This dataset is released for non-commercial academic research only (CC BY-NC 4.0). Access is reviewed manually. Please tell us who you are and how you intend to use the data. By requesting access you agree not to redistribute the videos, not to use them for commercial purposes, and to cite the associated publication in any work derived from them.

video-phys

Paired simulated and real video data for training physics-aware video generation models. The simulated releases ship per-frame velocity videos (vel_videos/) alongside the RGB videos, plus wind masks for a subset.

Mirror of http://vcv.stanford.edu/viscam/downloads/video_phys/. The archives here are the original, unmodified ones from that server; extracting them reproduces the original directory layout exactly, so the extracted root can be used directly as --dataset_base_path.

File	Size	Extracts to
`REAL_data_0320.zip.part00` … `part02`	20.5 GB total	`REAL_data_0320/` — 19,561 mp4
`wan_video.zip`	2.1 GB	`wan_video/` — 3,149 mp4
`example_data_0219.zip.part00` … `part04`	39.1 GB total	`example_data_0219/` — 3,280 mp4 + 3,280 vel
`example_data_0320.zip.part00` … `part10`	86.6 GB total	`example_data_0320/` — 6,465 mp4 + 6,465 vel + 1,634 wind masks
`example_data_0413.zip`	7.7 GB	`example_data_0413/` — 3,457 mp4 + 3,457 vel
`metadata_real_train_combined.csv`	18 MB	real training set, 22,091 rows
`metadata_sim_train_combined.csv`	12 MB	sim training set, 13,125 rows
`REAL_data_0320/metadata_filtered_test.csv`	18 KB	not present inside the zip — see below
`REAL_data_0320/metadata_filtered_train.csv`	15 MB	not present inside the zip — see below
`CHECKSUMS.txt`	—	sha256 of each full archive, for verifying a rejoin

Total ≈ 156 GB.

The three large archives are uploaded as ~8 GB slices. Concatenating the slices of an archive in order reproduces the original file bit-for-bit — verify with CHECKSUMS.txt. The slicing is purely a transport detail: Hugging Face caps single files at 50 GB, and uploads of multi-tens-of-GB blobs proved unreliable, whereas 8 GB slices upload cleanly.

The two loose CSVs

REAL_data_0320/metadata_filtered_test.csv and metadata_filtered_train.csv were added to the source server about a month after REAL_data_0320.zip was built, so they are not inside the archive. They are shipped here as loose files at the same relative path the archive would place them, so extracting the zip and keeping these two files side by side gives you the complete directory. metadata_filtered_test.csv is the validation split referenced by --val_real_dataset_metadata_path, so training will fail without it.

Download and extract

Access is gated with manual review — request access on the dataset page, then:

SNAP=/path/to/video-phys

hf auth login
hf download yzheng18/video-phys --repo-type dataset --local-dir "$SNAP"
cd "$SNAP"

# rejoin the sliced archives (slice order matters; the glob sorts correctly)
for z in REAL_data_0320 example_data_0219 example_data_0320; do
    cat "$z".zip.part* > "$z.zip" && rm "$z".zip.part*
done

# verify the rejoins before spending time on extraction
shasum -a 256 -c CHECKSUMS.txt

for z in REAL_data_0320 wan_video example_data_0219 example_data_0320 example_data_0413; do
    unzip -q "$z.zip" && rm "$z.zip"
done

If shasum -c reports anything other than OK for an archive, re-download the slices for that archive rather than trying to extract it — a partial slice produces a corrupt zip whose failure mode is confusing.

You need roughly 156 GB for the archives plus 156 GB for the extracted data; the loop above deletes each archive right after extracting it to keep the peak lower. Extraction restores REAL_data_0320/metadata_filtered_*.csv untouched, because those two files are not in the archive.

