Nativa

Nativa is a procedural 2D anime engine (skeleton rigs, keyframed poses, camera, render, lip-sync) written entirely as code. Llama 4 Maverick is its director. From one natural-language prompt, Llama writes the full production plan as structured JSON: story beats, a 10-shot list with camera framing, per-character blocking (position, depth, gait, emotion, gestures), timed dialogue, music mood and sound-effect cues. The engine validates the plan (timing, spacing, on-screen overlaps) and sends every error back; Llama repairs its own spec until it passes (5 attempts in the final demo). Then Llama uses vision: it reviews a contact sheet of rendered frames, scores the staging and proposes concrete fixes, and the scene is re-staged. The engine renders 1280x720 on a CPU; dialogue is voiced with Kokoro TTS and lip-synced with Rhubarb. Result: a 72-second, 10-shot, 3-character scene with 8 voiced, lip-synced lines, music and SFX, from one prompt, for well under $1 of API calls. Unlike video generators, every frame is editable and deterministic. Tagline: Write a scene in plain words; Llama directs it, our procedural engine animates it — no AI video, no GPU. Pitch video: https://github.com/Oliver369X/nativa-meta-hackathon/releases/download/hackathon-v2/pitch-v2.mp4 Built during Oct 26–Nov 4 2026: Llama director pipeline, validate-and-repair, vision critic, demo/pitch v2. Procedural engine code is in the submitted repo.