By requesting access you agree to the Data Sample Evaluation Terms: the sample sessions and metadata in this repository are for internal evaluation only. Models trained on samples are evaluation artifacts and may not be deployed, released, or used commercially. No redistribution of the data in any form, no publication without written consent from Origin Lab, and deletion within 30 days of download — including evaluation weights, or ceasing all use of them. Access to the full corpus is licensed separately, per agreement.

OriginLab Game Recordings v0.4.0

Human gameplay recorded inside the engine of 20 licensed games at 1080p and 60 FPS: RGB with and without the HUD, depth, surface normals, camera pose, keyboard and mouse input, in-engine events and game state, and audio. This release doubles the roster to 20 games and adds the first Unity title.

Previews and per-game statistics: app.originlab.ai/data.


Sessions	169
Hours	163.0
Games	20
Size	≈5.3 TB
Capture	1080p / 60 FPS CFR, in-engine SDK, shared frame clock

Dataset summary

Recorded inside the engine, not captured from the screen. All streams share one frame clock: frame k of the RGB is frame k of the depth, normals, pose and frame table, so the frame index is the join key.

Against other datasets

Other game datasets offer engine depth and pose for one title, or player input without geometry, or estimated geometry; simulators offer ground truth without human play. This release has all of them, across 20 titles, on one frame clock.

Dataset	Source	Sequences	FPS	Resolution	Frames	Depth	Camera pose	Surface normals	Player input	Game events, state, text	Audio	Optical flow	Fg. masks
Origin v0.4.0 (ours)	20 commercial titles, engine hooks	169 sessions	60	1920×1080	≈35M	✅ engine, 10-bit log code	✅ engine	✅ g-buffer	✅ keyboard + mouse	✅ events + state fields	✅	🔧	🔧
WildWorld (ECCV 2026) [1]	1 title (Monster Hunter Wilds), ReShade + engine hooks	3,434 clips, up to 30 min	30	1280×720	108M	✅ render buffer, 8-bit	✅ engine	❌	semantic action triplets, not input	✅ state + skeletons + VLM captions	❌	❌	❌
EgoCS-400K (2026) [2]	1 title (Counter-Strike), replay-rendered	400K videos	?	?	10,000 h	❌	✅ replay	❌	✅ inputs + view angles	✅ events + states + language	?	❌	❌
WorldCam-50h (2026) [3]	3 titles (CS, Xonotic, Unvanquished), screen capture	≈300 videos	30	?	50 h	❌	🟡 ViPE	❌	✅ keyboard + mouse	🟡 captions	❌	❌	❌
Sekai-Game (NeurIPS 2025) [4]	1 title (Lushfoil), UE5 hooks	60 h	30	1920×1080	6.5M	❌	✅ UE hooks	❌	❌	✅ location + captions	✅ AAC stereo	❌	❌
OmniWorld-Game (2025) [5]	game environments, self-captured with ReShade	96K clips (paper); public release: 484 videos, 11,482 clips	24 / 30	1280×720	18.5M (paper); 2.16M public	✅ render buffer via ReShade	🟡	❌	❌	🟡 captions	❌	🟡	🟡
TartanAir V2 (2023) [6]	63 simulator environments (AirSim / UE), no gameplay	?	?	640×640	?	✅	✅	❌	❌	✅ segmentation	❌	✅	✅ segmentation

✅ ground truth (engine, replay or recorder) · 🟡 estimated by a model · ❌ absent · 🔧 in progress · ? not stated. Read 2026-09-21; OmniWorld-Game rechecked 2026-09-30.

Supported tasks

Task	Why this dataset
Depth estimation	Engine depth for every RGB frame, not estimated.
World models & video prediction	Long, continuous sessions with dense action and camera conditioning signals.
Imitation learning	Frame-level keyboard/mouse actions paired with what the player saw.
Camera pose & ego-motion	Per-frame 6-DOF camera extrinsics straight from the engine.

Games

Titles are anonymized for licensing; real titles are disclosed under the full-dataset agreement. Game numbers are this release's own and do not match v0.3.0's.

Hours per game, one segment per session, with each game's size. Grey: game also in v0.3.0; blue: new. Every session is a new capture.

Figures in this section were measured on the 147 sessions (141.5 h) delivered as of 2026-09-30, before the final sessions landed; they will be refreshed on the full release.

SigLIP [7] features of 64 frames per session, projected with UMAP [8]. Hollow: v0.3.0; filled: v0.4.0; colour: title. Returning titles overlap their v0.3.0 regions; new titles open their own.

Contribution of each title

Distinct footage each title adds (leave-one-out Vendi score [9]), relative to removing the same amount of random footage (dashed line, 1.0). Above 1.0 the title adds frames the rest lacks; below 1.0 it repeats them. Bars: spread over draws; gaps under 0.1 are noise.

