DDIM-DSC: 4× Downscaling of Wind Velocities

DDIM-DSC is a custom-trained Denoising Diffusion Implicit Model (DDIM) designed for the downscaling of wind velocity fields from coarse- to high-resolution using reanalysis data.

It performs 4× spatial downscaling on 2-channel wind fields (u and v components), using ERA5 as low-resolution input and COSMO-REA6 as the high-resolution target.

---

📊 Data

• Input: ERA5 100 m wind components (u, v), 2 channels
• Target: COSMO-REA6 100 m wind components (u, v), 2 channels
• Sequence length: 3 (with temporal context across 3 timesteps)
• Total input channels: 8 (2 channels × 3 timesteps + 2 static channels)

---

🧠 Model Architecture

• Model type: DDIM (using diffusers)
• Scheduler: DDIMScheduler
• Conditioning: Concatenated temporal sequences
• Latent noise sampling: 10 per input
• Scale factor: 4×
• Input channels: 8
• Output channels: 2
• Note: The low-resolution input must be resized to high-resolution shape using bilinear interpolation before being passed into the model.

🚀 Usage

import torch
from diffusers import DiffusionPipeline

# load the custom DDIM pipeline
pipe = DiffusionPipeline.from_pretrained(
    "lschmidt/ddim-dsc",
    custom_pipeline="cond_ddim_pipeline",
    trust_remote_code=True
)

# create a sample low-resolution input --> shape: (sequence_length, channels, height, width)
lres_image = torch.randn((3, 2, 32, 32)).to(pipe.device)

# interpolate to match high-resolution


# run inference
outputs = pipe(image=lres_image)