ContentMachine

<div align="center">

❤️ Support This Project

An all-in-one AI pipeline for creating cinematic, documentary-style videos —
from a single topic to a fully packaged YouTube-ready project.

The Pipeline · Models & APIs · Real-World Cost · Getting Started · Features

</div>

🎥 Watch the Example

<p align="center"> <a href="https://www.youtube.com/watch?v=3BC8OXMzeF4"> <img src="https://img.shields.io/badge/▶_Watch_Full_Demo-YouTube-FF0000?style=for-the-badge&logo=youtube" alt="Watch Demo on YouTube"> </a> </p>

UI Demo

https://github.com/user-attachments/assets/2e440220-d31d-41e7-9adc-242fc97bd06b

---

API Status: I've personally tested this with Replicate and Gemini APIs — those are the battle-tested paths. fal.ai & elevenlabs support is implemented but not fully verified — it may have rough edges. PRs welcome!

---

What is ContentMachine?

ContentMachine automates the entire documentary video production workflow using state-of-the-art AI. Give it a topic, and it handles everything: researching real historical stories, planning scenes, generating images, creating video clips, writing narration scripts, generating voiceover audio, YouTube metadata, and thumbnails — all packaged into a clean ZIP ready for your video editor.

Built for content creators, documentarians, educators, and hobbyists who want to produce high-quality, cinematic content without a full production team.

I built this as a personal all-in-one pipeline — easy enough to run locally, flexible enough to swap AI providers, and powerful enough to produce publish-ready assets in one session.

---

The Pipeline

ContentMachine runs a step-by-step pipeline with a clean UI to monitor, pause, and resume at any stage.

Topic Input
    │
    ▼
┌─────────────────────────────────────────────────────┐
│  1. STORY GENERATION                                │
│     LLM finds 4 real, documented historical stories │
│     with cinematic potential → you pick one         │
└─────────────────────────────────────────────────────┘
    │
    ▼
┌─────────────────────────────────────────────────────┐
│  2. SCENE PLANNING                                  │
│     LLM builds a full cinematic shot list with      │
│     smart pacing: durations adapt per video model   │
└─────────────────────────────────────────────────────┘
    │
    ▼
┌─────────────────────────────────────────────────────┐
│  3. IMAGE GENERATION                                │
│     4 variations per scene (establishing, intimate, │
│     detail, atmospheric) — select the best one      │
│     All images saved as real PNG/JPG files in ZIP   │
└─────────────────────────────────────────────────────┘
    │
    ▼
┌─────────────────────────────────────────────────────┐
│  4. VIDEO GENERATION                                │
│     Image-to-video, 2 scenes at a time              │
│     Multiple models available — select best clip    │
│     Browse previous versions with ← → arrows        │
└─────────────────────────────────────────────────────┘
    │
    ▼
┌─────────────────────────────────────────────────────┐
│  5. AUDIO  (optional)                               │
│     ElevenLabs TTS narration + SFX per scene        │
└─────────────────────────────────────────────────────┘
    │
    ▼
┌─────────────────────────────────────────────────────┐
│  6. EXPORT                                          │
│     YouTube metadata · multi-select thumbnails      │
│     Full ZIP: videos + images/selected + images/all │
│     + audio + script + restorable project.json      │
└─────────────────────────────────────────────────────┘

Visual Style

The default aesthetic uses seamless glossy porcelain mannequins — figures always fully clothed in period-accurate outfits including explicitly named footwear (e.g. "iron-buckled brown leather knee boots"), no visible joints, stands, or supports. Photorealistic environments, ray tracing, cinematic lighting. A great starting point for YouTube-focused creators since it avoids depicting realistic scenes that may have been altered.

The visual style is fully customisable: expand Advanced — Customize System Prompts on the start page to edit the image prompt rules for any character type. Pair this with the Character Base Images feature (see below) to lock in a consistent look across every scene.

---

Supported Models & APIs

Note: Replicate and Gemini are the tested providers. fal.ai is a work in progress — contributions welcome.

