Chrono-TriClass-100k

Chrono-TriClass-100k is a three-class video-forensics dataset:

Class	Folder	Videos	What it is
Real	Real/	33,334	Unmodified Kinetics-400 clips (the Real class of Chrono-66k)
AI-Generated	AI_Generated/	30,682	Fully synthetic videos from 9 modern video generators
AI-Edited	AI_Edited/	33,329	Real Kinetics-400 clips partially manipulated by 27 editing methods in 8 families
Total		97,345

Status: generation is still running for AI-Generated and AI-Edited (target 33,333 each). This release holds every clip finished so far; the rest will be added in later commits.

The AI-Generated and AI-Edited classes were produced specifically for this dataset, with category-matched content: both use the same 11 Kinetics-400 action categories as the real videos, so a detector cannot separate classes by subject matter alone.

Structure

Real/
  part-1/ … part-4/                       # Kinetics-400 clips (10,000-file chunks, no meaning)
AI_Generated/
  <category>/<generator>/<video_id>.mp4   # 11 categories x 9 generators
  metadata.csv                            # one row per video (prompt, seed, model, QC scores, …)
  generators.json                         # generator registry
AI_Edited/
  <family>/<method>/<video_id>.mp4        # 8 families x 27 methods
  metadata.csv                            # one row per video (source clip, donor, model, …)

Labels come from the top-level folder: Real/ → real, AI_Generated/ → AI-generated, AI_Edited/ → AI-edited. metadata.csv files add finer labels (generator, family, method) and a split column (train / valid / test).

Categories (AI-Generated): basketball, cooking, cycling, dancing, driving, interview, running, soccer, swimming, talking, walking.

AI-Generated: models

Each generator was prompted with category-specific prompts (text-to-video), a generated keyframe (image-to-video), or a real Kinetics clip's depth/pose (video-to-video). More candidates than needed were rendered per (generator, category) cell; every candidate went through quality control and the best ones were kept (see Quality control).

Generator balance. Every category holds ~3,030 clips. VideoCrafter2 and Show-1, the two oldest models, contribute a smaller share; the remaining quota goes to the newer, higher-quality generators, weighted by quality and by how reliably each one passes quality control in that category. Each prompt is used for exactly one video across the whole class.

Generator	Type	Conditioning	Checkpoint	Videos
CogVideoX-5B	Diffusion transformer (DiT)	text	zai-org/CogVideoX-5b	6,306
HunyuanVideo	Transformer	text	hunyuanvideo-community/HunyuanVideo	3,854
Wan2.1-T2V-14B	Text-to-video	text	Wan-AI/Wan2.1-T2V-14B-Diffusers	3,582
VideoCrafter2	Latent diffusion	text	VideoCrafter/VideoCrafter2	2,178
Show-1	Pixel + latent diffusion	text	showlab/show-1-base (+ interpolation, sr1, sr2)	2,051
Pyramid Flow	Flow matching	text	rain1011/pyramid-flow-miniflux	2,222
Cosmos-Predict2 (Text2World)	World model	text	nvidia/Cosmos-Predict2-2B-Text2Image → nvidia/Cosmos-Predict2-2B-Video2World	2,182
FLUX.1-schnell → SVD-XT	Image-to-video	generated image	black-forest-labs/FLUX.1-schnell → stabilityai/stable-video-diffusion-img2vid-xt	4,409
Wan2.1-VACE (depth/pose-driven)	Video-to-video	real video's depth/pose	Wan-AI/Wan2.1-VACE-1.3B-diffusers (depth: Depth-Anything-V2-Small)	3,898

Show-1 and Cosmos-Predict2 were run at half their default sampling steps to fit the compute budget.

Quality control

Every rendered candidate was checked before selection:

1. Technical gate: decodes, at least 8 frames, not black or blown out (8 < mean luma < 247), not frozen (mean absolute frame difference ≥ 0.5).
2. Semantic gate: X-CLIP ranks the cell's own category in its top 2.
3. Ranking: survivors are ranked by X-CLIP probability of the target category, and the best ones up to the cell's quota are kept.

Rejected clips are never used to pad a cell. The scores (xclip_p_target, xclip_rank) are in AI_Generated/metadata.csv.

AI-Edited: families and models

Each AI-edited video starts from a real Kinetics-400 clip and changes only part of it: a face, the mouth, an expression, an object, a region, the background or the style. method in metadata.csv is the method slot and model is the model that actually produced the clip. Some original methods have no usable public inference release; their slot is filled by the closest modern model, as listed in the table.

