name stringlengths 3 22 | url stringlengths 26 71 | type stringclasses 3
values | categories stringlengths 32 94 | download stringlengths 19 79 | license stringclasses 8
values | use stringlengths 164 278 | prompt_template stringlengths 85 213 |
|---|---|---|---|---|---|---|---|
Kinetics-700 | https://github.com/cvdfoundation/kinetics-dataset | video | 700 human action categories, 650K clips | bash k700_2020_downloader.sh (youtube-dl from CVDF scripts) | CC BY 4.0 | Human action recognition. Covers every motion type: sports, instruments, daily activities, social interactions. 10-second clips from YouTube. Foundational dataset for video understanding. | human action: {action_label}. YouTube clip, 10 seconds, 24fps. Reference for motion understanding training. |
AVA | https://research.google.com/ava/ | video | 80 atomic visual actions, 430 movie clips (15-min each) | https://storage.googleapis.com/deepmind-media/Datasets/ava_kinetics_v1_0.tar.gz | CC BY 4.0 | Atomic visual actions in movie contexts — walking, sitting, standing, talking, listening, touching, holding, carrying. Spatio-temporal bounding boxes. Perfect for fine-grained motion understanding. | movie scene with human action: {action_label}. Hollywood film clip, 24fps, cinematic lighting, production quality. Reference for character action training. |
Something-Something V2 | https://developer.qualcomm.com/software/something-something-v-2-dataset | video | 174 basic hand-object interactions, 220K clips | Qualcomm Developer portal (free registration) | Qualcomm terms | THE contact dynamics dataset. Pushing, pulling, opening, closing, picking up, putting down, pouring, dropping, throwing, catching. Every hand-object interaction AI fails at. Essential for fine-grained motion training. | hand-object interaction: {action_label}. Close-up on hands performing precise action. Human hand, real object, natural motion. Reference for contact dynamics and fine-grained motion training. |
ShotBench | https://huggingface.co/datasets/Vchitect/ShotQA | image | 17 shot types, camera angles, lighting conditions | huggingface datasets load_dataset('Vchitect/ShotQA') | Research | Expert-level cinematic shot classification from films. Extreme close-up, close-up, medium, wide, establishing. Camera angles: low, high, birds-eye, Dutch. Lighting: golden hour, blue hour, neon, natural. Perfect for training cinematic understanding. | cinematic shot from film: {shot_type}, {camera_angle}, {lighting}. Hollywood production quality, 24fps, professional cinematography. |
CameraBench | https://github.com/sy77777en/CameraBench | video | Camera motion types: pan, tilt, dolly, zoom, crane, handheld, static, compound | GitHub repo scripts | Research | Camera movement understanding from video. Classifies every camera move type — pan left/right, tilt up/down, dolly in/out, crane up/down, zoom in/out, handheld, static. Essential for teaching Wan 3.0 camera vocabulary. | camera movement: {camera_type}. {description}. Cinematic quality, smooth motion, professional camera operation. |
EPIC-KITCHENS-100 | https://huggingface.co/datasets/awsaf49/epic_kitchens_100 | video | Egocentric hand-object interactions, 100 hours, 90K action segments | huggingface datasets load_dataset('awsaf49/epic_kitchens_100') | CC BY-NC 4.0 | First-person (egocentric) hand-object manipulation in kitchen. Cutting, peeling, opening, closing, pouring, stirring, washing, wiping. The most detailed hand-object interaction dataset. Perfect for contact dynamics training — exactly what AI fails at. | egocentric hand action: {verb} {noun}. First-person POV looking down at hands. Detailed hand-object contact, natural kitchen environment. Reference for fine-grained contact dynamics and hand manipulation training. |
RealCam-Vid | https://huggingface.co/datasets/MuteApo/RealCam-Vid | video | 100K video clips with metric-scale camera trajectories | huggingface-cli download MuteApo/RealCam-Vid | MIT | Metric-scale camera trajectory annotations. Intrinsics (fx, fy, cx, cy) and extrinsics (4x4 w2c matrices). Dynamic scene + dynamic camera focus. Best dataset for teaching Wan 3.0 precise camera control. | camera trajectory annotated: {short_caption}. Metric-scale camera motion with intrinsics and extrinsics. Reference for precise camera control training. |
Video-Detailed-Caption | https://huggingface.co/datasets/wchai/Video-Detailed-Caption | video | 1,027 videos with structured captions (camera, background, main object, detailed) | huggingface datasets load_dataset('wchai/Video-Detailed-Caption') | Apache 2.0 | THE captioning gold standard. Each video has: camera_caption (shot types, movements, transitions), short_caption, background_caption, main_object_caption, detailed_caption. Structured, detailed, training-quality captions already written. Better than anything Pexels can provide. | {camera_caption} {detailed_caption}. Reference quality video captioning for training. |
MovieNet | https://movienet.github.io/ | video | 1,100 Hollywood films, shot boundaries, scene boundaries, camera movements, character metadata | Annotations from GitHub, films sourced separately | Research (annotations only) | Hollywood film grammar at scale. Shot boundaries, camera movements, scene structure from 1,100 films. Teaches the brain real cinematic language from real movies — not YouTube clips, not stock footage. | Hollywood film shot: {shot_type}, {camera_movement}. Professional cinematography from feature film. 24fps, cinematic production. |
HACS | http://hacs.csail.mit.edu/ | video | 1.55M human action clips + 50K temporal action segments | python download_videos.py from GitHub repo | MIT research | Massive human action dataset from YouTube. Complements Kinetics with longer temporal segments. Better for understanding action sequences rather than single actions. | human action segment: {action_label}. YouTube clip, 24fps. Action sequence with temporal boundaries. |
Panda-70M | https://huggingface.co/datasets/multimodalart/panda-70m | video | 70M video-text pairs, 167K hours | 10M subset CSV from Google Drive, download with video2dataset | Research | Largest video-caption dataset. Use the 10M subset for pre-training caption understanding. Captions from multiple cross-modal teachers. Good for contrastive learning of video-language alignment. | {caption}. Video clip from Panda-70M dataset. Reference for video-language alignment training. |
DAVIS | https://davischallenge.org/ | video | 150 video sequences, dense object segmentation masks | Registration required at davischallenge.org | Research | Densely annotated video segmentation. Pixel-level masks for objects across frames. Essential for VFX training — rotoscoping, object isolation, background replacement. | video object segmentation: {object_class}. Dense segmentation masks across frames. Reference for VFX and object tracking training. |
BVI-HFR | https://data.bris.ac.uk/data/dataset/k8bfn0qsj9fs1rwnc2x75z6t7 | video | 22 native 120fps HD video sequences | Direct download from University of Bristol | CC BY 4.0 | Ground truth high frame rate footage. 120fps native — can be used as reference for frame interpolation training, slow-motion generation, and temporal quality assessment. | high frame rate footage: 120fps native. Smooth motion, no interpolation artifacts. Reference for frame interpolation and slow-motion training. |
improved-flux-prompts | https://huggingface.co/datasets/k-mktr/improved-flux-prompts | text | 17,307 FLUX-optimized prompts from CivitAI + Hermes3 enhancement | huggingface datasets load_dataset('k-mktr/improved-flux-prompts') | MIT | THE FLUX prompt reference dataset. 17K prompts optimized for FLUX.1 with subject, style, composition, lighting, color palette, mood, technical specifications. Every prompt is a masterclass in FLUX-specific prompt engineering. Use for Qwen creative director training. | FLUX-optimized prompt: {prompt}. Reference quality image prompt with subject, style, composition, lighting, and technical specifications. |
No dataset card yet
- Downloads last month
- 5