Code2Video-Compass: A Comprehensive Benchmark for General-Purpose Code-Based Video Generation across T2V, I2V, and V2V

Zhenghan Chen, Zekai Shao, Xin Lin, Leting Cheng, and Siming Chen

The official dataset repository of Code2Video-Compass

Overview  |  Task formats  |  Download and use  |  Evaluation  |  Data fields

1. Overview

_{The benchmark evaluates 15 multimodal large language models (MLLMs) on code-based text-to-video (T2V), image-to-video (I2V), and video-to-video (V2V) generation through a dual-level framework combining 12 objective metrics with five subjective metrics.}

Code2Video-Compass evaluates whether multimodal models can translate detailed visual and temporal requirements into executable programs that render complete videos. It assesses instruction following, visual quality, temporal coherence, and reference fidelity across diverse creative tasks.

Code2Video-Compass benchmarks 15 multimodal large language models (MLLMs) on 2,104 code-based video generation tasks across text-to-video (T2V), image-to-video (I2V), and video-to-video (V2V) settings. The benchmark spans 10 creative categories and employs 12 objective and five subjective metrics.

The tasks are organized into two collections:

• Source-paired tasks: 573 curated source videos, each associated with three corresponding instructions and selected visual references.
• Independent tasks: 385 separately constructed tasks, comprising text instructions, image-conditioned tasks, and clip-conditioned tasks.

Collection	T2V	I2V	V2V	Total
Source-paired	573	573	573	1719
Independent	145	120	120	385
Total	718	693	693	2,104

Each task includes its instruction, structured annotations, and an evaluation checklist.

2. Task Formats

Setting	Model input	What the model needs to produce
T2V — Text to Video	A detailed text instruction	Code that renders the specified visual content and event progression
I2V — Image to Video	A text instruction and a reference keyframe	Code that incorporates the reference state at the instructed stage of a complete animation
V2V — Video to Video	A text instruction and a reference clip	Code that incorporates the clip's content and dynamics into the instructed full sequence

I2V images specify a reference state to incorporate at the instructed stage. V2V clips provide selected source intervals. The task instruction and reference summary specify their role in the output.

3. Data Construction and Coverage

The source-paired collection selects complete animations through keyword-based retrieval, screening, deduplication, quality assessment, and manual review. The independent collection constructs additional tasks from curated text, image, and video inputs.

_{Data construction pipeline. (A) Independent tasks are developed from curated text seeds, images, and video clips. (B) Paired tasks share a complete source video across T2V, I2V, and V2V. (C) Both sets undergo prompt construction, QA checklist development, and annotation, yielding 2,104 tasks.}

Dataset distributions

_{Dataset distributions over 958 records: 573 source-paired groups and 385 independent tasks. Families, motion types, and subject types can have multiple labels.}

_{Category-wise construction, dynamic animation, and state-dependency complexity, and prompt lengths across input settings.}

Creative coverage

Family	Scope
1. Explainer Motion Graphics	Concepts, mechanisms, processes, and systems
2. Branding & Promotional Motion	Brand identities, product promotion, and campaigns
3. Kinetic Typography	Text-driven motion and editorial typography
4. UI & Product Motion	Interface workflows, dashboards, and HUDs
5. Data & Infographic Animation	Quantitative stories, relationships, timelines, and geographic information
6. Title Sequences	Openings, titles, closings, and credits
7. Broadcast & Show Graphics	News, sports, entertainment, and program packaging
8. Illustrative & Character Narrative	Character stories and illustrated scene narratives
9. Abstract & Experimental Motion	Procedural geometry, particles, and formal experiments
10. 3D Motion	Objects, products, environments, and procedural or simulated 3D

The 24-subtype taxonomy and definitions are available in taxonomy.json.

Complexity annotations

Axis	Description	Labels
C — Construction complexity	Complexity of the visual structures and elements to be constructed	C1–C3
D — Dynamic Animation Complexity	Complexity of changes, motion coordination, and temporal organization	D1–D3
S — State dependency span	How long an established object or state constrains later stages	S1–S3

4. Download and Use

Each task setting is provided in Parquet, JSONL, and CSV formats with identical records.

