Live now · v1.0 · source-available
fantasy-studio — preview● REC

Describe it.
Play it or film it.

Type a sentence — Fantasy Studio builds a playable 3D game with real physics, or directs a cinematic film in Blender. No code, no cloud, no per-render fees. Real meshes, real Cycles renders, on your machine. Yours forever.

Install on Windows · one line, PowerShell
irm https://raw.githubusercontent.com/bgrut/fantasy-studio/main/bootstrap.ps1 | iex
Full install guide (Win / Linux) →

// no cloud. no credits. no limits.

Building in public · follow on GitHub
~60s
Game Build
~90s
Film Render
0
Cloud · Credits
Yours Forever
fantasy-studio — polar bear cinematic render
90 seconds — type prompt, get cinematic render, no cloud.
Type a sentence.
Play the world.
Describe a game and Fantasy Studio builds it — a playable third-person world with real physics, objectives, enemies and pickups. Then edit it live: point at anything and type the change. No engine, no code, no cloud.
a fox on a dangerous mountain trek — collect 6 food, avoid the wolves, reach the shelter arctic survival — defeat 3 rivals, race to the beacon foxfire quest — gather 6 fireflies, reach the shrine

Prefer a movie over a match? Studio also directs cinematic films in Blender →

A guided tour of
the studio.
Type a prompt, AI directs, Blender renders — all local, all yours. Here's the full UI.
Fantasy Studio main canvas
Step 01

The Studio

Your canvas. Type any scene. The director plans casting, camera, lighting, and pacing.

Render tier selector
Step 02

AI Direction

Pick rendering quality. Iterate fast in Quick Preview, then commit to a Final Cinematic hero render.

Cast panel — pick or swap assets
Step 03

It Builds Your Cast

SDXL paints your subject from the prompt; TRELLIS.2 sculpts it into a textured, rigged 3D character. Prefer a library asset? Browse 300+ curated, healed assets as the fallback.

Scene controls
Step 04

Direct Every Detail

Lighting presets, camera modes, duration, brand color, template selection — all controllable.

Refine panel
Step 05

Refine With Words

"Lower the camera." "Switch to golden hour." "More dramatic." Talk to the director — the scene iterates.

Built like a
cinematic brain.
Every layer — from local AI direction to Cycles ray-tracing — engineered to produce studio-quality output from a single prompt.
🎬

Cycles Ray Tracing

Photorealistic path-traced rendering. Not AI hallucinated pixels — real light simulation with global illumination, caustics, and volumetrics.

🧠

AI Director (Local LLM)

A local Gemma 3 12B (via Ollama) plans every creative decision — camera, lighting, atmosphere, animation. Falls back to a deterministic director if the LLM is unavailable.

💻

100% Local Execution

Your GPU. Your renders. Your data. Zero cloud dependencies. No credits. No throttling. No subscription required to render.

🧬

Generates Your Subject in 3D

SDXL paints an identity-faithful reference from your prompt; TRELLIS.2 converts it to a textured GLB, auto-rigged and animated. A 300+ curated library plus Objaverse/Sketchfab fetch are the fallback path.

💬

Refine With Words

After your first render, adjust the scene with natural language. "Lower the camera." "Switch to golden hour." "More dramatic." The director iterates without starting over.

📦

Real Exports

MP4 video, animated GIF, PNG sequence, and the source .blend file. Re-light, re-frame, re-render in Blender forever.

Inside the AI.
Real engineering, not magic. Four local AI subsystems — paint your character's image, sculpt it into 3D, then direct and verify the film — each hand-tuned, debuggable, and running on your machine.
subsystem · 01

Scene Director

Local Gemma 3 12B via Ollama plans every scene. Picks camera, lighting, mood, and pacing from a one-line prompt. Falls back to a deterministic rule-based director when the LLM is unavailable, so renders never block on inference.

model: gemma3:12b · runtime: ollama · fallback: deterministic
subsystem · 02

Image Generator

A local SDXL model paints an identity-faithful reference image of your character straight from the prompt — on your GPU, no cloud. ControlNet-depth locks a clean pose, so a samurai stays a samurai.

model: SDXL Base 1.0 + ControlNet-depth · runtime: local GPU
subsystem · 03

Mesh Generator

A local TRELLIS.2 model lifts that image into a fully textured 3D character — a real GLB, sculpted from scratch. Mesh-hygiene + orientation gates clean it; a 300+ curated library is the fallback path.

model: TRELLIS.2-4B · output: textured GLB · fallback: library
subsystem · 04

Verification Gates

Every render runs seven structural checks before frames hit disk: bbox sanity, frustum, framing fill, primitive detection, orientation, grounding, hero tag. Bad outputs caught and aborted with a structured error — not silently shipped to you.

checks: 7 · abort: bbox_sane · oriented_correctly
From prompt to
production.
01

Prompt

Type anything. "Ferrari on a mountain road at golden hour."

02

AI Director

Local LLM (Gemma 3) extracts the scene — subject, motion, camera, lighting, mood.

03

Generate the Subject

SDXL paints an identity-faithful reference; TRELLIS.2 turns it into a textured 3D mesh. (Curated library is the fallback.)

04

Rig & Render

Auto-rigged, animated, lit — Blender Cycles renders it for real. MP4 + .blend out.

Not another
AI video tool.
AI Video GenWeb 3D ToolsFantasy Studio
Render qualityAI hallucinationsReal-time WebGPUCycles ray-traced (real light)
DeterministicNoYesYes
Runs locallyNoNoYes
Usage limitsCreditsCreditsUnlimited
Camera controlMinimalSomeCinematic
Export .blend sourceYes
Output tierConsumerWeb prototypeStudio CGI
Best for: playable single-subject games + cinematic shots (V1). Multi-subject composition is V2 work.
Live status.
Being built right now, in real time, on real GitHub. These tiles update from the public API.
GitHub stars
github.com/bgrut/fantasy-studio
Latest commit
checking…
Open issues
tracking the V1 polish list
Public source release
Live
source-available now
Where we're going.
Honest about timeline being aspirational. Quality compounds — we ship Now and Next before getting clever.
Now
Live now

V1 is live.

  • Playable game mode — physics, objectives, live editing
  • Single-subject cinematic films (MP4 + .blend)
  • Public source release (BSL 1.1)
  • Export to Godot 4 · share links
  • Local SDXL + TRELLIS.2 asset generation
Soon
V1.5 – V1.9 · 3–6 months

Capability spread.

  • Cloud render tier (BYO-GPU optional)
  • Sound effects + ambient audio
  • Voice prompt input
  • Sharing / portfolio features
  • Patreon-only premium templates
Future
V2.0+ · 6–12 months

The hard stuff.

  • Marketplace launch (creator revenue share)
  • Multi-subject scene composition
  • Character consistency across scenes
  • Unreal Engine renderer backend
  • AI Director v2 — treatment → casting → blocking
// timelines aspirational · dependent on traction · feedback shapes priority

Follow the build

Ready to direct?

It's live. Grab the source on GitHub — or drop your email for build updates.