forge/README.md

171 lines
7.9 KiB
Markdown
Raw Blame History

# FORGE AI: Unified Generative AI Platform
**FORGE AI** is an enterprise-grade, microservices-based platform designed to unify the world's most powerful generative AI models into a single, cohesive workflow for various media types. It provides a robust backend for orchestration and a modern, responsive frontend for creative professionals.
---
## 🌟 Executive Summary
Instead of managing subscription islands (Runway, Midjourney, Topaz, ElevenLabs), **FORGE AI** brings them all together. It allows for complex workflows like generating an image with **DALL-E 3**, upscaling it with **Topaz**, extending it into a video with **Google Veo**, and adding a voiceover via **ElevenLabs**—all within one interface.
---
## <20> Comprehensive Feature Matrix
### 1. 🎬 Video Generation (Multi-Provider)
The video module abstracts complexity between different providers, handling authentication, file upload, polling, and result retrieval automatically.
| Provider | Model | Capabilities | Optimal Use Case |
| :--- | :--- | :--- | :--- |
| **Runway** | **Gen-4 Turbo** | Image-to-Video | **High-Fidelity Animation**. Best for animating static marketing assets. Features **Smart Cropping** (auto-resize to 1280x768). |
| **Runway** | **Veo 3 / 3.1** | Text/Image-to-Video | **Versatile Generation**. Native 720p/1080p, 8-second clips. Good for general-purpose stock footage. |
| **Google** | **Veo Native** | Text-to-Video | **Enterprise Scale**. Direct Vertex AI integration for scalable generation. |
> **Key Feature**: **Smart Aspect Ratio Handling**. The backend automatically detects the required aspect ratio for the selected model (e.g., Gen-4's strict 1280:768 requirement) and resizes/crops input images on the fly to prevent API errors.
### 2. 🖼️ Image Generation (The "Omni-Model" Engine)
Access the latest models without switching tools.
* **OpenAI**:
* **GPT-Image-1**: The latest efficient model. Supports transparent backgrounds and variable quality.
* **DALL-E 3**: HD quality, vivid/natural styles.
* **Google Imagen**:
* **Imagen 4.0 (Standard/Ultra/Fast)**: Supports "Prompt Enhance" and "Person Generation" safety filters.
* **Stability AI**:
* **SD3.5 / SDXL**: Advanced control with **Negative Prompts** and **Image-to-Image** strength sliders.
* **Nano Banana (Gemini)**:
* **Gemini 2.5 Flash / 3 Pro**: High speed, supports up to 4K resolution and 21:9 aspect ratios.
* **Flux**: Black Forest Labs Flux Pro integration for photorealistic outputs.
* **Ideogram**: Version 2 integration, excellent for typography.
### 3. <20> Image Utilities
* **Professional Upscaling**:
* Integrated with **Topaz Photo AI SDK**.
* Capabilities: **Face Recovery**, **Denoising**, 2x/4x scaling.
* **Background Removal**:
* **Clipping Magic**: High-precision removal.
* **Bria AI**: Fast, commercially safe removal.
### 4. 🔊 Audio Intelligence
* **Text-to-Speech**:
* **ElevenLabs Integration**: Multilingual V2 model. High-quality voice synthesis.
* Configurable stability, similarity boost, and style.
* **Voice-to-Text**:
* **OpenAI Whisper**: Industry-leading transcription accuracy.
* Supports multiple languages and timestamp generation.
* **Sound Effects**:
* AI-generated SFX for video post-production.
### 5. 📝 Text & Utilities
* **Subtitle Processor**:
* Automated subtitle generation (VTT/SRT).
* **"Burn-in" Capability**: Hardcodes subtitles directly onto video frames using FFmpeg.
* **Prompt Studio**:
* Uses LLMs (Gemini/GPT-4) to refine simple prompts into detailed, artistic descriptions.
* Style presets: Cinematic, Anime, Photorealistic, etc.
* **Markdown/Mermaid**:
* Renders technical diagrams (Flowcharts, Gantt) from text descriptions.
---
## 🏗️ Technical Architecture
FORGE AI is built on a **Microservices Architecture** using **Docker Compose**.
### Backend (`forge-backend`)
* **Framework**: **FastAPI** (Python 3.11). High-performance, async-first API.
