Tsu7am1/Cinematic-DiT-Video-Dataset

Cinematic DiT Video Dataset

This dataset is publicly downloadable under a restricted research license. It is intended for non-commercial research on AI-generated video detection, media forensics, authenticity analysis, and content-safety evaluation.

Dataset Summary

This dataset contains 9,000 synthetic text-to-video samples generated from 3,000 Chinese cinematic prompts. Each source prompt has one video in each of three generation profiles. The prompt pipeline emphasizes photorealistic, documentary-like footage and suppresses painterly, CGI, overly stylized, and beautified output tendencies.

The dataset is intended for DiT-oriented generated-video detection research, controlled synthetic-data experiments, media forensics, and analysis of the gap between generated video and real-world visual dynamics. It is not presented as real camera footage.

Dataset Details

• Videos: 9,000 MP4 files
• Source prompts: 3,000
• Prompt organization: 300 scene categories x 10 shot designs
• Generation model: agnes-video-v2.0
• Generation interface: Agnes Video V2.0 API
• Language: Chinese scene descriptions with English photorealism anchors
• Video payload size: 57.945 GiB
• Repository package size: 30 uncompressed tar shards, approximately 58.003 GiB

The 10 shot designs combine camera platform, motion, framing, and capture style, for example handheld close tracking, drone overhead movement, static tripod framing, Steadicam orbiting, and high-speed camera shots.

The profile names are dataset labels rather than claims that the encoded files match broadcast-standard 1080p/720p/480p geometry. The Agnes API normalizes requested dimensions to backend-supported encoded sizes before returning the video. The exact normalization behavior is provider-defined, so public metadata records both requested values and ffprobe-measured encoded values. Researchers should use the actual metadata fields for preprocessing and avoid treating profile-specific aspect ratios as semantic signals.

Profile	Requested size	Actual encoded size	Frames	FPS	Duration
1080p	1920x1080	1920x1088	169	24	7.041667 s
720p	1152x768	1088x832	409	24	17.041667 s
480p	640x480	704x512	961	24	40.041667 s

The three videos associated with one prompt are independent generations. They are not spatially or temporally aligned versions of the same source video.

Repository Structure

README.md
LICENSE
metadata.parquet
metadata.csv
metadata.jsonl
prompts.csv
checksums.sha256
shard_index.json
validation_report.json
shards/
  train-00000.tar
  ...
  train-00029.tar

Each tar stores MP4 files under videos/<task_id>.mp4. The shard and file_name columns identify the containing tar and path inside that tar.

The Hugging Face Dataset Viewer is configured to preview metadata.parquet. The MP4 payloads are distributed through the tar shards listed in the metadata, because the Hub viewer may not reliably infer large WebDataset tar archives.

Metadata

Important columns include:

Field	Description
task_id	Unique video sample identifier
prompt_id	Unique source-prompt identifier, 0..2999
global_id	Unique source-prompt identifier, 0..2999
category	Chinese scene category
resolution	Requested generation profile
prompt_zh	Chinese prompt variant used for this video
enhanced_prompt	Photorealism-enhanced prompt submitted for generation
negative_prompt	Negative prompt submitted for generation
requested_width, requested_height	API request dimensions
actual_width, actual_height	Encoded MP4 dimensions from ffprobe
num_frames	Generation frame count for the profile
actual_frame_rate	Encoded frame rate from ffprobe
actual_duration_seconds	Encoded duration from ffprobe
file_size, sha256	Integrity metadata for the MP4
shard, file_name	Published tar and tar-internal video path

prompts.csv contains one row per source prompt and links each prompt to its three video task IDs. Its prompt_zh field contains the canonical prompt text for the prompt group, while prompt_variants_json records all source or repair variants observed for that group. Seven prompt groups contain two variants after repair. Per-video metadata preserves the prompt variant actually used for that video.

