Research map / Video / Video generation
Pretrained Video Models as Differentiable Physics Simulators for Urban Wind Flows
Janne Perini, Rafael Bischof, Moab Arar, Ayça Duran, Michael Kraus, Siddhartha Mishra, Bernd Bickel
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 |
|---|---|---|---|
| full fine-tuning of DiT instead of LoRA (rank 512) (ablation) | WinDiNet · Urban wind test set | VRMSE: 0.746 -> 0.702 | Improved |
| scalar embedding conditioning instead of text prompts (ablation) | WinDiNet · Urban wind test set | VRMSE: 0.702 -> 0.616 | Improved |
| add color adapter around frozen VAE (ablation) | WinDiNet · Urban wind test set | VRMSE: 0.616 -> 0.577 | Improved |
| fine-tune VAE decoder (ablation) | WinDiNet · Urban wind test set | VRMSE: 0.616 -> 0.531 | Improved |
| add physics-informed losses to decoder fine-tuning (ablation) | WinDiNet · Urban wind test set | VRMSE: 0.531 -> 0.520 | Improved |
| fine-tune VAE decoder (ablation) | VAE reconstruction · VAE reconstruction (test set) | VRMSE: 0.188 -> 0.084 | Improved |
| add color adapter (ablation) | VAE reconstruction · VAE reconstruction (test set) | W1: 0.0157 -> 0.0167 | Hurt |
| add physics-informed losses to decoder fine-tuning (ablation) | VAE reconstruction · VAE reconstruction (test set) | VRMSE: 0.084 -> 0.070 | Improved |
| train DiT from scratch on fine-tuned VAE (remove video pretraining) (ablation) | WinDiNet Dec. FT Physics · Urban wind test set | VRMSE: 0.52 -> 0.57 | Hurt |
| train both DiT and VAE from scratch (ablation) | WinDiNet Dec. FT Physics · Urban wind test set | VRMSE: 0.52 -> 0.92 | Hurt |
| use higher classifier-free guidance scale (above 1.0) (ablation) | WinDiNet · validation set grid search | VRMSE | No clear effect |
| use WinDiNet instead of OFormer as optimization surrogate | layout optimizer · Single-inlet inverse optimization (GT CFD) | discomfort fraction >5 m/s: 36.63% -> 12.80% | Improved |
Closest papers by meaning
- GeoWind2Plan: Mission-Time 3D Urban Wind Prediction for Energy-Efficient UAV Planning (NeurIPS 2026 · similarity 0.85)
- 3DID: Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization (NeurIPS 2025 · similarity 0.82)
- EngiBench: A Framework for Data-Driven Engineering Design Research (NeurIPS 2025 · similarity 0.81)
- Mosaic: A Benchmark Suite for Differentiable Physics Solvers (NeurIPS 2026 · similarity 0.81)
- DrivAerStar: An Industrial-Grade CFD Dataset for Vehicle Aerodynamic Optimization (NeurIPS 2025 · similarity 0.81)
- NeuralBES: A Differentiable, Control-Aware Emulator for Scalable Building Energy Modeling (NeurIPS 2026 · similarity 0.81)
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.
