---
license: cc-by-nc-4.0
tags:
  - diffusion
  - ddpm
  - 3d
  - volumetric
  - monai
  - generative
  - synthetic-data
  - atmospheric-science
---

# 3D Convective-Cloud Denoising Diffusion Probabilistic Model

An unconditional **3D Denoising Diffusion Probabilistic Model (DDPM)** that
generates synthetic **64×64×64 convective-cloud cells** (vertical-velocity
volumes). Roughly 190M parameters.

**What it is for:** producing synthetic 3D cloud cells to augment scarce
atmospheric-simulation data — for example as additional training data for 3D
cloud segmentation.

**What it is not:** this is not a weather-forecasting model and not a physical
simulation. It samples volumes resembling its training distribution; it does not
solve atmospheric dynamics.

**Project:** [3D Convective-Cloud Generation collection](https://huggingface.co/collections/HoqueMahmudul/3d-convective-cloud-generation-ddpm-190m-params-6a6ba230c89356ff617d481a)

> **Status:** research artifact from work with an associated manuscript **in
> preparation**. Methodology, training details, and evaluation are deliberately
> not described here and will appear in the publication.

## Requirements

This is a **MONAI** model, not a `transformers` model. You need
**`monai-generative`** (which provides the `generative` package) to build the
network before loading the weights:

```bash
pip install monai-generative==0.2.3 monai==1.5.2 torch safetensors huggingface_hub
```

A `requirements.txt` with the verified-working versions is included in this repo:

```bash
pip install -r <(curl -sL https://huggingface.co/HoqueMahmudul/convective-cloud-denoising-diffusion-probabilistic-model-3d/raw/main/requirements.txt)
```

> [!IMPORTANT]
> **Use `monai-generative`, not MONAI core.** MONAI core (>= 1.4) ships its own
> `DiffusionModelUNet`, `DDPMScheduler` and `DiffusionInferer`, and they look
> like drop-in replacements. **They are not weight-compatible with this
> checkpoint.**
>
> | | `monai-generative` 0.2.3 | MONAI core 1.5 |
> |---|---|---|
> | Attention keys | `...attentions.0.to_q.weight` | `...attentions.0.`**`attn`**`.to_q.weight` |
> | Constructor arg | `num_channels=[...]` | `channels=[...]` |
> | Strict load here | works | **52 missing / 52 unexpected keys** |
>
> Both classes have the *identical* parameter count (189,815,297), so if you hit
> the key error and "fix" it with `strict=False`, you will get a model with **52
> randomly-initialised attention tensors**. It will run and produce
> plausible-looking output, and it will be silently wrong.
>
> **Never load these weights with `strict=False`.** If a strict load fails, the
> environment is wrong — fix the environment.

## Files

| File | What it is |
|---|---|
| `best_model.safetensors` | The model weights (a plain PyTorch `state_dict`) |
| `requirements.txt` | Verified-working dependency versions |

## How to load and generate a sample

The configuration below is **required** — the weights are a bare `state_dict`, so
the network must be constructed exactly this way before loading.

```python
import torch, numpy as np
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from generative.networks.nets import DiffusionModelUNet
from generative.networks.schedulers import DDPMScheduler
from generative.inferers import DiffusionInferer

REPO = "HoqueMahmudul/convective-cloud-denoising-diffusion-probabilistic-model-3d"
device = "cuda"

# 1. build the network with exactly this configuration
model = DiffusionModelUNet(
    spatial_dims=3, in_channels=1, out_channels=1,
    num_channels=[256, 256, 512],
    attention_levels=[False, False, True],
    num_head_channels=[0, 0, 512],
    num_res_blocks=2,
)

# 2. load the weights -- strict=True is deliberate; see the warning above
weights = hf_hub_download(REPO, "best_model.safetensors")
model.load_state_dict(load_file(weights), strict=True)
model.to(device).eval()

# 3. the sampler must use these settings
scheduler = DDPMScheduler(
    num_train_timesteps=1000, schedule="scaled_linear_beta",
    beta_start=0.0005, beta_end=0.0195,
)
scheduler.set_timesteps(1000)
inferer = DiffusionInferer(scheduler)

# 4. sample one volume (1000 reverse steps -- use a GPU)
noise = torch.randn((1, 1, 64, 64, 64), device=device)
with torch.no_grad(), torch.amp.autocast("cuda"):
    x = inferer.sample(input_noise=noise, diffusion_model=model, scheduler=scheduler)

# 5. convert the [0,1] output back to raw units
#    clip the LOWER bound at 1, not 0: the model's background output is very
#    slightly negative, which would otherwise denormalise to 0 and leave you
#    with a background value the format does not use.
blob = x[0, 0].cpu().numpy()
blob = np.clip(blob * (3100 - 1) + 1, 1, 3100).astype(np.float32)

# 6. enforce the value gap the data format requires (see below)
gap = (blob > 1) & (blob < 300)
blob[gap] = np.where(blob[gap] < 150, 1, 300)

blob = np.round(blob).astype(np.uint16)   # 64^3 synthetic cloud cell
mask = blob >= 300                        # binary cloud mask
```

