---
license: cc-by-4.0
library_name: pytorch
datasets:
  - MedPhyGraph/support-graph-data
tags:
  - digital-twins
  - scene-graphs
  - dynamic-scene-graphs
  - robotics
  - support-relations
  - graph-maintenance
  - counterfactual-evidence
  - pytorch
---

# CF-SupportNet

CF-SupportNet is the learned edge-scoring component of **MedPhyGraph**  
(*TwinWorld: Visual Intelligence for Built Environment Digital Twins, ECCV 2026 Workshop*).

It is **not** the full MedPhyGraph framework by itself. MedPhyGraph also applies
deterministic **State Consistency** and **Union-Based Transition-Aware
Consistency** on top of CF-SupportNet scores to maintain support relations
across adjacent digital-twin states.

| Resource | Link |
|---|---|
| Model | **You are here** |
| Dataset | [MedPhyGraph/support-graph-data](https://huggingface.co/datasets/MedPhyGraph/support-graph-data) |
| Project | [medphygraph.github.io](https://medphygraph.github.io/) |

## Overview

CF-SupportNet scores candidate `SupportedBy` edges from structured scene
information.

The scorer combines:

- static geometric features
- candidate support-edge features
- analytic host-removal counterfactual evidence

The counterfactual signal is computed using an **analytic AABB-based geometric
proxy**. It should not be interpreted as a full rigid-body physics simulation.

CF-SupportNet operates downstream of scene perception. Rendered RGB images are
**not** model inputs and are **not** used as label sources.

## Architecture

The released CF-SupportNet checkpoints use the paper-frozen configuration:

| Property | Value |
|---|---|
| Framework | PyTorch |
| Architecture | GRU + MLP |
| Hidden size | 64 |
| Trainable parameters | 25,409 |
| Counterfactual weight `rho` | 1.0 |

## Paper-frozen checkpoints

This repository contains five released checkpoints:

| Checkpoint | Role |
|---|---|
| `health_dyphygraph_r1.0_seed0.pt` | Primary paper checkpoint |
| `health_dyphygraph_r1.0_seed1.pt` | Multi-seed checkpoint |
| `health_dyphygraph_r1.0_seed2.pt` | Multi-seed checkpoint |
| `health_dyphygraph_r1.0_seed3.pt` | Multi-seed checkpoint |
| `health_dyphygraph_r1.0_seed4.pt` | Multi-seed checkpoint |

**Seed 0** is the primary paper checkpoint.  
**Seeds 1–4** are provided for the paper's multi-seed evaluation.

The checkpoint files are paper-frozen artifacts and should not be modified when
reproducing the released results.

## Download

Download the complete model release with the Hugging Face CLI:

```bash
hf download MedPhyGraph/CF-SupportNet \
  --local-dir ./CF-SupportNet
```

Or with Python:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="MedPhyGraph/CF-SupportNet",
    local_dir="./CF-SupportNet",
)
```

To download only the primary checkpoint:

```python
from huggingface_hub import hf_hub_download

checkpoint_path = hf_hub_download(
    repo_id="MedPhyGraph/CF-SupportNet",
    filename="health_dyphygraph_r1.0_seed0.pt",
)

print(checkpoint_path)
```

## Relationship to MedPhyGraph

CF-SupportNet provides learned scores for candidate support edges.

MedPhyGraph then applies deterministic graph-maintenance components, including:

1. **State Consistency**
2. **Union-Based Transition-Aware Consistency**

These components are non-learned inference operations and are not encoded as
separate model checkpoints.

In other words:

```text
structured scene state
        │
        ▼
candidate support edges
        │
        ▼
CF-SupportNet
(learned edge scores)
        │
        ▼
State Consistency
        │
        ▼
Union-Based Transition-Aware Consistency
        │
        ▼
maintained SupportedBy graph
```

## Dataset

The corresponding public procedural training and evaluation data are available
at:

[**MedPhyGraph/support-graph-data**](https://huggingface.co/datasets/MedPhyGraph/support-graph-data)

The dataset release contains:

- the frozen candidate-edge dataset and split
- structured Procedural scene states
- the 136-case Procedural subset of the expanded transfer evaluation

NVIDIA Isaac for Healthcare assets and Isaac-derived structured scene states
are not redistributed in that repository.

## Scope

CF-SupportNet is intended to reproduce and study the learned scoring component
used in MedPhyGraph.

The model does not directly provide:

- scene perception
- RGB/image understanding
- object detection
- complete physical simulation
- the deterministic MedPhyGraph consistency modules
- NVIDIA Isaac for Healthcare assets or scenes

The complete MedPhyGraph method combines the learned scorer with deterministic
graph-maintenance logic.

## Integrity

The repository includes:

- **`checkpoint_manifest.json`** — metadata for the released checkpoints
- **`SHA256SUMS.txt`** — SHA256 hashes for checkpoint integrity verification

The primary seed-0 checkpoint has SHA256:

```text
e0b34529745399ecc5da5341ed7a162173611e12c8bd50dec121b0c575c5b789
```

## License

The released CF-SupportNet model weights are provided under the
**Creative Commons Attribution 4.0 International License (CC BY 4.0)**.

Please provide appropriate attribution when using or redistributing these
weights.

The license for the MedPhyGraph source code is separate from the model-weight
license.

## Citation

If you use CF-SupportNet, MedPhyGraph, or these released checkpoints, please
cite:

```bibtex
@inproceedings{gholizadeh2026medphygraph,
  title     = {MedPhyGraph: Counterfactual Support-Graph Maintenance for Dynamic Built-Environment Digital Twins},
  author    = {Gholizadeh HamlAbadi, Kamran and Vahdati, Monica and El Saddik, Abdulmotaleb},
  booktitle = {TwinWorld: Visual Intelligence for Built Environment Digital Twins, ECCV 2026 Workshop},
  year      = {2026}
}
```