PhyCoVE

PhyCoVE is a simulation-based benchmark for physics-aware video editing, with 90 scenes, 90 source videos, and 508 editing tasks. Given a source video and an instruction that changes physical properties or object presence, a model generates an edited video. Each task includes a reference edited video, instructions in English and Chinese, and simulation-derived ground-truth annotations.

The benchmark covers rigid, elastic, fluid, and mixed interactions. Examples include collisions, falling objects, cloth deformation, liquid pouring, and buoyancy. The videos of this main set are rendered from simulated scenes rather than captured from the real world.

The release also contains PhyCoVE-Real, a small real-recorded subset with 20 scenes, 20 source videos, and 25 editing tasks. It is stored under real/ and is reported separately from the main set (see PhyCoVE-Real).

Release contents	Count / format
Scenes	90
Source videos	90
Editing tasks / reference edited videos	508
Unique videos	598
Ground-truth files	598, one per source or edited case
Video format	MP4 / H.264, 1280 × 720, 24 fps
Clip length	4 seconds / 96 frames
Instruction variants per task	English / Chinese × vague / quantitative
Evaluation split	test — 508 paired-video tasks

Latest update — 2026-11-13

• Added 70 DELETE editing tasks across 59 existing scenes, bringing the benchmark to 508 editing tasks and 598 unique videos. All new tasks remove an object from the start of the clip and include reference videos, bilingual prompts, and ground-truth annotations.
• Updated affected prompts to identify objects by visible appearance, such as the star-patterned toy ball, white toy chest, and purple toy sports car.
• Added the PhyCoVE-Real subset under real/: 20 real-recorded scenes and 25 editing tasks, loaded with the real configuration.
• Re-rendered the source and five reference edited videos for metro_security_bag_curtain to fix the flashing scanner hood. Use the current videos and their updated video_sha256 values when refreshing a cached copy.

Task

For each editing task:

1. Read the source video and select one of the four instruction variants.
2. Generate an edited video that follows the requested physical intervention.
3. Compare the prediction against the reference edited video and, where appropriate, the accompanying ground truth.

Reference edited videos and their ground-truth files are evaluation targets. They should not be supplied as inputs to the editing model.

Edits use three operations:

Operation	Meaning	Tasks
SET	Change a physical parameter or initial condition	333
DELETE	Remove an object or material	157
ADD	Introduce an object	18
Total		508

There are 445 whole-clip edits and 63 edits that begin partway through a clip. The manifest records each intervention as an edit DSL statement, such as:

SET apple.mass TIMES 1.5
SET water.viscosity TIMES 10000
DELETE water AT FRAME 49

applies_from_frame is 1-based: frame 1 begins at 0 seconds; frame 49 begins at 2 seconds in a 24 fps clip. For an intervention starting at frame f, the intended application time is (f - 1) / 24 seconds. Frames preceding the intervention are intended to retain the source behavior.

Scene coverage

Interaction category	Manifest value	Scenes	Editing tasks
Rigid–rigid	rigid_rigid	40	236
Rigid–elastic	rigid_elastic	28	160
Rigid–fluid	rigid_fluid	16	80
Elastic–fluid	elastic_fluid	4	21
Elastic–elastic	elastic_elastic	2	11
Total		90	508

Properties include mass, friction, restitution, initial velocity, gravity, density, viscosity, stiffness, and scene-specific physical parameters. The exact parameter names and counts are in benchmark_manifest.json["counts"]["by_property"]; use each task's physics_diff and edit_dsl to interpret its intervention.

Download and use

Download the complete release

pip install -U huggingface_hub

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="ccmoony/PhyCoVE",
    repo_type="dataset",
    local_dir="PhyCoVE",
)

For a reproducible experiment, pass a specific Hub commit hash as revision and record it with the results.

