IP_Composer/README.md

1. Composition Generation script

Running the Script

Use the following command:

python generate_compositions.py --config path/to/config.json --create_grids

Parameters

• --config: Path to the configuration JSON file.
• --create_grids: (Optional) Enable grid creation for visualization of the results.

Configuration File

The configuration file should be a JSON file containing the following keys:

Explanation of Config Keys

• input_dir_base: Path to the directory containing the base images.
• input_dirs_concepts: List of paths to directories containing concept images.
• all_embeds_paths: List of .npy files containing precomputed embeddings for the concepts. The order should match input_dirs_concepts.
• ranks: List of integers specifying the rank for each concept’s projection matrix. The order should match input_dirs_concepts.
• output_base_dir: Path to store the generated images.
• prompt (optional): Additional text prompt.
• scale (optional): Scale parameter passed to IP Adapter.
• seed (optional): Random seed.
• num_samples (optional): Number of images to generate per combination.

2. Text Embeddings Script

This repository also includes a script for generating text embeddings using CLIP. The script takes a CSV file containing text descriptions and outputs a .npy file with the corresponding embeddings.

Running the Script

Use the following command:

python generate_text_embeddings.py --input_csv path/to/descriptions.csv --output_file path/to/output.npy --batch_size 100 --device cuda:0

Parameters

• --input_csv: Path to the input CSV file containing text descriptions.
• --output_file: Path to save the output .npy file.
• --batch_size: (Optional) Batch size for processing embeddings (default: 100).
• --device: (Optional) Device to run the model on.