OpenVINO Generative AI Scripts

This repository contains Python scripts for running and evaluating OpenVINO-based generative AI models.
They are adapted from official OpenVINO Jupyter notebooks and customized for our research objectives, including large-scale image generation and model quality benchmarking.

---

📌 Overview

• Purpose: Automate inference, image generation, and evaluation for multiple generative AI models using OpenVINO Runtime.
• Dataset: phiyodr/coco2017 (COCO 2017 prompts for text-to-image generation)
• Core Features:
  • Support for multiple models and precision formats (weights: int4, int8, fp16)
  • Automated folder creation for generated outputs
  • Large-batch inference: 300 images per run with prompts from COCO 2017
  • Pre-written evaluation scripts for IS, CLIP Score, FID, PickScore

---

🛠 Supported Models

Model Name	Variants (Weights)	Script Example
FLUX.1 Schnell	int4 / int8 / fp16	flux_int8.py
Stable Diffusion XL	int4 / int8 / fp16	sdxl_fp16.py
Stable Diffusion v1.5	int4 / int8 / fp16	stable-diffusion-v1.5-int4.py

⚠️ Special Note on SDXL Models

The official SDXL repository does not provide different weight formats by default.
You will need to manually download and export them with optimum-cli before running the scripts.

Example: Download SDXL int4 model

optimum-cli export openvino \
    --model stabilityai/stable-diffusion-xl-base-1.0 \
    --weight-format int4 \
    --dataset conceptual_captions \
    int4_sdxl/

After downloading, update the model_path in the corresponding Python script to point to your local folder, e.g.:

model_path = "int4_sdxl"

📊 Workflow

1. Run Model Script

   • Each script will:
     • Load the selected model in OpenVINO Runtime
     • Use prompts from phiyodr/coco2017
     • Generate 300 images per run
     • Save results in a timestamped folder under {model_name}/{weight}/images

2. Evaluate Generated Images

   • Use the scripts in evaluation/ to compute:
     • IS (Inception Score)
     • CLIP Score
     • FID (Fréchet Inception Distance)
     • PickScore