synthetic-image-caption-generator 0.0.1

Tool to generate captions to use as input for pre-trained image generation models, aligned with some pre-existing captions.

Synthetic Image Caption Generator

A tool that uses Qwen to generate image captions similar to those in a given dataset. The model learns from example prompts and generates new captions that match their style and structure.

This tool is designed to be used for deep generative dataset augmentation.

Features

• Uses Qwen from HuggingFace🤗 Transformers
• Supports few-shot learning with configurable number of examples
• Reads prompts from .txt files in a dataset directory
• Flexible CLI with multiple configuration options
• Can generate multiple captions in one run
• Output to file or stdout

Installation

To install from PyPI:

pip install synthetic-image-caption-generator

To install locally:

pip install -e .

Or install dependencies manually:

pip install transformers>=4.30.0 torch>=2.0.0 accelerate>=0.20.0

Usage

Basic Usage

Generate a caption using 5 example prompts from the dataset:

generate-captions /path/to/dataset

Specify Number of Examples

Use a different number of example prompts (e.g., 10):

generate-captions /path/to/dataset --num-examples 10

Choose a Model

Select a different Qwen model (e.g., smaller or larger):

# Use the smaller 7B model
generate-captions /path/to/dataset --model qwen2.5-7b

# Use the larger 72B model
generate-captions /path/to/dataset --model qwen2.5-72b

# Use Qwen3 14B
generate-captions /path/to/dataset --model qwen3-14b

Generate Multiple Captions

Generate 5 captions:

generate-captions /path/to/dataset --num-generate 5

Specify Object Information

Include information about what's in the image:

generate-captions /path/to/dataset --object-info "the image contains 3 elephants"

generate-captions /path/to/dataset --object-info "the main crop in the field is soybean"

Save to File

Save generated captions to a file:

generate-captions /path/to/dataset --num-generate 10 --output generated_captions.txt

Advanced Options

generate-captions /path/to/dataset \
  --model qwen2.5-14b \
  --num-examples 8 \
  --num-generate 5 \
  --temperature 0.8 \
  --max-length 300 \
  --object-info "a cityscape with tall buildings" \
  --output captions.txt

Command-Line Arguments

• dataset_dir (required): Path to directory containing .txt files with caption prompts
• --model: Qwen model to use (default: qwen2.5-32b). Options: qwen2.5-0.5b, qwen2.5-1.5b, qwen2.5-3b, qwen2.5-7b, qwen2.5-14b, qwen2.5-32b, qwen2.5-72b, qwen3-14b, qwen3-32b
• --num-examples: Number of example prompts to provide to the model (default: 5)
• --num-generate: Number of captions to generate (default: 1)
• --output: Output file to save generated captions (optional, prints to stdout if not specified)
• --temperature: Temperature for text generation (default: 0.7, higher = more creative)
• --max-length: Maximum length of generated caption (default: 256)
• --object-info: Information about objects/content in the image (e.g., "the image contains 3 elephants")

Dataset Format

The dataset directory should contain .txt files, each with one or more prompts. Each prompt should be on its own line.

Example dataset structure:

dataset/
├── captions1.txt
├── captions2.txt
└── captions3.txt

Example content of captions1.txt:

A serene landscape with mountains in the background and a lake in the foreground
An urban street scene with people walking and cars passing by
A close-up portrait of a person smiling at the camera

Requirements

• Python >= 3.8
• CUDA-capable GPU recommended
• Sufficient GPU memory depending on chosen model:
  • Small models (0.5B-3B): 4-8GB VRAM
  • Medium models (7B-14B): 12-24GB VRAM
  • Large models (32B): 24-40GB VRAM
  • Extra large models (72B): 40GB+ VRAM or quantization required

How It Works

1. The script reads all prompts from .txt files in the specified directory
2. It randomly samples a specified number of prompts as examples
3. These examples are formatted into a few-shot prompt for Qwen2.5-32B-Instruct
4. The model generates a new caption that matches the style and structure of the examples
5. The generated caption(s) are output to stdout or saved to a file

License

MIT