.39 -> 1.17 | Improved |
| use autoregressive scheduling instead of full-sequence (ablation) | Nexus · nuPlan scheduling comparison | overall time (sec): 4.96 -> 79.36 | Hurt |
| add step skipping to reduce sampling steps to 18 (ablation) | Nexus · nuPlan 8-second scene generation | generation time (sec): 2.79 | Improved |
| add Nexus-Data and classifier guidance (Nexus-Full) (ablation) | Nexus · nuPlan conditioned generation | ADE: 1.28 -> 1.12 | Improved |
| add 3x synthetic Nexus data to real data | lightweight planner · nuPlan closed-loop reactive | score: 48.11 -> 46.61 | Hurt |
Closest papers by meaning
- SceneDiffuser: Efficient and Controllable Driving Simulation Initialization and Rollout (NeurIPS 2024 · similarity 0.90)
- DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving (CVPR 2025 · similarity 0.89)
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.88)
- X-Scene: Large-Scale Driving Scene Generation with High Fidelity and Flexible Controllability (NeurIPS 2025 · similarity 0.88)
- SimGen: Simulator-conditioned Driving Scene Generation (NeurIPS 2024 · similarity 0.88)
- Generating Multimodal Driving Scenes via Next-Scene Prediction (CVPR 2025 · 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.