* **Task Queue**: **Celery** with **Redis**. Handles long-running jobs (video gen can take minutes) cleanly, decoupling HTTP requests from processing.
* **Database ORM**: **SQLAlchemy**. Interface for PostgreSQL.
* **Validation**: **Pydantic**. Strict usage of data contracts ensures API reliability.
* **File Handling**: Direct handling of binary assets, streaming uploads to disk storage.
### Frontend (`forge-frontend`)
* **Framework**: **Next.js 14** (App Router). Server-Side Rendering (SSR) for performance.
* **UI Library**: **React** + **Tailwind CSS**.
* **Components**: Custom design system using **ShadCN/UI** primitives (Dialogs, Selects, Toasts).
* **State Management**: React Hooks for polling job status and updating UI progress bars.
### Data Persistence
* **PostgreSQL 16**:
* `jobs`: Stores parameters, status, and API metadata for every request.
* `assets`: Tracks file paths, MIME types, metadata (width/height/duration).
* `users`: Authentication and profile data.
* **Redis**: In-memory message broker for Celery and caching layer.
* **Docker Volumes**:
* `postgres_data`: Persistent DB storage.
* `assets_data`: Shared volume for storing generated media files.
---
## ⚡️ Scalability & Performance
FORGE AI is architected to handle high user concurrency (e.g., 200+ simultaneous users) without degrading API performance.
* **Asynchronous Job Queue**: All heavy compute tasks (Video Generation, Upscaling) are offloaded to **Celery** workers via **Redis**. The API responds immediately with a `Queued` status, ensuring the interface remains snappy.
* **Horizontal Scaling**:
* The `forge-worker` service can be scaled horizontally to process more jobs in parallel.
* Command: `docker-compose up -d --scale worker=3` (Starts 3 worker containers).
* **Fault Tolerance**: If a worker crashes or an API fails, the job status is tracked in Postgres, and broken jobs can be automatically retried or flagged.
---
## 💾 Database Schema Overview
| Table | Description | Key Fields |
| :--- | :--- | :--- |
| `jobs` | The central unit of work. | `id`, `module`, `action`, `status` (pending/failed/completed), `input_data` (JSON), `api_provider` |
| `assets` | Files managed by the system. | `id`, `file_path`, `thumbnail_path`, `mime_type`, `source_job_id` |
| `users` | User accounts. | `id`, `email`, `hashed_password`, `role` |
---
## <20> Configuration (.env)
The system is highly configurable. Key variables include:
### API Keys (Critical)
* `RUNWAY_API_KEY`: For Gen-4 Turbo / Veo 3.
* `GOOGLE_API_KEY` / `GOOGLE_PROJECT_ID`: For Imagen / Vertex AI Veo.
* `OPENAI_API_KEY`: For DALL-E, GPT, Whisper.
* `ELEVENLABS_API_KEY`: For TTS.
* `TOPAZ_API_KEY`: For Upscaling.
### Infrastructure
* `DATABASE_URL`: Postgres connection string.
* `REDIS_URL`: Redis connection string.
* `CELERY_BROKER_URL`: Usually same as Redis.
---
## 🚀 Deployment & Operation
### Development Mode
```bash
# Start the full stack
docker-compose up -d --build
# View logs
docker-compose logs -f forge-backend
```
### Access Points
* **UI**: `http://localhost:3000`
* **API Documentation (Swagger UI)**: `http://localhost:8000/docs`
* **Database (Internal)**: Port `5432`
---
## <20> Troubleshooting Common Issues
### "Gen-4 Turbo - Validation Failed: Ratio"
* **Cause**: Runway's Gen-4 Turbo API is extremely strict about input image resolution (must be 1280x768).
* **Solution**: The backend now includes a **Smart Crop** pre-processor. It automatically resizes and crops your input to the exact pixel dimensions required. You do not need to manually edit images.
### "422 Unprocessable Entity"
* **Cause**: Usually missing required fields in the API payload.
* **Solution**: We recently relaxed the `prompt` requirement for Image-to-Video jobs. Ensure your frontend is sending the correct structure (refresh browser to clear cache).
---
## © 2025 FORGE AI Platforms
*Unified Creativity.*