Read a task with the authoritative manifest

This route uses only Python's standard library after the files have been downloaded:

import json
from pathlib import Path

root = Path("PhyCoVE")
benchmark = json.loads((root / "benchmark_manifest.json").read_text(encoding="utf-8"))
task = benchmark["edits"][0]

source_video = root / task["source_video"]
instruction = task["prompts"]["quantitative"]["en"]
reference_video = root / task["edited_video"]
ground_truth = json.loads((root / task["ground_truth"]).read_text(encoding="utf-8"))

print(task["global_id"])
print(instruction)
print(source_video, reference_video)

Load paired videos with Hugging Face Datasets

Use datasets>=4.0 for the paired-video metadata convention. The default configuration and Dataset Viewer read the 508 English-only editing-task rows in metadata_viewer.jsonl; shared source videos are referenced by those rows and do not create additional tasks.

pip install -U "datasets>=4.0"

from datasets import Video, load_dataset

tasks = load_dataset("ccmoony/PhyCoVE", split="test")

# Inspect paths and metadata without requiring a video decoder.
tasks = tasks.cast_column("source", Video(decode=False))
tasks = tasks.cast_column("video", Video(decode=False))
example = tasks[0]

print(example["id"])
print(example["prompt_quantitative_en"])
print(example["source"])  # input source video
print(example["video"])   # reference edited video

Automatic video decoding requires the decoder dependencies supported by your installed datasets version. The manifest-based loading example above works independently of those dependencies.

For a local copy, select the metadata and videos explicitly. This also prevents scene names containing words such as train from affecting automatic split inference:

tasks = load_dataset(
    "videofolder",
    data_files={"test": ["PhyCoVE/metadata.jsonl", "PhyCoVE/scenes/*/cases/*/video.mp4"]},
    split="test",
    drop_labels=True,
)

File layout

PhyCoVE/
├── README.md
├── benchmark_manifest.json
├── metadata.jsonl
├── metadata_viewer.jsonl        # English-only Dataset Viewer index
├── assets.json
├── real/                       # PhyCoVE-Real subset, same layout with its own indexes
├── scenes/
│   └── <scene>/cases/<case_id>/
│       ├── video.mp4
│       ├── prompts.json
│       └── ground_truth_transforms.json
└── splits/
    ├── all.txt
    ├── by_category/*.txt
    ├── by_kind/*.txt
    ├── by_property/*.txt
    └── by_timing/*.txt

All paths in the indexes are relative to the dataset root. Source videos are stored once per scene and reused by its editing tasks. Source case IDs are recorded in sources[].case_id; read them from the manifest rather than assuming a directory name.

Manifest fields

benchmark_manifest.json is the authoritative index. Its sources list contains the 90 source records; edits contains the 508 evaluation tasks. scene_metadata provides scene titles and interaction categories, and counts summarizes the release.

Editing-task field	Description
global_id	Unique task identifier, <scene>/<case_id>
scene, case_id, interaction_category	Scene, case, and physical interaction category
source_case_id, source_video	Input source case and its video path
edited_video	Reference edited video path
ground_truth	Reference case's ground-truth JSON path
video_sha256	SHA-256 of the reference edited video
prompts	Nested vague / quantitative × zh / en instruction strings
edit_dsl, edit_kind, object_id, property	Requested intervention; property may be null for presence edits
physics_diff, edit_summary, outcome_type	Existing parameter differences and descriptive metadata; inspect availability per task
applies_from_frame, timing	First affected frame and whole_clip / partway label
original_validation_status	Preserved authoring status for the original scene/case

metadata.jsonl is the full bilingual flat index, with one row per edit. metadata_viewer.jsonl is its English-only projection used by the default Hub configuration and Dataset Viewer; it omits Chinese prompt fields and scene titles and translates the two Chinese DSL display hints into English. All bilingual prompts remain available in metadata.jsonl, the manifest, and the per-case prompt files. Its id corresponds to manifest global_id, file_name points to the reference video, and source_file_name points to the source. VideoFolder exposes the video paths as video and source video columns. Instructions are flattened into prompt_vague_zh, prompt_vague_en, prompt_quantitative_zh, and prompt_quantitative_en; source_prompt_* fields describe the source clip.

