Fantasy Studio

Fantasy Studio is a local-first desktop app that turns one line of text into either a cinematic 3D video or a playable game — on your own hardware, with output you own outright. Type "a red Ferrari racing through the desert at sunset" and Studio produces a photoreal, textured, animated 3D scene rendered to MP4, plus the editable source scene you can re-light, re-frame, and re-render forever. Switch to game mode and type "a knight collecting gems in a forest" — about a minute later you're playing a self-contained 3D web game with a physics character controller, wandering NPCs, collectible objectives, a HUD, and a win screen, embedded right in the app. The game build needs no GPU at all. The core bet: AI should direct real tools, not replace them with uncontrollable pixels. Diffusion video tools like Sora, Runway, and Pika generate frames you can't edit — you can re-roll the dice, but you can't re-frame the shot or own the scene. Fantasy Studio builds actual 3D: real meshes, real physics, real render engines. Deterministic, directable, editable, and free per render after install. HOW VIDEO MODE WORKS — a nine-stage local pipeline: a local LLM (Gemma 3 / Qwen3 via Ollama) extracts a structured scene plan from your prompt, generates a pose-locked reference image of your subject, then converts it into a natively textured 3D mesh using TRELLIS.2 — which beat every alternative engine 4/4 in head-to-head tests, producing realistic fur, full armor, and glossy car paint instead of clay blobs. The mesh is cleaned, orientation-verified against the reference, rigged by motion archetype (legged subjects get auto-detected skeletons and looping gait cycles; vehicles get drive physics and wheel spin), placed into a real environment — OpenStreetMap-extruded cities or real-world terrain from elevation data — lit with camera-aligned key lighting, and rendered headless in Blender (EEVEE or Cycles) to MP4. Structural quality gates run throughout so it fails loudly instead of shipping a broken render. About six minutes per clip on a 16GB-VRAM GPU. HOW GAME MODE WORKS — the same local LLM extracts a GameSpec from your prompt, then Studio assembles a complete, offline, self-contained three.js game with Rapier physics: third-person character controller, NPC entities with wander/follow AI, collectible objectives, boundary walls, HUD and win screen. Character animations (idle/walk/run) are baked from real mocap data onto auto-rigged models. Every export passes a static verification gate — structure, libraries, JS syntax, animation integrity — before it's served. Builds in 30–60 seconds, runs anywhere a browser runs, and a Godot 4 export adapter is already headless-validated, with Unreal on the roadmap. Both modes share one asset system: procedural world-dressing props, a curated healed asset library, and auto-fetch fallbacks. And both ship real files — MP4, editable scene source, GIF, PNG sequences, playable game folders. Nothing is locked in a cloud account because there is no cloud: no API keys, no per-render fees, no rate limits, no data leaving your machine. Now shipping as a native desktop app (Tauri 2 shell) with the browser studio still available. Windows-first; Linux supported via upstream setup. Commercial-safe by construction: built entirely on permissively licensed components (TRELLIS.2, three.js, Rapier — MIT/Apache; SDXL under OpenRAIL++). Ship two attribution lines in your credits ("Built with DINOv3" and "© OpenStreetMap contributors") and everything you generate — videos and games — is yours to sell. Built for creators, indie filmmakers, YouTubers, marketers, and now indie game devs who want playable prototypes from a sentence. Best today at single-subject cinematic shots and compact objective-based games; multi-subject composition and deeper game genres are the explicit next phase, not hidden limitations. Source-visible under BSL 1.1 (auto-converts to Apache 2.0 in four years), built in public, solo dev. Part of FantasyLab.ai — local-first AI tools where the model directs real software and the user owns the output.