LLM — Story, Scene Planning, Scripts, Metadata

| Provider | Models | |---|---| | fal.ai (WIP) | Claude 3.5 Sonnet | | Gemini (direct) | Gemini 3 Flash (recommended), Gemini 3.1 Pro, Gemini 3 Pro, Gemini 2.5 Flash, Gemini 2.5 Pro | | Replicate | Gemini 2.5 Flash, Gemini 3 Flash, Gemini 3.1 Pro, Claude 3.5 Sonnet |

Image Generation

| Provider | Models | |---|---| | fal.ai (WIP) | Flux Pro, Flux 2 Pro, Flux Schnell, Nano Banana Pro, Qwen Image 2512, Z-Image Base, Ideogram V3, SD 3.5 Large | | Replicate | Flux 2 Pro, Flux 1.1 Pro, Nano Banana Pro (Gemini), Imagen 4 | | Gemini (direct) | Gemini 3 Pro Image Preview (2K native output) |

Video Generation

| Provider | Model | Notes | |---|---|---| | fal.ai (WIP) | LTX-2 image-to-video | Not fully verified | | Replicate | LTX-2 Pro | With generated audio, 6–10s | | Replicate | LTX-2 Fast | 6–20s in 2s steps, favours 12–20s | | Replicate | Kling v3 | 3–15s, standard/pro mode, AI audio | | Replicate | Kling v2.5 Turbo Pro | 5s or 10s only |

Audio / TTS

| Provider | Capability | |---|---| | ElevenLabs | Scene-by-scene narration voiceover + SFX generation | | Local TTS | Bring your own (QWEN TTS, Kokoro, etc.) — zero cost |

---

Real-World Cost

A 4:30 minute documentary video produced with ContentMachine cost me approximately $28 USD.

| Component | Provider / Model Used | Notes | |---|---|---| | Story + Scene Planning + Scripts | Gemini 3 Flash Preview (Gemini API) | Very cheap | | Scene Images + Thumbnail | Nano Banana Pro / gemini-3-image-preview (Replicate) | Medium | | Video Clips | LTX-2 Pro (Replicate) | Largest cost driver | | Narrator TTS | QWEN TTS (local) | Free |

Tips to reduce cost:

• Use gemini-2.5-flash (non-preview) for LLM — higher quota, fewer rate limits
• Use fal.ai LTX-2 instead of Replicate LTX-2 Pro for cheaper video (once fal.ai is fully verified)
• Use Flux Schnell for faster, cheaper image generation
• Use a free local TTS tool for zero audio cost
• Use LTX-2 Fast on Replicate for longer scenes at a similar price point

---

Getting Started

Prerequisites

• Node.js 18+
• API keys for at least one LLM provider and one image provider

Install & Run

# Clone
git clone https://github.com/Saganaki22/ContentMachine
cd ContentMachine

# Install all dependencies
npm install

# Start both backend and frontend
npm run dev

App runs at http://localhost:5173. Backend API at http://localhost:3000.

Configure API Keys

Open the Settings panel (gear icon, top right). Paste your API keys — they are saved in your browser's localStorage and automatically pushed to the backend on each session startup. No .env file required for local use.

| Provider | Link | |---|---| | fal.ai | fal.ai/dashboard/keys | | Replicate | replicate.com/account/api-tokens | | Gemini | aistudio.google.com/api-keys | | ElevenLabs | elevenlabs.io/app/settings/api-keys |

Build for Production

npm run build
npm run start

---

Project Structure

ContentMachine/
├── backend/
│   ├── server.js                Express API server (200mb body limit)
│   └── routes/
│       ├── claude.js            LLM: stories, scene plans, prompts, scripts, metadata
│       ├── images.js            Image generation: fal.ai / Replicate / Gemini
│       ├── videos.js            Video generation + status polling
│       ├── elevenlabs.js        TTS narration + SFX generation
│       ├── thumbnail.js         Thumbnail image generation
│       ├── export.js            ZIP packaging (streams to browser)
│       ├── session.js           Auto-save sessions to output/ folder
│       └── settings.js