## Understanding the output

You need these conventions to interpret a generated volume:

| Property | Value |
|---|---|
| Shape / dtype | `64 × 64 × 64`, `uint16` |
| Background | `1` |
| Cloud voxels | `300` – `3022` (vertical velocity, scaled units) |
| Value gap | nothing lies strictly between `1` and `300` |
| Voxel size | 0.5 km isotropic (so 20 voxels = 10 km) |
| Normalisation | `(raw - 1) / (3100 - 1)`, clipped to `[0, 1]` |

The model operates in the normalised `[0, 1]` space, so step 5 above is required
to return to raw units, and step 6 restores the value gap that real volumes have.

## Post-processing — extracting a single cloud cell

Steps 5 and 6 of the snippet give you a valid volume, but a raw sample often
contains several disconnected pieces (or none). To get a usable single cloud
cell you need to threshold, find connected components, keep the largest, and
blank the rest. Requires `scipy`:

```python
import numpy as np
from scipy import ndimage

def extract_cloud(blob, min_voxels=6, max_voxels=None):
    """Keep only the largest 26-connected cloud component; blank everything else.

    Returns (cleaned, mask, size).
      cleaned : uint16 volume containing at most one cloud component
      mask    : boolean cloud mask for `cleaned`
      size    : voxel count of the kept component; 0 means "discard this sample"
    """
    mask = blob >= 300
    if not mask.any():
        return blob.copy(), mask, 0                  # empty sample

    # 26-connectivity: faces, edges AND corners count as connected
    labels, n = ndimage.label(mask, structure=np.ones((3, 3, 3)))
    sizes = ndimage.sum(mask, labels, range(1, n + 1))
    keep = int(np.argmax(sizes)) + 1
    largest = int(sizes.max())

    cleaned = blob.copy()
    cleaned[labels != keep] = 1                      # fragments -> background

    if largest < min_voxels:
        return cleaned, cleaned >= 300, 0            # noise only -> discard
    if max_voxels is not None and largest > max_voxels:
        return cleaned, cleaned >= 300, largest      # implausibly large -> review
    return cleaned, cleaned >= 300, largest


cleaned, mask, size = extract_cloud(blob)
if size == 0:
    print("empty or noise-only sample -- generate another")
else:
    print(f"single cloud cell, {size} voxels "
          f"({size * 0.5**3:.1f} km^3 at 0.5 km voxels)")
```

`min_voxels` rejects noise-only samples and `max_voxels` (optional) flags blobs
that fill an implausible share of the volume. Both are yours to choose — pick
them from the size range your application needs. Expect to discard a large
fraction of samples; see the note below.

## Practical notes before you use this

- **Sampling is variable and low-yield.** Only a minority of raw samples are
  usable single cloud cells; the rest come out empty, as noise fragments, or
  occasionally as a blob filling the whole volume. **Plan on generating a surplus
  and filtering** with `extract_cloud()` above, discarding samples where it
  returns `size == 0`. This is the intended workflow, not a sign of a defect.
- **The size distribution is skewed.** Generated cells tend to be smaller and
  weaker than those in the source data, so a raw synthetic set is not a
  distribution-matched substitute for real data.
- **One source dataset.** Behaviour outside that simulation regime is unknown.
- **Cost.** Each sample takes 1000 reverse diffusion steps. A GPU is strongly
  recommended.

## Intended use and limitations

Intended for **research and educational use** — synthetic data augmentation and
generative-modelling experiments on volumetric atmospheric data.

No perceptual or physical-fidelity metrics are published for this model, so **no
fidelity or accuracy claims are made**. Do not use it for operational
forecasting, hazard assessment, or any decision-making application.

## Data and credit

Trained on convective-cloud cells extracted from atmospheric-simulation output.
**The dataset is not included or redistributed here.**

The underlying simulations come from **Dr. Xiaowen Li's lab at Morgan State
University**; please credit that lab as the source of the simulation data if you
use this model. The diffusion model in this repository was developed by
Mahmudul Hoque.

## License

**CC BY-NC 4.0** — [Creative Commons Attribution-NonCommercial 4.0
International](https://creativecommons.org/licenses/by-nc/4.0/).

You may use, share, and adapt these weights for **non-commercial** purposes with
attribution. Commercial use is not permitted. Please also credit the source of
the simulation data as described above.

## References

- Ho, Jain & Abbeel, *Denoising Diffusion Probabilistic Models*, NeurIPS 2020
  ([arXiv:2006.11239](https://arxiv.org/abs/2006.11239))
- [MONAI GenerativeModels](https://github.com/Project-MONAI/GenerativeModels)

## Contact

Mahmudul Hoque, Morgan State University — <mahoq1@morgan.edu>