Generation and Repair Pipeline

The construction pipeline includes atomic task assignment, per-credential rate limiting, retry with backoff, resumable state files, failure classification, download verification, and post-run hole repair. Every prompt passes through a deterministic enhancement stage that removes style-triggering camera color terms and adds photorealistic positive and negative constraints.

After the main run, failed samples were audited by task ID. Ninety-nine failures were retried without prompt changes. Persistent content-policy failures were generated successfully with more neutral wording where needed. The resulting canonical prompt groups and their variants are integrated into prompts.csv; the per-video metadata preserves the actual generation prompt for reproducibility.

Validation

• 9,000 task records, completed states, and MP4 files
• No missing, extra, empty, or unfinished samples
• 9,000/9,000 MP4 files readable by ffprobe
• Exactly 3,000 videos for each generation profile
• Encoded frame rate is 24 FPS for every video
• Encoded frame counts are 169, 409, and 961 for the three profiles
• SHA-256 recorded for every video and every shard
• 18 review samples inspected across six prompts and all three profiles

Intended Uses

• Training, validating, and benchmarking AI-generated video detectors
• Studying synthetic-to-real distribution gaps for media authenticity research
• Analyzing temporal consistency, generation artifacts, and forensic signals
• Content-safety, provenance, and authenticity evaluation
• Reproducing fault-tolerant synthetic dataset construction workflows

Prohibited Uses

• Training, fine-tuning, distilling, or evaluating text-to-video, image-to-video, video-to-video, or other generative video models
• Commercial use or incorporation into commercial products or services
• Rehosting, mirroring, redistributing, sublicensing, or repackaging the videos or shards outside the official Hugging Face repository
• Creating, improving, or distributing deceptive, impersonating, or harmful synthetic media
• Presenting generated videos as real-world recordings or evidence of real events

Limitations and Risks

• Every video is machine-generated and may contain physical, semantic, temporal, anatomical, text-rendering, or camera-motion artifacts.
• Photorealistic prompting does not make the samples equivalent to real footage.
• All samples come from a single closed-source generation model. Detectors trained only on this dataset may overfit model-specific artifacts, encoder behavior, prompt style, or resolution-profile geometry, and may require domain adaptation before generalizing to other video generators.
• The prompt taxonomy is manually designed and does not represent all Chinese language use, cultures, locations, people, or filming styles.
• Generated people, brands, text, faces, and scenes may accidentally resemble real entities or protected material. Users should perform their own review for their intended jurisdiction and use case.
• Independent multi-resolution generations must not be treated as aligned video triplets for super-resolution evaluation.
• The dataset should not be used to misrepresent generated footage as authentic evidence or recordings of real events.

Access and License

The dataset is distributed under the accompanying Research and Content Safety Use License. This is a restricted research license, not an open-source or open-data license.

The files are publicly downloadable from Hugging Face, but use remains subject to the accompanying license. The maintainers may remove, restrict, or pause access when misuse is suspected, when required by platform policy, or when needed to respond to a rights-holder, service-provider, or legal request.

This license grants only the rights that the dataset maintainers can grant. It does not override third-party rights, platform terms, provider terms, or applicable law. Users are responsible for confirming that their intended use is allowed in their jurisdiction and under any terms that apply to their work.

Project Leadership and Contributions

Project lead and dataset creator: Tsu7am1

Responsibilities: project conception, prompt taxonomy, photorealistic prompt pipeline, worker architecture, generation operations, failure recovery, quality control, metadata design, release packaging, and documentation.

Citation

@dataset{tsu7am1_2026_cinematic_dit_video,
  author    = {Tsu7am1},
  title     = {Cinematic DiT Video Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  version   = {1.0.0},
  url       = {https://huggingface.co/datasets/Tsu7am1/Cinematic-DiT-Video-Dataset}
}

Related Resources

• Construction code: https://github.com/Ad3n1ne/Cinematic-DiT-Video-Dataset
• Stable release and DOI: not assigned yet
• Contact: Hugging Face user Tsu7am1