IRIS: Intrinsic Relational Identity via Slots

We introduce IRIS (Intrinsic Relational Identity via Slots), a self-supervised worldmodel that discovers, tracks, and understands all entities composing a video purelyfrom internal evidence without recognition, external labels, or pretrained priors.Each entity is assigned a latent slot whose state evolves as a continuous trajectoryin a learned latent space L = Rd. Identity is not dened by visual similarity but bybehavioral coherence: the temporal regularity of a slot's trajectory, the structuralconsistency of its relational ngerprint, and its separation from all other slots. Wederive a self-supervised loss L = λ1Ltemp + λ2Lrel + λ3Ldist from rst principles andprove that, under mild regularity assumptions, (i) every local minimum of L inducesa bijection between slots and true entities, and (ii) every non-minimal critical pointis a strict saddle guaranteeing convergence of perturbed gradient descent to avalid entity decomposition. As a primary application, IRIS powers POV-Cinema:a pipeline that reconstructs any lm or anime as an immersive rst-person videofrom the perspective of any chosen character, using exclusively frames and audiopresent in the original content, enabling a new class of VR viewing experiences.