Game 6, Game 19 and Game 18 add the most distinct appearance (1.52, 1.41 and 1.41). Game 20 adds well under its size (0.65): its sessions are races that look alike.

Scene coverage, v0.3.0 to v0.4.0

Share of a batch's 1.6 s windows that land in scene cells (DINOv2 [10], fixed grid sized on the CS:GO footage of DIAMOND [11, 12]) that a half of v0.3.0 never visits. Every batch matches that half in hours (about 72 h). Bars: 5–95% of 200 draws.

The new titles bring new scenes at about three times the rate of a same-size v0.3.0 batch: 2.7% of their windows against 0.8%. The titles also in v0.3.0 add 1.1%. Adding v0.4.0 to v0.3.0 raises the number of grid cells visited by 38%.

Camera motion

Engine camera pose over 168,456 three-second windows. Left: share of windows in six bins of speed and turn rate (thresholds under each bin, set by hand). Right: median turn rate, 20.5°/s against 12.9 for OmniWorld-Game [5] and 4.5 for RealEstate10K [13]. Those references use estimated poses, which inflates their rates.

Dataset composition

Figures in this section were measured on the 147 sessions (141.5 h) delivered as of 2026-09-30, before the final sessions landed; they will be refreshed on the full release.

Hand labels on 2 frames per session (294 frames; v0.3.0: 4 per session) for setting, time of day, fog and terrain, by one labeller not blind to the release. Intervals come from resampling sessions.

Condition	v0.3.0	v0.4.0	Terrain, of frames with a clear terrain	v0.3.0	v0.4.0
Outdoor, of frames with a clear setting	73%	70%	Urban / built	18%	23%
Night, of outdoor frames	13%	13%	Grassland / open	14%	17%
Fog or haze, of outdoor daytime frames	15%	23%	Interior (furnished rooms)	8%	17%
			Forest	17%	14%
			Industrial / facility	17%	12%
			Desert	16%	10%
			Cave / underground	3%	3%
			Snow / alpine	3%	3%
			Water / coast	4%	2%

The new titles move the mix toward urban scenes, open grassland and furnished interiors, and they are foggier (30% of outdoor daytime frames against 15%). The outdoor and night shares are unchanged.

9.6% of sampled frames closely match a frame from another session (cosine similarity 0.95 or more), against 4.9% in v0.3.0, mostly from two driving titles: Game 20 (43%) and Game 3 (25%).

Where the hours go

Every frame is assigned to one of five kinds from the capture's event flags. Idle is playable footage with no key held and no mouse movement.

Kind	Share
Raw play	72.0%
Idle	22.0%
Menu	3.2%
Loading	1.8%
Cutscene	1.0%

Per title, idle runs from 4% to 42%; loading, menus and cutscenes are each concentrated in one title (21%, 25%, 14%). Idle frames are labelled, not removed. Game 19 has no mouse movement in its input, so its idle share (29%) includes mouse-looking; counting that as play, it is 15%.

Dataset structure

One folder per session (video/, depth/, telemetry/, tables/), no archives. About 33 GB per hour of footage.

No predefined splits: metadata/sessions.parquet indexes the sessions (game, duration, sync status, sizes) for cutting your own.

Data files and fields

File	Purpose
`video/prehud.mp4`	1080p H.264, 60 FPS CFR, HUD removed, in-engine capture with game audio. Every stream starts at the shared frame 0
`video/posthud.mp4`	The frame exactly as the player saw it, HUD included, on the same 60 FPS clock
`video/normals.mp4`	Per-pixel surface orientation rendered by the engine (world frame), same clock
`video/mosaic.mp4`	The four visual streams composed as quadrants of one video — a playable sync proof, not extra data
`video/mosaic_layout.json`	Which mosaic quadrant holds which stream
`depth/depth.hevc`	10-bit HEVC elementary stream, log-encoded relative depth; frame k is RGB frame k
`depth/depth_meta.jsonl`	Per-frame depth params: depth_transform, range, FOV
`depth/decode_contract.json`	Session decode contract: near plane (makes depth metric), pinhole intrinsics, per-stream frame accounting (fps, dup counts), and the world frame
`telemetry/camera.jsonl`	Camera pose keyed by frame index and the Windows high-resolution timestamp, field `qpc` (≈2 samples per frame): position, pitch/yaw/roll
`telemetry/input.jsonl`	Mouse (position and dx / dy deltas), keyboard (key_down / key_up with modifiers), scroll, and window-focus events, frame-indexed
`telemetry/events.jsonl`	In-engine action events: a schema kind (weapon_fired, enemy_killed, item_pickup, ...) plus the engine's own label, frame-indexed
`telemetry/state.jsonl`	Sampled game state (health, active tool, ...): field, unit, and value per sample, frame-indexed
`telemetry/annotation.jsonl`	AI mechanic-detection intervals (label, category, confidence) where available
`telemetry/world.json`	World telemetry: engine, world-to-meters scale, handedness, gravity
`telemetry/gameclock.jsonl`	In-game clock samples, where the title exposes a readable clock (roughly two thirds of sessions)
`telemetry/inventory.jsonl`	New in this release: one record per changed inventory slot (toolbar or bag) with item id, resolved item name, count, ammo, and decay, frame-indexed
`telemetry/quest.jsonl`	New in this release: one record per quest or objective change with resolved quest and objective names, progress / goal, and state, frame-indexed
`telemetry/ability.jsonl`	New in this release: one record per ability or stat change with resolved name, level, value, and state, frame-indexed
`tables/frames.parquet`	One row per video frame, training-ready: camera pose, held-keys bitmask, mouse deltas, state columns, and event flags, all pre-joined on the frame index
`tables/events.parquet`	One row per in-engine event with resolved labels
`tables/conversion_manifest.json`	Every binning rule and cap used to build the tables, so they are re-runnable
`session.json`	Manifest: files + sizes, fps, the shared-clock alignment statement, frame accounting, and the sync audit (sync_status / sync_report)

