Introducing JointDiff: a Kognia model using AI to transform performance analysis

Think of a “What if engine” for football tactical data. What if the winger stayed wide rather than came inside? What if the possession order changed and the full back received the ball earlier in a sequence? 

JointDiff is the latest Kognia innovation. Guillem Capellera, part of the AI team at Kognia, provides further insight into this latest work and how the research can transform the approach teams take to performance analysis.

JointDiff is a foundational generative AI architecture that aims to solve one of the hardest problems in sports analytics: the future simulation of gameplay in team sports. The architecture enables realistic simulations under text constraints, opening new possibilities for data augmentation and strategic analysis.

Generative models struggle to jointly model continuous signals (player trajectories) and discrete events (e.g. ball possession). JointDiff unifies both within a single diffusion framework, and we unlock controllable generation via natural language and player-level constraints. 

While traditional models attempt to model player tracking overseeing the event data, JointDiff advances diffusion modeling to unify them. It treats team sports as the joint modeling of players and ball movement and the most basic event like who is possessing the ball, as a “one”. By understanding the deep relationship between a player's physical trajectory and the events they trigger, JointDiff can generate or predict highly realistic sequences of play that respect both the laws of physics and the tactical logic of the game.

Initial research to show key findings

Our research proposal, which has been accepted at the prestigious ICLR 2026, demonstrated that modeling player trajectories jointly with discrete possessor events leads to physical and tactical improved outcomes. Meanwhile "siloed" data modeling is the enemy of accuracy.

Some of our initial key learnings include:

• Coordination is key: High-fidelity simulation requires modeling all 22 players and the ball, jointly with discrete ball-possession events. JointDiff excels at capturing the ball physics and dynamics, ensuring that ball physics and agent intent are intrinsically linked. This holistic approach yields significant performance gain over existing state-of-the-art algorithms, which model trajectories in isolation and consequently fail to maintain the tactical coherence required for professional-grade sports analytics. 

• Realism vs. randomness: Unlike other generative models that produced "robotic" or jittery movements, JointDiff produces smooth, human-like trajectories that can be recognised as authentic.

• The Power of constraints: We introduced an operation that allows us to "guide" the AI generation. By giving the model a specific constraint (e.g. "The play must end in a shot from the left wing"), the model can fill in the most statistically likely movements that lead to that outcome.

How this Kognia research will impact the professional game and enhance performance analysis

JointDiff moves Kognia beyond descriptive analytics (what happened) and into prescriptive simulation (what could have happened). We expect teams to use this "engine" in three main ways:

• Dynamic "what if" analysis: Coaches can modify a single variable, such as a defender’s starting position, and the AI will simulate how the rest of the play would likely unfold in response

• Synthetic pre-match testing: Before a game, analysts can run 10,000 simulations of an opponent’s build-up patterns against their own defensive structure to identify high-risk zones

• Data enhancement: For particular matches where tracking data is low-quality or "noisy," JointDiff can act as a "smart filter," automatically correcting errors and filling in gaps to provide professional-grade insights from any video source

If we look at some more practical use cases:

Example 1: Optimising defensive shape: “How would this have played out if we had a back three instead of a back four?” A common conversation in coaching and analysis meetings, and something that now can be simulated to address the once-hypothetical question.

Example 2: The tactical ghost: By allowing coaches to manually adjust player positioning within a real game clip, this feature creates a "living" tactical board for testing alternative defensive strategies. The AI then automatically re-simulates the entire sequence, predicting how both the opposition and teammates would have realistically reacted to that specific tactical change.

Example 3: Closing the data gap: In many situations, ball-player tracking data can be inconsistent. JointDiff can take possessor event data (e.g., "Possession from Player A to Player B") and automatically correct the 2D trajectories for all 22 players and the ball to a high level of accuracy. This allows Kognia to provide tactical and physical metrics even when the raw sensor data is incomplete.

JointDiff represents a major leap for sports analytics. By bridging the gap between descriptive and prescriptive analysis, this generative AI engine is set to become the foundational component of football analytics, providing teams with an unparalleled tactical advantage and reshaping how the future of performance is simulated and understood.

To find out more about JointDiff and the work we are doing at Kognia, get in touch: info@kogniasports.com