Painted pottery is among the most iconic cultural material remains of China’s Neolithic Age. Beyond serving as tangible evidence of early aesthetic consciousness, it provides crucial clues for reconstructing prehistoric cultural sequences and investigating cultural exchange and ethnic interaction. The diverse forms and intricate decorative patterns of painted pottery unearthed across China together constitute a vast “prehistoric visual archive.”

‘Single-image 3D reconstruction’

Among existing digital approaches, active 3D scanning and multi-view stereo reconstruction are the main methods for acquiring high-precision three-dimensional data on cultural artifacts. Although these technologies are relatively mature, their application faces clear bottlenecks. Both are essentially object-dependent techniques that require direct contact with artifacts or close-range photography. This entails high equipment costs, time-consuming data collection, and strict demands on lighting conditions and surface materials. Against this backdrop, the concept of “single-image 3D reconstruction” emerged.

Technical pathway

Using AI to infer the 3D shape of an object from a single photograph was once considered a formidable challenge, but recent years have seen remarkable progress in this field. Whether these methods can be effectively applied to painted pottery with complex decorative patterns, however, has remained uncertain, and it has also been unclear which method performs best.

To address these issues, we evaluated and fine-tuned several representative single-image 3D reconstruction techniques and applied them to painted pottery. The results indicate that a hybrid strategy—combining a diffusion-based multi-view image synthesis model with a large-scale feed-forward reconstruction network—achieves the highest reconstruction fidelity.

Within this technical pathway, the system first learns patterns of geometric transformation under different lighting conditions and viewing angles by training on large datasets of 3D objects, a process known as “learning geometric priors.” When a two-dimensional photograph of painted pottery is input, the diffusion-based generative model does not simply process pixels but engages in probabilistic prediction. Drawing on learned patterns, it estimates how the object might appear from side, rear, and top-down perspectives, generating a set of synthesized multi-view images. These images are then processed by the feed-forward reconstruction network, which performs stereo merging through spatial algorithms to produce a 3D mesh model that integrates both geometry and surface texture. The entire process requires no manual intervention and can be completed within seconds or minutes.

Crucially, this technology goes beyond simple image stitching. It represents a form of deep-learning-based “knowledge reasoning” that allows the system to handle occluded regions and logically infer unseen portions of an object. In this sense, it marks an important step in archaeological digitization, moving from “passive recording” toward “intelligent generation.”

Practical outcomes

To test the applicability of this approach in archaeological research, we selected representative samples from the Complete Collection of Painted Pottery Unearthed in China for experimentation. The samples ranged from simple rounded-bottom bowls to complex large jars with double handles.

The results show that AI performs with remarkable accuracy in reconstructing geometric forms. The resulting 3D models were able to faithfully restore vessel shapes, making them suitable for relatively reliable analyses of morphological evolution and typological sequencing.

Pattern reconstruction, by contrast, displays a combination of “visual plausibility” and residual “detail” uncertainty. For areas visible in the original photograph, the model produces high-fidelity 3D mappings. For occluded areas, it intelligently extends and fills in patterns based on stylistic features inferred from the visible surfaces. While these reconstructions are coherent and visually convincing—well suited to digital museum exhibitions and public education—they require caution when used in specialized academic research. In such contexts, it is essential to distinguish clearly between original data and generated data.

The technology also continues to face difficulties when applied to highly asymmetrical or severely damaged artifacts. This underscores that AI-based single-image 3D reconstruction of cultural relics cannot yet fully replace traditional object-dependent recording methods. Instead, it should serve as a supplementary tool to bridge the gap between physical artifact research and two-dimensional documentation.

Tu Dongdong is an assistant professor from the Institute for Humanities at ShanghaiTech University.