Du, YilunHe, Hairong2026-07-0720262026-06-022026He, Hairong. 2026. Bringing Dunhuang Dance to Life. Bachelors Thesis, Harvard University Engineering and Applied Sciences.32451812https://dash.harvard.edu/handle/1/42743979How did people dance over a thousand years ago? Across the Dunhuang Caves (4th–14th centuries) in Northwestern China, hundreds of mural paintings depict celestial dancers in mid-motion — arms extended, torsos twisted, ribbons suspended in the air. These depictions capture breathtaking poses, but the movement between them, where dance truly lives, has been lost. This thesis asks how we can decode dynamic motion from static visual records. Unlike some other cultures, including India and Japan, ancient Chinese dance does not have a living tradition passed down across generations. The movement style itself is not recoverable from the images alone. This project therefore draws on all available resources — religious and secular, visual and verbal — to explore the range of possibilities for how the dances depicted in the murals might have looked in motion. The thesis contains both expository and original components. Chapters 1 and 2 are primarily expository. Chapter 1 examines the visual imagery of dance in Dunhuang murals as both religious imagination and records of medieval secular dance culture. It argues that dance imagery in the Dunhuang murals fits into the narrative of music and dance as an important kind of religious offering, while also reflecting the dance traditions of contemporary social life from the court to popular culture, especially in the Tang Dynasty. Chapter 2 turns to textual sources, especially Buddhist texts and Tang poetry. It develops a historically grounded vocabulary of movement and style that can help decode the visual imagery of dance. Together, these chapters show that the murals record the artists’ sustained effort to take the almost impossible challenge of capturing and encoding dynamic movement in two-dimensional static images, and that textual descriptions of dance can help culturally reimagine and reconstruct these motions. The original contribution of the thesis lies in Chapter 3. As I elaborate there, my technical contributions are threefold: first, I formulate Dunhuang dance reconstruction as a sparse-pose and text conditioned motion generation problem and introduce an end-to-end pipeline for producing possible, expert-evaluated in-between motions from historical pose pairs; second, I construct a high-quality, original Dunhuang dance pose dataset with provenance-aware grouping, textual pose descriptions, and character-consistent pose image renderings for conditioned large-scale generation; and third, I propose a test-time motion selection framework that combines 3D kinematic scoring with historically informed style-based ranking, which improves the quality of the reconstructions in both physical plausibility and style alignment with the Dunhuang dance culture. More broadly, this project opens up new possibilities for dancers, historians, museums, and the broader public to engage with this performance tradition both as cultural heritage and a living catalyst for artistic creation today.application/pdfenComputer scienceThesis or Dissertation2026-07-07