🌊 DynaDiff

Official implement of ICML'26 "Adaptation of Dynamics to Environmental Shifts via Weight-space Diffusion", a novel approach for explicitly modeling the conditional dependence of the predictive model on environments.

 

🚀 Demo (Cylinder Flow)

DynaDiff relies only on common deep learning libraries, including torch, tqdm, torchmetrics, and sklearn.

Experience our pretrained model by generating weights for a new environment in just 10 seconds using only 1.7GB GPU memory!

Steps to Run the Demo

1️⃣ Download Checkpoints📂(3.96GB): download link

2️⃣ Execute: Open and run demo.ipynb

1. Select Environment

   Default setting: Re=210.0, r=10.0, Re=240.0, r=19.0, Re=270.0, r=24.0

   Download additional environments📂(0.70GB): download link

2. Generate Model Weights

   The model generates weights in approximately 10 seconds with only 1.7GB GPU memory.

3. Evaluate Model Weights

4. Visualize Results

3️⃣ Train from scratch

• Download Model Zoo📂(about 18GB): available post-acceptance

• Data Preprocess

  python fno_LDM/train/prepare.py

• Train Weight VAE

  python fno_LDM/train/_cy_vae.py --zoo_size 1

• Train Weight Diffusion

  python fno_LDM/train/_cy_ldm.py --zoo_size 1

 

📁 Repository Structure

.
├── README.md
├── asset
│   ├── result.png
│   └── result2.png
├── data
│   └── cy_ (Additional files available for download)
│       ├── Re_210.0_r_10.0.npy
│       ├── Re_270.0_r_24.0.npy
│       └── Re_290.0_r_19.0.npy
├── demo.ipynb
├── fno_LDM
│   ├── model
│   │   ├── __init__.py
│   │   ├── diffusion.py
│   │   ├── diffusion_utils.py
│   │   ├── dit.py
│   │   ├── fno.py
│   │   ├── fno_graph_construct.py
│   │   ├── fno_graph_vae.py
│   │   └── graph_normalize.py
│   └── train
│       ├── _cy_ldm.py
│       ├── _cy_vae.py
│       ├── evaluate.py
│       ├── prepare.py
│       └── utils.py
├── weights (Download required)
│   ├── ldm.pt
│   └── vae.pt
└── zoo
    └── cy_
        └── fno
            └── minmax_dict.pkl
            └── origin
            │   ├── Re210.0_r10.0
            │   │   └── seed0
            │   │       └── epoch1000.pt
            │   ├── Re240.0_r19.0
            │   │   └── seed0
            │   │       └── epoch1000.pt
            │   └── Re270.0_r24.0
            │       └── seed0
            │           └── epoch1000.pt
            └── Additional training files (Available post-acceptance)

 

📌 Notes

• This implementation provides a demo of Cylinder Flow system .
• The weights provided in the Model Zoo are limited to one per environment for demonstration purposes. The full model zoo will be released after acceptance.
• The complete version, including additional environments and datasets, will be made available upon manuscript acceptance.