Synthetic Data Augmentation for Multi-Task Chinese Porcelain Classification: A Stable Diffusion Approach

TL;DR

This work tackles data scarcity in archaeological porcelain classification by using LoRA-tuned Stable Diffusion to generate synthetic images that augment a real dataset for a four-task CNN (dynasty, kiln, glaze, type). Through controlled experiments with mixed real-synthetic data at ($95:5$) and ($90:10$) ratios, the study finds task-dependent gains, with the strongest improvements in type and dynasty classifications, while glaze (texture) can degrade due to limitations in texture synthesis. A key contribution is a rigorous quality-control and prompt-engineering framework that demonstrates both the potential and limits of synthetic data in digitizing cultural artifacts, along with practical guidelines for when synthetic augmentation is warranted and how to balance diversity with authenticity. The findings indicate a threshold around $8$–$9 ext{ extpercent}$ synthetic data for noticeable gains and emphasize morphology-focused improvements, urging caution for texture-critical tasks while offering a path toward more robust, domain-aware synthetic data strategies in archaeology.

Abstract

The scarcity of training data presents a fundamental challenge in applying deep learning to archaeological artifact classification, particularly for the rare types of Chinese porcelain. This study investigates whether synthetic images generated through Stable Diffusion with Low-Rank Adaptation (LoRA) can effectively augment limited real datasets for multi-task CNN-based porcelain classification. Using MobileNetV3 with transfer learning, we conducted controlled experiments comparing models trained on pure real data against those trained on mixed real-synthetic datasets (95:5 and 90:10 ratios) across four classification tasks: dynasty, glaze, kiln and type identification. Results demonstrate task-specific benefits: type classification showed the most substantial improvement (5.5\% F1-macro increase with 90:10 ratio), while dynasty and kiln tasks exhibited modest gains (3-4\%), suggesting that synthetic augmentation effectiveness depends on the alignment between generated features and task-relevant visual signatures. Our work contributes practical guidelines for deploying generative AI in archaeological research, demonstrating both the potential and limitations of synthetic data when archaeological authenticity must be balanced with data diversity.

Figures (18)

• Figure 1: The Chinese porcelain glaze colour categories from the dataset
• Figure 2: The Chinese porcelain form categories from the dataset
• Figure 3: The Chinese porcelain pattern categories from the dataset
• Figure 4: Hierarchical Label System for Porcelain Classification
• Figure 5: Adaptive Dataset Splitting Strategy Based on Combination Size