Prompts

Source prompts describe the unedited scene. Editing prompts describe the intervention:

• Quantitative: names the requested numerical change and application frame.
• Vague: describes the requested change and its timing in qualitative terms.
• English and Chinese: both variants are included for every source and edit.

Prompt wording is preserved from the authoring suites. The release conversion does not rewrite the instructions.

Ground truth

Every case includes a ground_truth_transforms.json with the pcve_ground_truth/1 format and 96 frame records. Annotations include scene/case identity, timing, camera information, and available object transforms and presence states. Additional physical quantities and provenance vary with the scene and solver; inspect the actual JSON before using a field in a metric. These annotations do not represent a universal per-particle fluid or deformable-body state.

Selection indexes

splits/all.txt lists all 508 task IDs. The files in by_category, by_kind, by_property, and by_timing select tasks along different axes; they are not independent train/validation/test partitions. A task appears in one file within each axis and in selections across the other axes.

PhyCoVE-Real

real/ holds a real-recorded subset in the same layout as the main set, with its own benchmark_manifest.json, metadata.jsonl, metadata_viewer.jsonl, and splits/. It has 20 scenes, 20 source videos, and 25 editing tasks (21 DELETE, 3 ADD, 1 SET). The videos use the main set's format: MP4 / H.264, 1280 × 720, 24 fps, 96 frames.

• Pairs. Each source video and its reference edited video are separate recordings of the same setup, aligned on event timing. Trajectories, poses, and reflections differ between takes, so a pair is not a pixel-exact counterfactual. Trimming, cropping, frame sampling, and the exposure normalisation applied to 8 clips to remove camera auto-exposure jumps are recorded in real/editing_provenance.json.
• Ground truth. The files record frame timing, the mapping to original frames, object presence, and manually annotated events. 3D transforms, camera calibration, and physical parameters were not measured and are null, so trajectory-based physics metrics do not apply. Report video metrics for this subset separately from the main set.
• Edit DSL. Besides the main set's statements, three forms describe recorded edits: DELETE a; DELETE b removes several objects together (all are listed in object_ids), ADD x or ADD x AT <region> adds an object whose position is described only in the prompts, and SET x.p INCREASE / DECREASE gives the direction of an unmeasured change.

real_tasks = load_dataset("ccmoony/PhyCoVE", "real", split="test")

Evaluation and reproduction scope

This repository distributes benchmark data for model evaluation: source and reference videos, bilingual prompts, per-case annotations, indexes, and asset provenance. assets.json records asset sources, licenses, and checksums; it does not contain the 3D assets themselves.

Simulation caches, Blender scene files, solver environments, historical experiments, logs, intermediate render frames, and the scene-generation/evaluation code are not included. Rebuilding the simulated scenes requires the separate authoring code, its runtime dependencies, and the original or generated scene inputs. Downloading this dataset alone does not reproduce the simulations from scratch.

For experiments, report the dataset revision, task selection, prompt language/flavor, number of successfully generated videos, and any temporal or spatial preprocessing. Treat missing predictions and unavailable annotation fields explicitly. Perceptual similarity alone does not establish that a requested physical change was followed.

The authoring validation labels are preserved, including pending-review or unrecorded states. The main set is synthetic, and performance on it may not transfer directly to real-world footage.

License and attribution

PhyCoVE is released under the MIT License. Third-party asset attributions and their original license declarations are retained in assets.json; the MIT license does not replace those third-party terms.

No separate paper citation has been supplied for this release. When referencing it, identify PhyCoVE, its Hugging Face repository, and the revision used. Dataset questions and reproducibility issues can be reported in the repository's Community tab.