Progressive Prompt Detailing for Improved Alignment in Text-to-Image Generative Models (SCoPE)

Ketan Suhaas Saichandran*, Xavier Thomas*, Prakhar Kaushik, Deepti Ghadiyaram
(*Equal Contribution)

📄 Paper (arXiv)

This repository contains the code for SCoPE, a training-free method to improve prompt-image alignment in text-to-image diffusion models by progressively interpolating between coarse-to-fine prompt embeddings during the denoising process.

SCoPE can be easily applied on top of existing diffusion pipelines without any retraining.

View run.sh for an example to run our pipeline

🔧 Setting up t2v_metrics for VQA Scoring

To use the VQA scoring model (clip-flant5-xxl), follow these steps to install t2v_metrics in a dedicated conda environment.

📁 Step 1: Clone the repository

cd scorers/
git clone https://github.com/linzhiqiu/t2v_metrics
cd t2v_metrics

🐍 Step 2: Create and activate a new conda environment

conda create -n t2v python=3.10 -y
conda activate t2v

📦 Step 3: Install dependencies

conda install pip -y
pip install torch torchvision torchaudio
pip install git+https://github.com/openai/CLIP.git

🛠️ Step 4: Install t2v_metrics locally

pip install -e .

You’re now ready to use the VQA scoring functionality in get_scores.py.