100% Free · 100% Local · Apple Silicon

Kids' Educational
Videos, on Autopilot

Gurukul AI turns any topic into a Pixar-style animated educational video — script, images, narration, animation, subtitles — fully local, zero cloud costs.

Get the Code

How It Works Locally

Open http://localhost:7860 after running python app.py

localhost:7860
1. Generate Script2. Scene Images3. Narration4. Animate5. Assemble6. Subtitles⚡ /selfimprove
photosynthesis
✓ Web research first
10 scenes
🧠 Generate Script with Gemma
Stage 0: Researching topic online...
→ Wikipedia: Photosynthesis is...
→ DuckDuckGo: 5 facts found
Stage 1: Generating with Gemma 3 4B...
Island: Photosynthesis Forest Island
Scene 1: Aerial view of glowing green...
Scene 2: A single leaf magnified...
✓ Script saved (10 scenes)
localhost:7860 — ⚡ /selfimprove
coin slowly flipping in golden light
Scene
3
Min score
7.0
Max tries
3
▶ Run Single Model
Stage 1 [Director] — Gemma 4...
Expanded: A golden coin tumbles slowly...
Stage 2 [Creator] — ltx-2b attempt 1
Generated: scene3_ltx2b_a1.mp4 (341KB)
Stage 3 [Critic] — Qwen2.5-VL...
Motion: 6 Visual: 7 Prompt: 8 Cinema: 7
Overall: 7.0 ✓ Accepted!
Saved to training dataset.

From Topic to Video in 6 Steps

Everything runs on your Mac. No API keys. No subscriptions.

🌐01ResearchDuckDuckGo + Wikipedia facts
🧠02ScriptGemma 3 4B writes 10 scenes
🖼️03ImagesFLUX Dev generates landscapes
🔊04AudioKokoro TTS narration
🎬05AnimateLTX / Wan2.2 video models
✨06PolishSubtitles + xfade transitions
⚡
/selfimprove — Agentic Pipeline
Auto-escalates models until quality passes
🎭 DirectorGemma 4 expands prompt
→
🎬 CreatorGenerates video
→
🔍 CriticQwen2.5-VL scores 1–10
→
🔄 RefinerAuto-retry better model
→
✨ PolisherTopaz 4K upscale
Model escalation
LTX-2B → Wan2.2 Fun-5B → LTX-13B → Wan2.2 I2V-A14B
Quality threshold
Configurable min score (default 7.0/10)
Output
Training dataset + model leaderboard

9 Animation Models Built In

From instant Ken Burns to 20-min cinematic Wan2.2 — pick your quality/speed tradeoff

Ken BurnsINSTANT
ffmpeg zoom+pan, zero GPU
< 5s/scene★★☆☆☆
LTX Video 2BFAST
ComfyUI, great for iteration
~40s/scene★★★☆☆
LTX Video 13B
High quality landscapes
~11min/scene★★★★☆
LTX-2.3 22B GGUFNEW
Newest LTX, speed+quality
~4-6min/scene★★★★☆
Wan2.2 Fun-5B GGUFGGUF
Object motion, stable
~8-10min/scene★★★★☆
Wan2.2 I2V-A14B GGUFBEST
Hero shots, dual-stage
~15-20min/scene★★★★★

Stolen from OpenMontage. Made Free.

Three premium features, zero cost, fully local

🔤

Word-Level Subtitles

mlx-whisper transcribes on-device. Karaoke-style yellow highlight on current word. No OpenAI key.

WhisperX → mlx-whisper
🎞️

Animated Transitions

ffmpeg xfade between every scene — dissolve, wipeleft, slideright. No Remotion, no JS bundle.

Remotion → ffmpeg
🌐

Research-First Scripts

DuckDuckGo + Wikipedia facts fed to Gemma before writing. Episodes are accurate, not hallucinated.

Web research → free
🖥️

Optimised for Mac Studio M4 Max 36GB

All models tested on Apple Silicon MPS. Sequential generation to avoid memory spikes. FLUX Schnell for previews, FLUX Dev for finals. GGUF quantisation for stability over BF16.

Start Making Videos

Clone the repo, start the Gradio app, type a topic.

$ git clone https://github.com/LakshmiSravyaVedantham/gurukul-ai
$ cd gurukul-ai && pip install -r requirements.txt
$ python app.py # open http://localhost:7860
View on GitHub