What is new vs v0.3.0

	v0.3.0	v0.4.0
Games	10	20, at most 10 h per game
Engines	Unreal	Unreal + the first Unity title
Inventory	not captured	per-slot changes with resolved item names (`telemetry/inventory.jsonl`)
Quests	not captured	quest / objective changes with progress and state (`telemetry/quest.jsonl`)
Abilities	not captured	ability / stat changes with level and value (`telemetry/ability.jsonl`)
Game state fields	30	36
Event kinds	74	76

Scenes: the ten new titles reach scenes v0.3.0 lacks at about three times the rate of a same-size v0.3.0 batch.
Inputs: no new key combinations, turns or input sequences beyond what a v0.3.0 batch of the same size adds (0.17% of half-second blocks against 0.31%). New places, familiar controls.

In-engine events & game state

Events and game state are on the shared frame clock, and the tables join both onto the per-frame grid. The format differs by engine:

// Unreal titles: the event carries the engine's function label
{"frame":203057,"qpc":1200000000001,"kind":"weapon_fired",
 "source":"ue_processevent","confidence":255,"unit":"raw",
 "arg_i":0,"arg_f":0.0,"label":"OnPrimaryFireShot"}

// Unity titles: events come from memory polling; label is null
{"frame":1210,"qpc":383217012849,"kind":"jump",
 "source":"memory_poll","confidence":200,"unit":"raw",
 "arg_i":0,"arg_f":2.696,"label":null}

Dataset creation

Consenting, paid players record long guided sessions with our in-engine SDK. Depth and camera pose come from the engine at capture time: measured, not estimated. Every stream is aligned to the shared clock before delivery; ticks a stream did not capture are dropped from all streams and listed in session.json under frame_accounting.alignment.

How to use it

Needs ffmpeg and numpy. RGB is H.264 mp4. Decode depth to 16-bit frames with ffmpeg -f hevc -i depth/depth.hevc -f rawvideo -pix_fmt gray16le -, then:

import numpy as np
K = 4000.0  # log-depth constant; see depth_transform in depth/depth_meta.jsonl

def decode_depth(luma):                    # uint16 frame from ffmpeg
    q = luma.astype(np.float32) / 65535.0
    return (2.0 ** (q * np.log2(1.0 + K)) - 1.0) / K   # 0 = nearest, 1 = farthest

valid = luma < 65535   # max value = sky or far clip; mask it out
# metric planar depth: multiply by near_plane_cm from depth/decode_contract.json

Pull one session:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="originlab/game-recordings-v4",
    repo_type="dataset",
    allow_patterns=["metadata/*", "sessions/d5389dcf-d386-4f62-943c-478593f202cc/*"],
    local_dir="./v4",
)

What is in this repository

sessions/<uuid>/: one folder per session, in the delivery layout (169 sessions, 163.0 h).
samples/<uuid>/: a few quick-start sessions (a subset of sessions/).
metadata/sessions.parquet: the session index shown in the dataset viewer.
metadata/files.parquet: every file and its size, to check a download.
assets/previews/: the mosaic preview. tools/: session loader and resumable downloader.

Teams on AWS can request in-cloud delivery instead: contact Origin Lab after approval.

