ChildDiffusion: Unlocking the Potential of Generative AI and Controllable Augmentations for Child Facial Data using Stable Diffusion and Large Language Models

TL;DR

ChildDiffusion is a novel framework that leverages Stable Diffusion and Large Language Models to generate diverse, high-quality synthetic child facial images with controllable features, addressing privacy concerns and enabling downstream AI applications.

Contribution

The paper introduces a high-level generative framework for creating customizable, photorealistic child facial datasets using text prompts and image transformations, with validation through ethnicity classification tasks.

Findings

Generated diverse child facial images with high realism

Successfully created a synthetic ethnicity dataset of 2.5k samples

Enhanced child ethnicity classification accuracy using synthetic data

Abstract

In this research work we have proposed high-level ChildDiffusion framework capable of generating photorealistic child facial samples and further embedding several intelligent augmentations on child facial data using short text prompts, detailed textual guidance from LLMs, and further image to image transformation using text guidance control conditioning thus providing an opportunity to curate fully synthetic large scale child datasets. The framework is validated by rendering high-quality child faces representing ethnicity data, micro expressions, face pose variations, eye blinking effects, facial accessories, different hair colours and styles, aging, multiple and different child gender subjects in a single frame. Addressing privacy concerns regarding child data acquisition requires a comprehensive approach that involves legal, ethical, and technological considerations. Keeping this in view this framework can be adapted to synthesise child facial data which can be effectively used for numerous downstream machine learning tasks. The proposed method circumvents common issues encountered in generative AI tools, such as temporal inconsistency and limited control over the rendered outputs. As an exemplary use case we have open-sourced child ethnicity data consisting of 2.5k child facial samples of five different classes which includes African, Asian, White, South Asian/ Indian, and Hispanic races by deploying the model in production inference phase. The rendered data undergoes rigorous qualitative as well as quantitative tests to cross validate its efficacy and further fine-tuning Yolo architecture for detecting and classifying child ethnicity as an exemplary downstream machine learning task.