World & spatial models
01For world model labs
Action-conditioned video: every frame paired with the exact input that produced it, plus the game’s own state. The training signal that turns a video model into a world model.
A data product from BUFF
Real human gameplay,captured in sync.
Gameplay video, every keystroke and mouse movement at 1 kHz, and the game’s own events, locked to the same frame. Recorded by real players on BUFF.
Sample session · illustrativeIllustrative
A scripted session in the same format as a real recording. Open the viewer to see real sessions.
1Player 1· mainFortnite · Battle Royale
FRAME000000· 30 FPSVIEWrebuilt from input
MAP · 1 Hz
Aim—
1 kHz sampling
ItemShotgun · 4
Acc36%ping 6
Players left70
No game event yet at this point
0:02.6 / 0:17.0Open the real viewer
The view above is rebuilt from mouse and keyboard input alone. In the viewer, the same data plays under the real video.
2M
players on BUFF
250K
play every day
700+
titles classified
1 kHz
input sampling
- FortniteGame events
- Counter-Strike 2Game events
- Apex LegendsGame events
- ValorantInput only
- PUBG: BattlegroundsGame events
- Marvel RivalsGame events
- GTA VInput only
- Overwatch 2Game events
- Rocket LeagueGame events
- Sea of ThievesInput only
- MinecraftGame events
- War ThunderInput only
- PalworldGame events
- RobloxInput only
What's in every recording
Five layers, one clock.
The same layers the viewer shows under every video. Nothing is stitched together afterwards: each layer carries the frame it belongs to.
What the player saw
Gameplay video
1080p at 30 fps in H.264, with in-game audio and no microphone. Every part carries its own frame, dropped-frame and lagged-frame counts.
1080p30H.264game audioper-part counts
004925
004926
004927
004928
004929
004930
004931
004932
004933
004934
33.3 ms / frame
1920 × 108030 fpsH.264 · CBRAAC game audiomicrophone off
Captured to order through a QuestEye trackingTeam voiceController & gamepadSkill-tier labelsIntent annotationsCustom scenarios
Inspect the data
Pause anywhere. Check every layer.
Real sessions with video, depth, camera and inputs on one clock. Step frame by frame and compare.
Multiplayer · 4 players · 1 session
Every player. One clock.
When a squad plays the same match on BUFF, you get every player’s video, input and game events, lined up on the same timeline.
IllustrativeSame moment on all 4 · within 0.15 s
14:07:55.0 UTC
Session timeline · UTCeach bar is one player’s recording
- P1
- P2
- P3
- P4
Every point of view
Each player in the match records on their own machine. Framestime groups those recordings into one session.
One wall clock
A corrected clock and per-frame stamps line every view up, so the same moment matches across all of them within 0.15 s.
Joins and leaves included
Each recording keeps its own start and end, so you can see exactly where every view overlaps with the others.
The rewards app gamers already use, and the source of every frame.
Built on BUFF
Not scraped. Recorded by the people who play.
Framestime is BUFF’s data product. The same app that pays players for their game time records it, so every session comes with consent, a known source and the data behind the video.
- 01
BUFF app
Players earn for the games they already play
Two million gamers use BUFF for rewards; 250,000 play on any given day, across hundreds of titles.
- 02
Capture
Sessions are recorded with the player’s consent
The BUFF app records video, every input, the game’s events, the device and the sync data, on the player’s own machine.
- 03
Framestime
Every recording is checked and classified
Sync and frame quality are validated, and each title is tagged by genre, perspective and play mode.
- 04
Your team
Review it, then take delivery
Watch sessions in the Framestime viewer with the data under the video, then receive the set by API or cloud bucket.
Open the viewer
A real session
Real footage, with the data under it.
A Fortnite recording in the viewer, with input and game events following the video.
Quests
If it isn’t there yet, we commission it.
Need a scenario the corpus doesn't have yet? We brief it to BUFF players as a Quest and ship the data in about four weeks.
01
Week 1
Define
Scope a brief together. Volume, edge cases, success metrics.