Considerations for using the data

Repeated frames. Some frames repeat the previous one (median 13% per session, worst 57%); counts per stream are in depth/decode_contract.json as dup_count.
Depth resolution. Upscaled titles render depth below 1920×1080; read each session's size and intrinsics from depth/decode_contract.json.
Stream gaps. Game 19: input only in telemetry/input.jsonl. Game 20: pose on about a quarter of frames. Game 3: no events. Games 3 and 6: state only in telemetry/state.jsonl. Game 14: state on 70% of frames.
Depth units. Relative, not metric, on Game 8 build 1.2.38, Game 18 and Game 20.
Depth follows the engine's depth buffer, so fog, glass and particles vary by title.
Mechanic annotations are model-generated.
No microphone audio or player-identifying information ships; real game titles are redacted from text files.

Licensing

All gameplay is recorded under non-exclusive licenses with the rights holders and captured by consenting, compensated players.

Samples in this repository ship under the Data Sample Evaluation Terms accepted when you request access: internal evaluation only. Models trained on samples are evaluation artifacts and may not be deployed, released, or used commercially. No redistribution, no publication without written consent, and deletion within 30 days of download, including evaluation weights.

The full dataset is licensed per agreement, with production training and deployment rights defined in your contract. Contact Origin Lab to license. See LICENSE.md for the complete terms.

Methods and references

Comparison table: each dataset's paper or repository [1–6]. Frame map and contribution: SigLIP [7], UMAP [8], Vendi score [9]. Scene coverage: DINOv2 [10] on a grid sized with the CS:GO footage of DIAMOND [11, 12]. Camera references: OmniWorld-Game [5], RealEstate10K [13].

Z. Li, Z. Meng, S. Shi, et al. WildWorld: A large-scale dataset for dynamic world modeling with actions and explicit state toward generative ARPG. arXiv:2603.23497, 2026.
R. Guo, D. Liang, Y. Liu, et al. EgoCS-400K: An egocentric gameplay dataset for world models. arXiv:2606.18180, 2026.
J. Nam, Y. Hong, C.-H. P. Huang, et al. WorldCam: Interactive autoregressive 3D gaming worlds with camera pose as a unifying geometric representation. arXiv:2603.16871, 2026.
Z. Li, C. Li, X. Mao, et al. Sekai: A video dataset towards world exploration. arXiv:2506.15675, 2025.
Y. Zhou, Y. Wang, J. Zhou, et al. OmniWorld: A multi-domain and multi-modal dataset for 4D world modeling. arXiv:2509.12201, 2025.
W. Wang, D. Zhu, X. Wang, et al. TartanAir: A dataset to push the limits of visual SLAM. IROS, 2020. Version 2: tartanair.org.
X. Zhai, B. Mustafa, A. Kolesnikov, L. Beyer. Sigmoid loss for language image pre-training. ICCV, 2023. arXiv:2303.15343.
L. McInnes, J. Healy, J. Melville. UMAP: Uniform manifold approximation and projection for dimension reduction. arXiv:1802.03426, 2018.
D. Friedman, A. B. Dieng. The Vendi score: A diversity evaluation metric for machine learning. TMLR, 2023. arXiv:2210.02410.
M. Oquab, T. Darcet, T. Moutakanni, et al. DINOv2: Learning robust visual features without supervision. TMLR, 2024. arXiv:2304.07193.
E. Alonso, A. Jelley, V. Micheli, et al. Diffusion for world modeling: Visual details matter in Atari. NeurIPS, 2024. arXiv:2405.12399.
T. Pearce, J. Zhu. Counter-Strike deathmatch with large-scale behavioural cloning. IEEE Conference on Games, 2022. arXiv:2104.04258.
T. Zhou, R. Tucker, J. Flynn, G. Fyffe, N. Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM Transactions on Graphics (SIGGRAPH), 2018. arXiv:1805.09817.

Citation

@misc{originlab2026gameplaycore,
  title  = {OriginLab Gameplay-Core: Frame-Synced RGB-D Gameplay with
            Actions, Camera Pose, and Mechanic Annotations},
  author = {Origin Lab},
  year   = {2026},
  url    = {https://app.originlab.ai}
}

Downloads last month: 600

Total file size:

5.5 TB

Papers for originlab/game-recordings-v4

EgoCS-400K: An Egocentric Gameplay Dataset for World Models

Paper • 2606.18180 • Published Jun 16 • 16

WildWorld: A Large-Scale Dataset for Dynamic World Modeling with Actions and Explicit State toward Generative ARPG

Paper • 2603.23497 • Published Mar 24 • 92

WorldCam: Interactive Autoregressive 3D Gaming Worlds with Camera Pose as a Unifying Geometric Representation

Paper • 2603.16871 • Published Mar 17 • 61

OmniWorld: A Multi-Domain and Multi-Modal Dataset for 4D World Modeling

Paper • 2509.12201 • Published Sep 15, 2025 • 107

Sekai: A Video Dataset towards World Exploration

Paper • 2506.15675 • Published Jun 18, 2025 • 67