Visual Authenticity Workbench

AI image authenticity analysis and model evaluation platform.

Problem

AI-generated images are now common in content moderation, media verification, legal review, and asset management workflows. A single detector score is not enough for a real application: teams also need upload records, model versions, thresholds, latency, image hashes, history, and reports that clearly explain the limits of the result.

This project turns open-source AI image detectors into an engineering system. It does not claim to be a universal truth engine.

Scope

MVP supports image-only detection:

• Upload one JPG, PNG, or WebP image.
• Store the file and image metadata.
• Create a detection task in a Java backend.
• Call a versioned Python model-service contract. Local development uses the heuristic runtime; real weights and CUDA verification are server-pending.
• Store model output, threshold, version, latency, and image hash.
• Generate a report and detection history.
• Show model health and registry information.

Out of scope for MVP:

• Video deepfake detection.
• Audio or text detection.
• Legal-grade forensic certification.
• User billing, tenants, RBAC, or complex audit workflows.
• Custom model training or stacking meta-learners.

Video and expert-ensemble visuals are development showcases, not formal product capabilities. See the product capability matrix for implementation evidence and deferred boundaries.

Architecture

React + TypeScript frontend
  -> Spring Boot Java backend
    -> PostgreSQL
    -> local file storage
    -> Python FastAPI model service
      -> Nonescape Mini image detector

The Java backend owns business logic. Python services own model inference.

Tech Stack

• Frontend: React 18, TypeScript, Vite, Framer Motion
• Backend: Java, Spring Boot, Maven, Spring Data JPA, Flyway
• Database: PostgreSQL
• Model service: Python, FastAPI, PyTorch, safetensors
• Deployment: Docker Compose

Planned Models

MVP model:

• Nonescape Mini: https://github.com/e3ntity/nonescape

Later comparison models:

• Nonescape Full
• ClipBased-SyntheticImageDetection: https://github.com/grip-unina/ClipBased-SyntheticImageDetection

Development

Run the local environment check first:

powershell -ExecutionPolicy Bypass -File tools/check-local-env.ps1

Full local stack with Docker Compose:

docker compose -f infra/docker-compose.yml up --build

Frontend:

npm install
npm run dev
npm run build

Java backend:

cd backend-java
mvn test
mvn spring-boot:run

Model service:

cd model-services\nonescape-mini
..\..\.venv-model-service\Scripts\python -m uvicorn app.main:app --host 127.0.0.1 --port 5010

Detailed Windows, WSL, and SSH-server setup notes are in docs/local-development.md. For the evaluation workflow demo, follow docs/fullstack-evaluation-demo.md.

Documentation

Start with docs/README.md for the documentation map, project worklog, improvement roadmap, and writing standards.

Reliability Note

Detection results are auxiliary signals. They should not be used as the sole basis for high-stakes decisions.