Posttiva

Multi-Agent System for Automated Advertisement Poster Generation

Posttiva is an AI-driven advertising assistant that turns a furniture product image and a small set of user preferences into a complete advertisement poster. The system coordinates specialized Image, Text, and Layout agents to preserve the real product, generate a context-aware background, write marketing copy, and compose the final poster.

Academic project: Bachelor of Artificial Intelligence Sciences, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University (PNU), Riyadh, 2026.

Demo

System Architecture

The pipeline is organized into three cooperating agents:

1. Image Agent

• SAM-HQ ViT-L for human-in-the-loop product segmentation.
• Stable Diffusion 2 Inpainting for generating a new interior background around the isolated furniture item.
• Product-preservation and quality checks are used to keep the uploaded item visually unchanged while regenerating the surrounding scene.

2. Text Agent

• GPT-4o-mini Vision extracts structured product attributes such as type, material, color, and style.
• GPT-4o-mini generates a headline, short description, and call-to-action based on both visual attributes and user campaign settings.
• Claude Sonnet orchestrates the text workflow and can trigger regeneration when quality requirements are not met.

3. Layout Agent

• Qwen2-VL-7B-Instruct reasons over the composed image to select text zones, typography, colors, and placement.
• A custom rendering engine creates the final poster while avoiding product/text overlap.
• CLIP ViT-L/14 scores visual quality and Claude Opus performs final poster evaluation/selection.

Key Results

The final system was evaluated on 50 furniture products:

User Flow

1. Create an account / sign in.
2. Upload a furniture image and optionally a brand logo.
3. Select the target product interactively with SAM-HQ.
4. Configure the advertisement type, product focus, room, style, lighting, discount/urgency, and poster frame.
5. Generate and approve advertisement text.
6. Generate a product-preserving background.
7. Compose and evaluate the final poster.
8. Customize, save, and download the result.

Tech Stack

AI / ML: Python, PyTorch, Hugging Face Transformers, Diffusers, SAM-HQ, Stable Diffusion 2 Inpainting, GPT-4o-mini, Claude, Qwen2-VL, CLIP, OpenCV, scikit-learn

Backend: FastAPI, Express.js, PostgreSQL

Frontend: HTML, CSS, JavaScript

Repository Structure

.
├── Posttiva/                 # Frontend SPA
│   ├── index.html
│   ├── app.js
│   ├── api.js
│   └── config.js
├── Posttiva_Backend/         # FastAPI + AI multi-agent pipeline
│   ├── agents/
│   │   ├── image_agent.py
│   │   ├── text_agent.py
│   │   └── layout_agent.py
│   ├── models/loader.py
│   ├── main.py
│   └── requirements.txt
├── db-server/                # Express/PostgreSQL auth + project persistence
│   ├── server.js
│   ├── db.js
│   ├── schema.sql
│   └── package.json
└── assets/                   # README visuals

Local Setup

Prerequisites

• Python 3.10+ (3.11 recommended)
• Node.js 18+
• PostgreSQL
• An NVIDIA CUDA-capable GPU is strongly recommended and is required by the current CLIP/layout code path.
• OpenAI and Anthropic API credentials

The first AI run downloads large pretrained model weights, including the SAM-HQ checkpoint and Qwen/Stable Diffusion resources.

1. Clone the repository

git clone <your-repository-url>
cd posttiva

2. Configure the AI backend

cd Posttiva_Backend
python -m venv .venv

Activate the environment, then install dependencies:

pip install -r requirements.txt

Copy the environment template:

cp .env.example .env

Add your API keys to .env, then start FastAPI:

uvicorn main:app --host 0.0.0.0 --port 8000

Health check: http://localhost:8000/health

3. Configure PostgreSQL and the database server

Create a PostgreSQL database named posttiva, then apply the included schema:

psql -d posttiva -f db-server/schema.sql

Install Node dependencies and create the environment file:

cd db-server
npm install
cp .env.example .env
node server.js

The database API runs at http://localhost:3001 by default.

4. Run the frontend

In another terminal:

cd Posttiva
python -m http.server 8080

Open http://localhost:8080.

The checked-in config.js points to the local AI backend (:8000) and database backend (:3001). For deployment, replace those values with your hosted endpoints.

API Overview

AI backend (FastAPI)

Database backend (Express)

Provides registration/login, password reset, profile updates, project save/update/delete, feedback, and logout endpoints under /api/....

Notes

• Model checkpoints, generated images, caches, .env files, and node_modules are intentionally excluded from Git.
• The current FastAPI implementation keeps AI pipeline state in memory and is designed around the academic/demo workflow rather than multi-user production serving.
• The project is currently focused on furniture advertisements. Future work identified in the study includes additional product categories, multi-product posters, and Arabic/multilingual support.

Team

Developed as a graduation project by:

• Reema Saad Alkathiry
• Layan Hisham Bin Shaheen
• Shouq Khalid Aldossari
• Leen Thabet Khashugji

Supervised by Prof. Norah Alghamdi.

Reema's Contributions

• Image Agent development, including product segmentation and background-generation work.
• AI output evaluation/testing and evaluation-interface work.
• Contributions to frontend integration and the end-to-end project workflow.

Academic / Portfolio Use

This repository presents the implementation of a university graduation project. Pretrained models and external APIs remain subject to their respective licenses and terms. No separate open-source license is granted by this repository unless the project team adds one explicitly.