02
Week 2
Mobilize
Quest deploys across the right titles. Capture begins.
03
Week 3
Deliver
API or cloud bucket. Cosmos format on request.
04
Week 4
Measure
Joint eval on your benchmark. Iterate or scale.
Who uses this
Four kinds of teams, one corpus.
Humanoid & embodied AI
02For robot foundation models
Pre-train on human spatial intuition at scale, then fine-tune on your teleop. Navigation, aiming and reaction, from people moving through 3D worlds.
Autonomy & defense
03For AV, drones and autonomy
Edge cases too dangerous to capture on real roads or in real airspace. We design them as Quests and deliver the data in about a month.
Frontier & computer-use agents
04For agent training teams
Screen, every keystroke and click, and what the software reported back, all on one clock. Human-software interaction at the scale of a consumer app.
Public companyListed on the TASEGDPR / CCPACompliant
Player consentPer session
Player revenue shareBuilt in
How we compare
Other ways to source gameplay data, and what comes standard here.
| Capability | Clip-sharing platforms | Licensed video marketplaces | Model research labs | Framestime |
|---|---|---|---|---|
| Keyboard & mouse input, per frame | No | Partial | No | 1 kHz, frame-anchored |
| The game's own events & state | Pixels only | No | No | Native, per title |
| Device & capture context | No | No | Partial | Every session |
| Several players, same match | No | No | Partial | Same lobby, synced |
| On-demand scenarios | No | Catalog only | No | 4-week Quests |
| Player consent, per session | Partial | Creators only | Partial | Built into BUFF |
| Player revenue share | No | Creators only | No | Built-in |
| Public-company governance | No | No | No | TASE-listed |
Keyboard & mouse input, per frame
Clip-sharing platformsNo
Licensed video marketplacesPartial
Model research labsNo
Framestime1 kHz, frame-anchored
The game's own events & state
Clip-sharing platformsPixels only
Licensed video marketplacesNo
Model research labsNo
FramestimeNative, per title
Device & capture context
Clip-sharing platformsNo
Licensed video marketplacesNo
Model research labsPartial
FramestimeEvery session
Several players, same match
Clip-sharing platformsNo
Licensed video marketplacesNo
Model research labsPartial
FramestimeSame lobby, synced
On-demand scenarios
Clip-sharing platformsNo
Licensed video marketplacesCatalog only
Model research labsNo
Framestime4-week Quests
Player consent, per session
Clip-sharing platformsPartial
Licensed video marketplacesCreators only
Model research labsPartial
FramestimeBuilt into BUFF
Player revenue share
Clip-sharing platformsNo
Licensed video marketplacesCreators only
Model research labsNo
FramestimeBuilt-in
Public-company governance
Clip-sharing platformsNo
Licensed video marketplacesNo
Model research labsNo
FramestimeTASE-listed
Leadership
The team.
ND
Nir David
GM, Framestime
Leads commercialization and partnerships for Framestime. Bridges BUFF's data platform with the AI labs that need it most.
TS
Tomer Simon, PhD
Advisor
Chief Scientist at Microsoft Israel R&D. PhD from Tel Aviv University, executive education at Saïd Business School, Oxford. Expert in AI systems, data mining, and the intersection of computing and human behavior.
OG
Ophir Gertner
Co-founder
Two decades leading digital transformation and enterprise partnerships. Builds strategic alliances between technology companies and Fortune 500 enterprises.
ED
Elay De-Beer
Co-founder
Seasoned product leader with deep experience in AI-driven platforms and gaming technology. Driving Framestime's product vision from data infrastructure to market-ready AI.
OS
Ophir Sarapi
Co-founder
Combines deep operational expertise with years of leadership experience across gaming, technology, and high-growth digital businesses.
Get in touch
Tell us what you’re training, and what isn’t working yet.
One call is usually enough to tell whether the corpus already covers your case, whether a Quest makes sense, or whether neither is the right answer.
No spam. We reply within one business day.