✳ HOMEPORTFOLIO · 2026
← BACK TO INTERACTIVEPROJECT / 01
INTERACTIVE / 2026

Nanxiang Murals.

An Evidence-Informed Human–AI Workflow for Digital Preservation

An aerial view of Nanxiang Village, Guangxi.
FIG. 01 / NANXIANG MURALS An aerial view of Nanxiang Village, Guangxi.
PROJECT TYPE
First Author
DATE
2026
MEDIUM & METHODS
Digital Heritage / Human–AI Collaboration / Generative AI
FIRST AUTHOR · ACCEPTED AT ACM MULTIMEDIA ASIA 2026
01 / FIELD CONTEXT

A living archive, facing loss.

Murals in Nanxiang Village, Guangxi, are part of the buildings and everyday environments that hold them. Fading pigments, surface loss, and fragmented documentation make their preservation an urgent challenge.

This research asks how generative AI can support digital preservation while keeping the evidence, uncertainty, and human decisions behind each reconstruction visible.

Historic buildings within the village fabric.
FIG. 02Historic buildings within the village fabric.
Architectural context and field documentation.
FIG. 03Architectural context and field documentation.
02 / FIELD DOCUMENTATION

Begin with the surviving evidence.

The field corpus contains 1,303 high-resolution image files documenting murals and their architectural context. Collection-level descriptions offer historical context, but do not establish what belonged inside every missing region.

Source documentation is kept separate from generated candidates. Recording the setting and surviving paint provides a basis for later reference selection and review.

Documentation activity inside a historic building.
FIG. 04Documentation activity inside a historic building.
Lighting equipment used during on-site documentation.
FIG. 05Lighting equipment used during on-site documentation.
03 / THE MURAL ARCHIVE

Different subjects. Shared questions.

Architectural scenes, flowers and birds, landscapes, and narrative figures require different kinds of evidence. These source photographs show the variety of the documented paintings; they are not AI-generated reconstructions.

SDT-06-01-02 · Source mural photograph.
FIG. 06SDT-06-01-02 · Source mural photograph.
SDT-03-02-01 · Source mural photograph.
FIG. 07SDT-03-02-01 · Source mural photograph.
DFD-01-03-01 · Source mural photograph.
FIG. 08DFD-01-03-01 · Source mural photograph.
DFD-01-02-03 · Source mural photograph.
FIG. 09DFD-01-02-03 · Source mural photograph.
04 / RESEARCH QUESTION & WORKFLOW

What makes a completion defensible?

A visually plausible image does not tell us which sources support it. The workflow treats each generated candidate as a visual hypothesis, linking it to source images, a mask, an independently selected reference, explicit constraints, generation settings, and a human decision.

  1. 01

    Document

    Record images, context, condition, and field sources.

  2. 02

    Diagnose

    Define the masked region and identify uncertainty.

  3. 03

    Constrain

    Select a reference; state the evidence, allowed continuations, and forbidden additions.

  4. 04

    Generate

    Keep model settings, prompt versions, and random seeds traceable.

  5. 05

    Verify & preserve

    Accept, revise, or reject; archive candidates and their review history.

Allow

Continue features supported by surviving boundary paint and the selected evidence.

Forbid

Do not add unsupported people, inscriptions, buildings, or symbolic objects.

Uncertainty

Revise the evidence record or defer generation when support is insufficient.

05 / CONTROLLED COMPARISON

Comparing visual hypotheses.

Ten murals, three generation conditions, and three fixed seeds produced 90 candidates. The comparisons below show artificially masked inputs, context-only generation (A), visual-reference guidance (B), and evidence-informed generation (C). Complete originals were used only for evaluation.

SDT-10-01-01 · Controlled landscape comparison.
FIG. 10SDT-10-01-01 · Controlled landscape comparison.
Artificially masked inputA · Context onlyB · Visual referenceC · Evidence-informedComplete target · Evaluation only
DFD-01-02-03 · Reference and constraints refined before generation.
FIG. 11DFD-01-02-03 · Reference and constraints refined before generation.

From Figure 2 of the paper; fixed seed 1144340710. The floral example uses a reference and constraints refined before generation. These are controlled experiments, not records of physical restoration.