README.md

metadata

title: CineGen CPU
emoji: 🖼
colorFrom: purple
colorTo: red
sdk: gradio
sdk_version: 5.44.0
app_file: app.py
pinned: false
short_description: automate the process of short movie creation
tags:
  - mcp-in-action-track-creative

CineGen AI Director is an AI agent designed to automate the process of short movie creation. It transforms a simple text or image idea into a fully realized video production by handling scriptwriting, storyboard generation, character design, and video synthesis using a multi-model approach.

• Sponsor Platforms: Uses Google Gemini (story + character prompts) and Hugging Face Inference Client with fal.ai hosting for Wan 2.2 TI2V video renders;
• Autonomous Agent Flow: StoryGenerator → CharacterDesigner → VideoDirector pipeline runs sequentially inside a single Gradio Blocks app, with MCP-friendly abstractions (StoryGenerator, CharacterDesigner, VideoDirector) designed for tool-call orchestration.
• Evaluation Notes: Covers reasoning (Gemini JSON storyboard spec), planning (scene/character tables that feed downstream steps), and execution (queued video renders with serialized HF jobs).

Artifacts for Reviewers

• Social Media Proof: Replace <SOCIAL_LINK_HERE> with your live tweet/thread/LinkedIn post so judges can verify community sharing.
• Video Recording: Upload a walkthrough of the Gradio agent (screen + narration) and swap <DEMO_VIDEO_LINK> with the shareable link.

🚀 Key Features

• End-to-End Automation: Converts a single sentence idea into a complete short film (approx. 30s-60s runtime).
• Intelligent Storyboarding: Breaks down concepts into scene-by-scene visual prompts and narrative descriptions.
• Character Consistency System:
  • Automatically identifies main characters.
  • Generates visual reference sheets (Character Anchors).
  • Allows users to "tag" specific characters in specific scenes to ensure visual consistency in the video generation prompt.
• Multi-Model Video Generation: Supports multiple state-of-the-art open-source video models via Hugging Face.
  • Robust Fallback System: If the selected video model fails (e.g., server overload), the system automatically tries alternative models until generation succeeds.
• Interactive Editing:
  • Edit visual prompts manually.
  • Add, Insert, or Delete scenes during production.
  • Regenerate specific clips or character looks.
• Client-Side Video Merging: Combines individual generated clips into a single continuous movie file directly in the browser without requiring a backend video processing server.

🤖 AI Models & API Usage

The application orchestrates two primary AI services:

1. Google Gemini API (@google/genai)

Used for the "Brain" and "Art Department" of the application.

• Logic & Scripting: gemini-2.5-flash
  • Role: Analyzes the user's idea, generates the title, creates character profiles, and writes the JSON-structured storyboard with visual prompts.
  • Technique: Uses Structured Output (JSON Schema) to ensure the app can parse the story data reliably.
• Character Design: gemini-2.5-flash-image
  • Role: Generates static reference images for characters based on the script's descriptions.
  • Role: Acts as the visual anchor for the user to verify character appearance before video generation.

2. Hugging Face Inference API (@huggingface/inference)

Used for the "Production/Camera" department.

• Video Generation Models:
  • Wan 2.1 (Wan-AI): Wan-AI/Wan2.1-T2V-14B (Primary/Default)
  • LTX Video (Lightricks): Lightricks/LTX-Video-0.9.7-distilled
  • Hunyuan Video 1.5: tencent/HunyuanVideo-1.5
  • CogVideoX: THUDM/CogVideoX-5b
• Provider: Defaults to fal-ai via Hugging Face Inference for high-performance GPU access.