Family	Method slot	Model used	Videos
Face swap	inswapper	InsightFace inswapper_128 (buffalo_l)	1,952
	simswap	DreamID-V (diffusion-transformer video face swap)	3,481
	faceshifter	REFace (diffusion face swap; FaceShifter has no released inference checkpoint)	817
Facial reenactment	fomm	First Order Motion Model	1,750
	pirenderer	Thin-Plate Spline Motion Model (TPSMM)	1,750
	face2face	LivePortrait	1,750
Lip-sync	wav2lip	Wav2Lip (GAN)	1,313
	musetalk	MuseTalk	1,313
	sadtalker	SadTalker	1,312
	videoretalking	LatentSync 1.6	1,308
Expression / attribute editing	ganimation	LivePortrait (expression retargeting)	3,583
	styleganex	StyleGANEX (age / attribute editing)	1,167
Object insertion / removal	propainter_object	ProPainter	900
	object_wiper	DiffuEraser	900
	anyv2v_object	Wan2.1-VACE-1.3B	850
	videocomposer	Wan2.1-VACE-1.3B	850
Video inpainting	propainter_inpaint	ProPainter	900
	e2fgvi_hq	E2FGVI-HQ	750
	sttn	STTN	650
	fuseformer	FuseFormer	700
Background manipulation	bg_real_composite	SAM 2 segmentation + real background composite	750
	bg_flux_image	SAM 2 + FLUX.1-schnell generated background	750
	bg_svd_video	SAM 2 + SVD-XT generated video background	750
	bg_propainter_recon	ProPainter (moving foreground removed, background reconstructed)	750
Video-to-video	tokenflow	TokenFlow	785
	insv2v	Wan2.1-VACE-1.3B	785
	anyv2v_style	Wan2.1-VACE-1.3B	763

REFace (the faceshifter slot) could not process many Kinetics clips reliably, so the rest of that slot's face-swap quota was produced with DreamID-V, the strongest face-swap model in the set; those clips are listed under simswap and keep reassigned_from: faceshifter in their generation metadata.

Family totals: face swap 6,250 · expression/attribute 4,750 · lip-sync 5,246 · object insertion/removal 3,500 · facial reenactment 5,250 · background 3,000 · inpainting 3,000 · video-to-video 2,333.

For face swap, reenactment and lip-sync, the identity, driving motion or audio comes from a different Kinetics clip (driving_clip_id / audio_clip_id in metadata.csv), so every edit is a real manipulation and not a re-encode of the same clip. The bg_propainter_recon rows have no split value yet.

Usage

from huggingface_hub import snapshot_download

snapshot_download("madhav-yrc/Chrono-TriClass-100k", repo_type="dataset",
                  local_dir="Chrono-TriClass-100k")

# only the AI-Edited lip-sync videos
snapshot_download("madhav-yrc/Chrono-TriClass-100k", repo_type="dataset",
                  local_dir="Chrono-TriClass-100k",
                  allow_patterns=["AI_Edited/lip_sync/*", "AI_Edited/metadata.csv"])

import pandas as pd
gen = pd.read_csv("Chrono-TriClass-100k/AI_Generated/metadata.csv")
edit = pd.read_csv("Chrono-TriClass-100k/AI_Edited/metadata.csv")

Intended use and limitations

The dataset is intended for non-commercial research on detecting AI-generated and AI-manipulated video, including three-way classification (real / generated / edited) and per-generator or per-method attribution. It is not intended for producing deceptive media or for identifying people. Real and source content comes from YouTube via Kinetics-400 and shows real people.

Licence and upstream terms

Released under CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International): research and other non-commercial use only, with attribution. Commercial use is not permitted.

The videos also remain subject to their upstream terms:

• Kinetics-400 clips come from YouTube and are subject to YouTube's terms and the rights of the original uploaders.
• REFace (the faceshifter slot) uses a checkpoint trained on CelebAMask-HQ, which permits non-commercial research use only; those videos inherit that restriction.
• Generator and editing model licences vary (Apache-2.0, MIT, research-only and model-specific licences); see each model's page linked above.

These upstream terms are compatible with, and in places stricter than, the CC BY-NC 4.0 licence.

Related

• Chrono-66k - two-class predecessor: Kinetics-400 real vs. VidProM (Pika, VideoCrafter2) AI-generated videos.