AI-Powered UGC Video Generation Platform

The Challenge

Botanix Batana Oil, an e-commerce company, faced a critical bottleneck in their marketing strategy: creating authentic user-generated content (UGC) videos for their product ads was prohibitively expensive and slow.

Traditional UGC Creation Costs:

• Beginner creators: $50-100 per video
• Intermediate creators: $200-500 per video
• Expert creators: $500-1,000+ per video

For a growing e-commerce business needing dozens of video variations for A/B testing across Meta ads, these costs quickly became unsustainable. Additionally, the turnaround time (days to weeks) made it impossible to respond quickly to market trends or test new creative angles.

The Solution: Intelligent Automation

I designed and implemented a fully automated AI video generation pipeline that creates realistic, UGC-style product videos at a fraction of the cost and time.

Platform Demo

Watch the platform in action: upload a product image, provide a creative prompt, and generate multiple UGC-style video variations automatically.

System Architecture

The system uses n8n as the orchestration engine, coordinating multiple AI services into a seamless pipeline:

1. User Input & Image Processing

• Custom frontend accepts product images and creative prompts
• Webhook trigger initiates the n8n workflow
• Images uploaded to Cloudinary for CDN hosting
• GPT-4o Vision analyzes the product image to extract:
  • Brand name and visual identity
  • Color schemes and design elements
  • Product features and context

2. Intelligent Scene Generation

• LangChain AI Agent with structured output parsing
• Generates realistic UGC-style video prompts based on:
  • Product analysis from GPT-4o
  • User’s creative direction
  • Best practices for authentic UGC content
• Creates multiple scene variations with specific:
  • Dialogue (what the person says)
  • Actions (how they interact with the product)
  • Camera angles (selfie-style, handheld, etc.)
  • Emotional tone and character personality
  • Setting and environment details

3. Parallel Video Generation

• Multiple scenes processed simultaneously for speed
• Sora 2 model for image-to-video generation
• Veo 3 model as alternative engine (not shown in workflow)
• Each scene rendered with realistic human movements
• Portrait orientation (9:16) optimized for social media

4. Automatic Video Assembly

• ffmpeg via Fal.ai API stitches scenes together
• Seamless transitions between clips
• Final output: cohesive UGC-style video ad
• Ready for immediate upload to Meta Ads Manager

Key Technical Innovations

Prompt Engineering Excellence

• Developed comprehensive system prompts for UGC-style realism
• Trained AI agent to understand product context and brand voice
• Structured output ensures consistent, parseable results
• Multi-scene storytelling with natural flow and continuity

Workflow Optimization

• Parallel video processing reduces total time to 7-9 minutes
• Conditional logic handles both uploaded images and URL inputs
• Automatic retries and error handling for API reliability
• Webhook-based async architecture prevents timeouts

Cost-Effective Architecture

• Pay-per-use API model keeps costs variable
• No infrastructure overhead (serverless n8n workflows)
• Cloudinary free tier for image hosting
• Strategic use of cheaper models where quality permits

Business Impact

Cost Transformation

• Before: $50-1,000 per video (human creators)
• After: $0.50-1.00 per video (AI generation)
• Savings: 98-99% reduction in production costs

Speed to Market

• Before: 3-7 days turnaround for creator delivery
• After: 7-9 minutes for complete video generation
• Impact: Test new creative concepts same-day

Scale & Flexibility

• Generate unlimited variations for A/B testing
• Quick iteration on messaging and visuals
• Rapid response to trending topics or seasonal campaigns
• Internal tool enabling marketing team autonomy

Technical Deep Dive

n8n Workflow Components

Webhook Trigger

• CORS-enabled endpoint for frontend integration
• Accepts JSON payload with prompt and image data
• Returns immediate acknowledgment while processing async

Image Handling Flow

• Conditional branching: uploaded file vs URL
• Base64 to binary conversion for uploads
• Cloudinary integration with preset configurations
• Normalized output format for downstream nodes

AI Agent Configuration

• LangChain integration with custom tools
• “Think” tool for complex reasoning steps
• Structured JSON output with scene arrays
• Error handling and validation

Video Generation Loop

• Split out scenes for parallel processing
• Batch processing with delays to respect API limits
• Status polling with exponential backoff
• Result aggregation for final assembly

Model Selection Strategy

Sora 2: Primary choice for:

• High-quality facial expressions and movements
• Better understanding of product interaction
• Consistent style across multiple scenes

Veo 3: Alternative for:

• Different aesthetic preferences
• Backup when Sora API has high load
• Cost optimization on high-volume days

GPT-4o: Essential for:

• Accurate product/brand analysis
• Vision capabilities for design extraction
• Reliable structured text output

Lessons Learned

1. Prompt engineering is critical - The quality of AI agent instructions directly impacts video realism and brand alignment

2. Parallel processing matters - Processing scenes simultaneously reduced total time from 30+ minutes to under 10 minutes

3. Error handling is essential - API rate limits, timeouts, and generation failures require robust retry logic

4. User testing drives refinement - Marketing team feedback led to improvements in scene variety and dialogue naturalness

Future Enhancements

• Voice synthesis integration for audio narration
• Brand voice fine-tuning for consistency
• Multi-language support for international markets
• Advanced analytics dashboard for performance tracking
• Template library for common product categories

---

Status: Live and actively used for internal content production

Client: Botanix Batana Oil

Deployment: n8n Cloud (workflow), Custom hosting (frontend)