VideoStudio AI
PERSONAL PROJECT · FULL-STACK & AI APPLICATION
VideoStudio AI
A modular platform for creating, voicing and organising AI-generated videos from a prompt or a reference image.
The project connects a PHP web application to several video engines, speech services and local or cloud tools through one bilingual and secure user journey.
Multi-model
ComfyUI, Runpod and cloud providers are exposed through one capability-based catalogue.
Asynchronous pipeline
A project queue and dedicated worker handle long-running video workloads.
End-to-end creation
Images, video, narration, voices, music, assembly and media management live in one product.
The challenge
AI video tools expose very different contracts, formats and constraints. VideoStudio AI provides a consistent interface that hides this complexity without locking the application into a single model.
For the user
Choose an intent, a compatible model and an output format, then follow generation without handling the underlying APIs.
For the developer
Add a provider or model through stable interfaces without rewriting the business workflow.
An interface driven by model capabilities
The creation form automatically adapts the available formats, resolutions and required media to the selected engine. It supports text-to-video, image-to-video, short clips and stories assembled from several scenes.
One creation form controls different local and cloud engines.
From prompt to final video
- Describe the project. The user supplies a prompt, a reference image and, when needed, an audio track or narration.
- Validate capabilities. The catalogue checks the model, aspect ratio, resolution and required inputs.
- Queue the work. The project is saved immediately; the browser does not wait for generation to finish.
- Generate and assemble. The worker calls the provider, builds multi-scene Shorts and performs post-processing with FFmpeg.
- Deliver the result. The video, thumbnail, status and actionable errors are available in the private media library.
Search, grid or list display, previews and project status in one library.
Generation runs in the background with visible progress and actionable errors.
Implemented capabilities
Text and image to video
Multiple models, aspect ratios, resolutions and execution locations.
Multi-scene Shorts
Ollama storyboarding, sequential generation and transition-based assembly.
Talking characters
Audio-driven animation and lip synchronisation from a reference image.
Narration and audio
Speech synthesis, performance styles, music and automatic FFmpeg mixing.
Voice library
Microphone recording, authorised sample import and voice candidate design.
Local operations
Worker and Ollama monitoring from the administration area, with silent Windows startup.
Voice as a first-class part of the product
The platform goes beyond animated imagery. It also manages narration, private voice samples and voice design, with explicit consent required before any cloning operation.
This feature connects in-browser audio capture, file validation, speech services and owner-scoped storage.
Import, microphone recording and consent checks in one flow.
Architecture and technology choices
| Layer | Role in the project |
|---|---|
| PHP 8.2+ · Natora Core | Modular Router → Middleware → Controller → Service → Repository architecture. |
| Smarty 5 · JavaScript | Bilingual, responsive and accessible interface under a strict CSP. |
| MariaDB · PDO | Projects, voices, status and ownership persisted through prepared queries. |
| Asynchronous worker | Locking, heartbeat, controlled retries, cancellation and actionable failures. |
| ComfyUI · Runpod | Interchangeable providers for local or on-demand cloud model execution. |
| Ollama · FFmpeg · TTS | Storyboarding, multi-scene assembly, audio mixing, narration and voice design. |
Key engineering decisions
A provider-independent core
Provider interfaces isolate external APIs. Business rules remain usable without knowing Runpod, Gemini or ComfyUI.
Private media by design
Sources, voices and renders live outside the public directory. Every read operation revalidates ownership.
Machine-specific configuration
Sensitive settings remain in untracked .local.php files, separate from application defaults.
An observable experience
The interface exposes queueing, processing, failure and cancellation instead of hiding long-running operations.
What this project demonstrates
- End-to-end product design, from an open requirement to UX and operations.
- Integration of heterogeneous AI APIs behind stable application contracts.
- Asynchronous processing, state management, large files and remote failure diagnosis.
- Attention to security, voice consent, access control and private storage.
- French and English product development with responsive behaviour down to mobile.
- Ability to broaden the product scope without introducing unnecessary architectural complexity.
An evolving project
VideoStudio AI is a working application that I continue to expand around one consistent goal: support a broad range of video creation workflows while keeping the architecture understandable and maintainable.