VidFM3D: How Much 3D Do Video Foundation Models Encode?

Installation

# create conda environment
conda create -n vidfm3d python=3.11 cmake=3.14.0 -y
conda activate vidfm3d

# install PyTorch (adjust versions according to your system)
conda install pytorch torchvision torchaudio pytorch-cuda=12.4 nvidia/label/cuda-12.4.0::cuda-toolkit -c pytorch -c nvidia

# install PyTorch3D from source (the compilation will take a while)
# export MAX_JOBS=6 # un-comment this if your machine is low on RAM (e.g., 16GB) when compiling PyTorch3D
pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable" --no-build-isolation
# unset MAX_JOBS

# install requirements
pip install -r requirements.txt

# install vidfm3d as a package (so you can import vidfm3d and use it in your own project)
pip install -e .

export CUDA_HOME=/usr/local/cuda-12.4
pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable"

conda install "mkl<2024.1" "intel-openmp<2024.1" -c conda-forge -y

Dataset Preparation

CO3Dv2

1. Download the raw CO3Dv2 dataset following the official instructions. Place (or symlink) the downloaded categories under vidfm3d/data/CO3D/CO3D-data/.

2. Extract frames: crops each sequence to a fixed aspect ratio (16:9 by default), subsamples to 81 frames, and rejects poorly-cropped sequences:

python -m vidfm3d.data.processing.co3d.extract_frames \
    --raw_root vidfm3d/data/CO3D/CO3D-data \
    --out_root vidfm3d/data/CO3D/CO3D-raw \
    --stride 1 --num_frames 81 \
    --trunc_thresh 0.25 --resize_to 960 540

3. Extract ground-truth point maps using VGGT:

python -m vidfm3d.data.processing.process_co3d --root vidfm3d/data/CO3D

DL3DV

1. Download DL3DV-10K from HuggingFace and place (or symlink) it under vidfm3d/data/DL3DV/DL3DV-10K/. The dataset already comes with extracted frames, no additional frame extraction step is needed.

2. Extract ground-truth point maps using VGGT:

python -m vidfm3d.data.processing.process_dl3dv --root vidfm3d/data/DL3DV

Feature Extraction

Use the following command to extract features (using WAN2.1-1.3B as example):

python -m features.run_co3d \
  --vfm wan \
  --subset all \
  --model-id Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
  --prompt "" \
  --output-layers 20 \
  --t 749

python -m features.run_dl3dv \
  --vfm wan \
  --subset all \
  --model-id Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
  --prompt "" \
  --output-layers 20 \
  --t 749

Different VFMs may require different arguments; see the top of features/run_co3d.py or features/run_dl3dv.py for per-model examples.

Training

Train model with chosen experiment configuration from configs/experiment/. For example, to train probe for WAN2.1-1.3B on both datasets:

python vidfm3d/train.py experiment=co3d/wan job_name=wan
python vidfm3d/train.py experiment=dl3dv/wan job_name=wan

You can override any parameter from command line following Hydra override syntax. All the logging can be found on wandb.

Evaluation & Visualization

Quantitative Metrics

Export final metrics from local W&B logs to CSV:

python scripts/parse_results.py \
  --groups dl3dv,co3d \
  --runs wan \
  --metrics "val/Auc_30,val/pmap_mse_aligned,val/loss_depth"

--metrics supports wildcard patterns (e.g. val/*). Per-run CSVs are written to logs/metrics/<group>/<run>.csv and a joint CSV to logs/metrics/<group>/joint.csv.

Qualitative Visualization

The training val loop skips saving viser artifacts. To generate them, run the test loop explicitly:

python vidfm3d/train.py experiment=co3d/wan task_name=co3d-eval job_name=wan train=false test=true ckpt_path=/path/to/checkpoints/last.ckpt

Results are saved to logs/{task_name}/runs/{task_name}_{job_name}/ (e.g. logs/co3d-eval/runs/co3d-eval_wan/), with two subdirectories:

• viser_viz/ — interactive point cloud and camera visualization
• viz/val/epoch_0/ — depth maps, confidence maps, and images

To view a 3D scene, port-forward and run:

python scripts/viser_view.py --scene logs/co3d-eval/runs/co3d-eval_wan/viser_viz/batch_0_sample_0_pred
python scripts/viser_view.py --scene logs/co3d-eval/runs/co3d-eval_wan/viser_viz/batch_0_sample_0_gt

For apples-to-apples comparison across methods, copy the pose string from one viser session and paste it into others to lock the viewpoint.

Figure Plotting

The radar plot compares multiple methods at once: run parse_results.py with all desired methods in --runs first so the joint CSV contains one row per method. Then:

python scripts/plot_radar.py \
  --csv logs/metrics/co3d/joint.csv \
  --out results/radar_co3d.png

Citation

@article{huang2025vidfm3d,
  title   = {How Much 3D Do Video Foundation Models Encode?},
  author  = {Huang, Zixuan and Li, Xiang and Lv, Zhaoyang and Rehg, James M.},
  booktitle = {arXiv preprint arXiv:2512.19949},
  year    = {2025}
}

Acknowledgments & Licenses

This project builds on code from the following awesome repositories, which retain their original licenses:

• Fast3R: training & data infrastructure.
• VGGT: architecture and data processing.

Foundation models evaluated in this work, each under its own license:

• WAN 2.1
• CogVideoX
• Open-Sora
• Aether
• V-JEPA
• DINOv2
• Fast3R

Please refer to each repository for its full license and terms of use.