ChitraGati: A Grammar-Constrained Intelligent System For AI-Assisted Transcreation of Indian Folk Visual Art into Heritage Animation

Digitisation has served Indian folk art in one narrow sense.There are more photographs of Warli walls, Madhubani kohbars, Pattachitra pattas and Paithani pallus than at any earlier point in history.It has served far less well the part a practitioner would call essential: the rule system that decides what may be drawn, where it may sit, in which colour, and in what order.This paper treats that rule system as a first-class computational object.We define the Folk Visual Grammar Schema (FVGS), a typed and machinereadable encoding of a tradition's motif lexicon, admissible spatial syntax, chromatic constraints, compositional invariants and permissible motion ranges, and we use it not as training data but as an inferencetime constraint on generative models.Around this schema we propose ChitraGati, a five-layer intelligent system for heritage animation.A perception layer performs folk motif recognition using a self-supervised vision transformer backbone with prototypical few-shot heads and open-set rejection, so that an unfamiliar motif is flagged rather than forced into the nearest known class.A generative layer performs AI-assisted animation generation: style-adapted latent diffusion with low-rank adapters and structural control produces assets, while a compact diffusion model over Motion Grammar Tokens, a discrete vocabulary of culturally attested kinetic primitives, produces motion in parameter space rather than pixel space, keeping output editable, auditable and stylistically bounded.A governance layer binds every generated asset to community-issued Traditional Knowledge Labels and cryptographically signed provenance manifests, and enforces a communityauthored exclusion list for sacred imagery as a hard constraint rather than a soft prior.A pedagogy layer supplies an AI-based educational evaluation framework that traces learner mastery across a heritage concept graph, calibrates automatically generated assessment items, and closes the loop by feeding learning signals back to the generator so that segmentation, pacing and motif salience adapt to the learner.We introduce the Grammar Fidelity Score, a decomposable and computable measure of cultural fidelity, and the Artisan Concordance Index, which tests that measure against practitioner judgement instead of assuming agreement.The paper specifies the architecture, the guidance and loss formulations, a corpus and consent protocol, and a pre-registered evaluation plan with explicit success criteria.No empirical results are claimed at this stage; the contribution is a design, and specifically a design for letting generative systems move folk art without dissolving it.