To train on only the real data, you can skip the example_data_* archives:

hf download yzheng18/video-phys --repo-type dataset --local-dir "$SNAP" \
    --include "REAL_data_0320*" "wan_video.zip" "metadata_real_train_combined.csv"

The two combined training CSVs

Both have a video column holding paths relative to the extracted root, and each one spans several releases:

CSV	prefix	rows
`metadata_real_train_combined.csv`	`REAL_data_0320/`	18,942
	`wan_video/`	3,149
`metadata_sim_train_combined.csv`	`example_data_0219/`	3,280
	`example_data_0320/`	6,439
	`example_data_0413/`	3,406

So training on either combined CSV needs several archives, not just one — metadata_real_train_combined.csv needs both REAL_data_0320.zip and wan_video.zip, and metadata_sim_train_combined.csv needs all three example_data_* archives. With everything extracted, all 22,091 real and 13,125 sim rows resolve, including all 13,125 video_vel and 1,629 mask_wind references.

⚠️ Filename collisions across releases

example_data_0219, example_data_0320 and example_data_0413 each number their videos from 00000.mp4, but identically-numbered files are different clips. For instance example_data_0219/videos/00000.mp4 (a black smoke explosion) and example_data_0320/videos/00000.mp4 (a brown bear on a wooden platform) are unrelated, with different checksums and different prompts. Always keep the directory prefix from the video column, and never merge these directories.

Columns

metadata_real_train_combined.csv: video, prompt, original_video_path, category, run_id, model, seed

metadata_sim_train_combined.csv: video, prompt, video_vel, mask_wind, wind_direction, wind_normalized_strength, wind_prompt, physical_prompt_coarse, physical_prompt_detailed, dissipation, turbulence, confinement, disturbance, buoyancy, fire_scale, has_ball, turbulence_scale, buoyancy_scale, disturbance_scale, divergence_scale, density_scale, temperature_scale

video_vel is populated for all 13,125 sim rows; mask_wind and wind_direction for 1,629 and 1,627 rows.

Use with DiffSynth Wan2.1 training

The extracted layout matches the original server paths, so point the base paths at the extracted root:

SNAP=/path/to/video-phys

accelerate launch --config_file examples/wanvideo/model_training/full/accelerate_config_14B.yaml \
  examples/wanvideo/model_training/train.py \
  --dataset_base_path            $SNAP/ \
  --dataset_metadata_path        $SNAP/metadata_sim_train_combined.csv \
  --real_dataset_base_path       $SNAP/ \
  --real_dataset_metadata_path   $SNAP/metadata_real_train_combined.csv \
  --val_dataset_base_path        $SNAP/example_data_0320/ \
  --val_dataset_metadata_path    $SNAP/example_data_0320/metadata_new_test.csv \
  --val_real_dataset_base_path   $SNAP/REAL_data_0320/ \
  --val_real_dataset_metadata_path $SNAP/REAL_data_0320/metadata_filtered_test.csv \
  --height 480 --width 832 \
  --extra_inputs "video_vel,end_image" \
  --data_file_keys "image,video,video_vel" \
  --model_id_with_origin_paths "Wan-AI/Wan2.1-I2V-14B-480P:diffusion_pytorch_model*.safetensors,Wan-AI/Wan2.1-I2V-14B-480P:models_t5_umt5-xxl-enc-bf16.pth,Wan-AI/Wan2.1-I2V-14B-480P:Wan2.1_VAE.pth,Wan-AI/Wan2.1-I2V-14B-480P:models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth" \
  --trainable_models "dit" --remove_prefix_in_ckpt "pipe.dit." \
  --initialize_model_on_cpu

License

CC BY-NC 4.0 — non-commercial research use with attribution.

Portions of the real footage derive from third-party sources (see original_video_path in metadata_real_train_combined.csv, which references FluidNexus data); those retain their original terms.

Citation

@misc{video-phys,
  title  = {video-phys: paired simulated and real video data for physics-aware video generation},
  author = {Zheng, Yang and others},
  year   = {2026},
  note   = {https://huggingface.co/datasets/yzheng18/video-phys}
}

Downloads last month: 64

Total file size:

156 GB

You need to agree to share your contact information to access this dataset