Generative AI

AI Content Factory

Automated video production pipeline based on multiple orchestrated generative AI models.

Status: In production
Category: Generative AI
Reference: Case #6

From briefing to finished video clips: specialized AI models for text, image and video generation are combined into an automated, controllable production process.

Technical challenge: Multiple specialized generative AI models were to be orchestrated into an automated, repeatable video production process – instead of operating models individually and manually.
Architecture: A German story is broken down by an LLM (Ollama qwen2.5vl) into a storyboard of three clips; an image model (Qwen-Image-Edit) generates start/end images with consistent continuity, and a first-last-frame-to-video model (Wan2.1-FLF2V-14B) generates the clips on a GPU instance.
Engineering challenges: Consistent characters across the clips (end image N = start image N+1), asynchronous processing despite long LLM runtimes, scale-to-zero GPU operation and orchestration via SQS instead of synchronous API calls.
Demonstrated capabilities: Generative AI, model orchestration, workflow automation and production pipelines with an asynchronous job model and GPU operation.

The AI Content Factory (Storyboard Generator) is a Riegel Systems engineering case in the field of generative AI. The core idea: multiple specialized AI models are not “operated” individually, but orchestrated into an automated production pipeline.

The pipeline

Pipeline

Briefing
Storyboard
Scene Planning
Start Image
End Image
Video Prompt
Image-to-Video
Final Clips

Implemented in practice

A storyboard of three 6-second clips is generated automatically from a German story:

Story → Storyboard: An LLM (Ollama qwen2.5vl) breaks the story down into three scenes and generates, per clip, a start-image prompt, end-image prompt, motion prompt and negative prompt – as strict, schema-validated JSON.
Image generation: An image-editing model (Qwen-Image-Edit-2511, Apache-2.0) generates the four images with consistent continuity – the end image of one clip is also the start image of the next.
Image-to-video: A first-last-frame-to-video model (Wan2.1-FLF2V-14B) generates the final clips from start image, end image and prompt on a GPU instance (L40S) with scale-to-zero.
Approval workflow: The previews are reviewed and edited in an admin UI before the video jobs are triggered.

AI Content Factory: Orchestrated story-to-video pipeline from the LLM storyboard to the final clips — Orchestrated pipeline: LLM → storyboard → images → approval → image-to-video

Engineering focus

Model orchestration: Each model takes on the task it is best suited for – text analysis, image editing, video generation.
Asynchronous job model: Long LLM and GPU runtimes are decoupled via SQS instead of being squeezed into restrictive API timeout windows.
Prompt generation: Automatically generated, precise prompts drive the models deterministically instead of manually.

Key takeaway

Production-ready generative AI systems are systems integration: normalize inputs, validate outputs, handle failures and monitor costs. The Storyboard Generator shows how individual models become a reliable, repeatable production process.

Technologies & concepts

A similar engineering project?

If your project needs comparable technical depth, we discuss feasibility, architecture and effort – concretely rather than in pitch format.

Discuss your project