Chitra: GenAI Video Pipeline
Multi-stage GenAI pipeline that converts text articles into short-form videos using LLM summarization, scriptwriting, image generation, AI avatars, and human-in-the-loop QA
Overview
AI Chitra was an experimental content production system designed to answer a specific question: can we reliably convert text-first content (news articles and story prompts) into high-quality, publishable videos with a largely automated workflow? The project combined LLM-driven understanding and writing with generative visuals and avatar-based presentation, backed by human review before publishing.
Motivation
High-quality video creation is expensive and slow. Most content pipelines require scripting, creative direction, asset sourcing, editing, and publishing—often taking hours per video. AI Chitra explored whether a multi-stage GenAI pipeline + targeted human checkpoints could reduce cost/time while maintaining quality for:
- News shorts from current articles
- Narrative storytelling (e.g., India mythology & history for kids) in a consistent style
Product
- Article ingestion & summarization: selected articles parsed and summarized into structured key points
- Script generation: LLM-written scripts optimized for short-form pacing and clarity
- Visual prompt generation: image prompts generated per script segment/scene
- Generative image creation: scene images via Midjourney and DALL·E
- AI avatar anchoring: avatar-based presenter for an "anchor-style" delivery
- Human-in-the-loop quality gates:
- Verify narrative correctness and flow
- Curate/override generated images where needed
- Final QA before publishing
Technical Stack
- LLM layer: OpenAI + Gemini for extraction and summarization; OpenAI for scriptwriting + image prompt generation
- Generative visuals: Midjourney + DALL·E for scene assets
- Avatar/video tooling: AI avatar platform to generate presenter segments
- Orchestration: ingest → summarize → script → prompt → images → avatar → assemble → QA → publish
Deployment
- Operated as a semi-automated internal production workflow
- Outputs were manually published after QA
- Supported two tracks (news + storytelling) with repeatable "content runs"
Outcomes
- Bijli News: 52 videos, 40K+ views in 30 days
- Chitra Stories: 4 videos, 1K+ views