Read the local release

Run these commands from the huggingface/ directory:

pip install datasets

from datasets import load_dataset

t2v = load_dataset("parquet", data_files="all/t2v/*.parquet", split="train")
i2v = load_dataset("parquet", data_files="all/i2v/*.parquet", split="train")
v2v = load_dataset("parquet", data_files="all/v2v/*.parquet", split="train")

print(t2v[0]["prompt"])
print(i2v[0]["reference_frame_summary"])

Read from Hugging Face

The dataset card defines three configs: t2v, i2v, and v2v. Load a setting directly from the dataset repository:

dataset = load_dataset("code2video-compass/Code2Video-Compass", "i2v", split="train")

The train/ and all/ folders contain identical complete task tables. Hugging Face loads each task config through the train split.

Reference media

Task tables provide prompts and, for conditioned tasks, reference URLs and timestamps. Retrieve the reference media from its listed source; documentation figures are in assets/. Timestamps are relative to the original source video: I2V uses keyframe_timestamp_s, and V2V uses keyclip_start_s and keyclip_end_s. Video timestamps apply to video-derived I2V references and V2V clips.

5. Pairing and File Organization

The HF repository is organized as follows:

README.md
assets/
  code2video-compass.svg
  code2video-compass.png
  overview.jpg
  distributions.jpg
  complexity.jpg
all/
  t2v/       # 718 tasks
  i2v/       # 693 tasks
  v2v/       # 693 tasks
train/       # Same records; HF-style loading copies
  t2v/
  i2v/
  v2v/
pair_index.csv
taxonomy.json
schema.json
manifest.json
data_issues.csv

Source-paired tasks share a group ID such as C2VC-0001; their task IDs end in -t2v, -i2v, and -v2v. Join corresponding tasks by sample_id. Independent tasks have separate IDs. pair_index.csv lists the settings available for each group.

6. Evaluation

The evaluation framework reports automatic measurements and model-assisted assessments:

Part	What is evaluated
Execution and efficiency	Whether code produces a video, and output-token consumption
Automatic video measurements	Technical quality, aesthetic quality, motion smoothness, and local/global text–video alignment
Reference fidelity	Keyframe perceptual/structural fidelity for I2V; clip sequence, appearance, and motion fidelity for V2V
Instruction following	Checklist QA against the requirements of the corresponding task
Holistic quality	NC: content coherence and integrity; VD: visual design and graphic richness; MQ: motion quality and richness; VE: viewer engagement

Model outputs and scores are published separately from the task tables.

7. Data Fields

The T2V, I2V, and V2V tables contain 14, 17, and 18 columns, respectively, with qa as the final column. Multi-valued annotations are stored as ordered lists of strings.

Shared text fields

Field	Description
sample_id	Group ID shared by corresponding source-paired tasks
task_id	Unique task ID, including its T2V/I2V/V2V suffix
task	Input setting: t2v, i2v, or v2v
prompt	Task-specific instruction provided to the model
global_description	Complete overall description of the video
video_categories	All video-family names, as a native list of complete strings
subtypes	All subtype names, as a native list of complete strings
subjects	Primary subject type labels as a native list, such as ["Graphic"]
motion_types	All motion types, as a native list
events	All visual event descriptions, as a native list in order
construction_complexity	C label (C1–C3)
dynamic_complexity	D label (D1–D3)
dependency_span	S label (S1–S3)

I2V reference fields

Field	Description
reference_url	Source media URL or source dataset item locator
reference_frame_summary	Complete original description of the reference image
keyframe_timestamp_s	Recorded keyframe time in seconds for video-derived images; null for standalone images

V2V reference fields

Field	Description
reference_url	Source media URL or source dataset item locator
reference_clip_summary	Complete original description of the reference clip
keyclip_start_s, keyclip_end_s	Recorded clip interval in seconds relative to the original source video used for this task

QA checklist — final column

qa contains the shared questions and the specific questions for this task only. It is an ordered list of records with section, dimension, name, question, and expected_answer. T2V includes FullVideoSpecific, I2V includes KeyframeSpecific, and V2V includes ReferenceClipSpecific, following the shared questions.

8. Release Notes

manifest.json records counts and data-file hashes; schema.json records the field types.

Reference media are linked through source URLs or dataset locators, subject to their availability and licensing terms.

Citation

If you find Code2Video-Compass useful, please consider citing our paper:

@misc{chen2026code2videocompasscomprehensivebenchmarkgeneralpurpose,
      title={Code2Video-Compass: A Comprehensive Benchmark for General-Purpose Code-Based Video Generation across T2V, I2V, and V2V},
      author={Zhenghan Chen and Zekai Shao and Xin Lin and Leting Cheng and Siming Chen},
      year={2026},
      eprint={2610.28416},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.28416}
}