Research map / Image generation / Fast diffusion sampling
Diffusion Distillation with Direct Preference Optimization for Efficient 3D LiDAR Scene Completion
An Zhao, Shengyuan Zhang, Zejian Li, Ling Yang, Pei Chen, Jiale Wu, Haoran Xu, AnYang Wei, Perry Pengyun Gu, Lingyun Sun
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 preference-pair DPO term to score distillation (ablation) | LiDiff student (8 NFE) · SemanticKITTI (Refined) | Chamfer Distance: 0.419 -> 0.354 | Improved |
| plain score distillation vs teacher (ablation) | LiDiff · SemanticKITTI (Refined) | Chamfer Distance: 0.375 -> 0.419 | Hurt |
| distill 50-step LiDiff into 8-step student with preference alignment | LiDiff · SemanticKITTI (Refined) | time (s): 17.87 -> 3.38 | Improved |
| reduce sampling steps from 8 to 4 (ablation) | Distillation-DPO · SemanticKITTI (Refined) | Chamfer Distance: 0.354 -> 0.369 | Hurt |
| reduce sampling to 1 step (ablation) | Distillation-DPO · SemanticKITTI (Refined) | Chamfer Distance: 0.354 -> 0.601 | Hurt |
| increase initial noise scale lambda from 1.1 to 1.5 (ablation) | Distillation-DPO pair generation · SemanticKITTI (Refined) | Chamfer Distance: 0.354 -> 0.409 | Hurt |
| decrease lambda from 1.1 to 1.05 (ablation) | Distillation-DPO pair generation · SemanticKITTI (Refined) | Chamfer Distance: 0.354 -> 0.358 | Hurt |
| use a stronger teacher (LiDiff refined with Diffusion-DPO) (ablation) | Distillation-DPO · SemanticKITTI (Refined) | Chamfer Distance: 0.354 -> 0.343 | Improved |
| choose preference pairs by JSD instead of CD (ablation) | Distillation-DPO · SemanticKITTI (Refined) | Chamfer Distance: 0.354 -> 0.382 | Hurt |
| use two auxiliary diffusion models (win and lose) instead of one (ablation) | Distillation-DPO · SemanticKITTI (Refined) | Chamfer Distance: 0.372 -> 0.354 | Improved |
| Distillation-DPO vs refined LiDiff at 8 NFE | Distillation-DPO · SemanticKITTI (Refined) | EMD: 23.16 -> 23.66 (vs LiDiff 50 NFE) | Hurt |
Closest papers by meaning
- Scaling Diffusion Models to Real-World 3D LiDAR Scene Completion (CVPR 2024 · similarity 0.89)
- Towards Realistic Scene Generation with LiDAR Diffusion Models (CVPR 2024 · similarity 0.88)
- Three Pillars Improving Vision Foundation Model Distillation for Lidar (CVPR 2024 · similarity 0.87)
- Fine-grained Image-to-LiDAR Contrastive Distillation with Visual Foundation Models (NeurIPS 2024 · similarity 0.86)
- LiFi: LiDAR Generation from Multi-View Images via Geometric and Semantic Collaborative Guidance (NeurIPS 2026 · similarity 0.86)
- Distilling Monocular Foundation Model for Fine-grained Depth Completion (CVPR 2025 · similarity 0.86)
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.
