Background: This study tackles the challenge of reconstructing ephemeral wartime exhibitions in 1940s China from extremely sparse and degraded visual records. Photographic exhibitions during the Second Sino-Japanese War were pivotal propaganda forums, yet traditional historiography struggles to recreate how these displays looked and felt to contemporaries given the fragmentary archives. By addressing this archival scarcity, our work seeks to illuminate how images were arranged and experienced in situ, thereby shedding new light on spectatorship and collective visual memory under the extreme conditions of war. We approach this problem with an interdisciplinary  digital historical forensics methodology that fuses advanced computer vision techniques with deep archival research. While our case studies center on WWII China, the framework is designed to be scalable and applicable to any historical context where visual evidence is minimal, thus pushing the boundaries of digital humanities research on lost or partially preserved cultural experiences.

Methodological Innovation:  At the heart of our approach is a three-phase   digital historical forensics  pipeline that integrates AI-driven image processing with humanistic verification. The pipeline’s innovation lies in combining state-of-the-art super-resolution and image-matching algorithms with rigorous historical analysis to reconstruct exhibition content that would otherwise remain speculative. The three phases are as follows:

Figure 1. Hierarchical Exhibition Reconstruction Pipeline. Schematic of our three-phase workflow: (1) image enhancement via super-resolution, (2) iterative image matching with Hierarchical Region Pursuit, and (3) human-in-the-loop historical contextualization.

• AI Super-Resolution Enhancement: We employ deep learning techniques—specifically latent diffusion models (LDMs)—to upscale and restore low-quality historical photographs. Many surviving exhibition photos are tiny, blurry, or poorly lit, far below the resolution needed for analysis. Using LDM-based super-resolution, we progressively denoise and sharpen these images, effectively recovering details and textures that standard upscaling (bilinear or bicubic interpolation) cannot provide. This step ensures that even images initially only a few dozen pixels across are clarified to a level where both human researchers and machine algorithms can discern meaningful features. Importantly, our use of diffusion models aims to preserve documentary integrity while hypothesizing plausible high-frequency details in the absence of any ground-truth high-resolution originals.
• Hierarchical Region Pursuit (HRP) Image Matching:  We introduce a novel image search algorithm called   Hierarchical Region Pursuit, built on a transformer-based detector-free matching network (ASpanFormer). This algorithm forensically identifies which original photographs are depicted on the walls in the exhibition snapshots by cross-referencing them against a large archival database of wartime images. Crucially, HRP is designed to handle the partial, oblique, or occluded views of photos in exhibition documentation that confound traditional whole-image matching techniques (like SIFT or CNN feature matching). In our exhibition photos, multiple pictures are often clustered in collages or photographed at angles; HRP’s region-by-region pursuit was able to match photographs based on partial glimpses – effectively finding a needle in a haystack by examining each straw. Experiments confirm that HRP dramatically increases the recovery of correct matches: a baseline without HRP identified far fewer images, whereas with HRP enabled we matched   many more  photographs correctly to their archival originals. In essence, HRP serves as an adaptive sliding-window search guided by AI, substantially boosting both effectiveness and efficiency of image retrieval.
• Human-in-the-Loop Verification and Contextualization:  The final phase is an iterative cycle of expert historical analysis that scrutinizes and interprets the algorithmic outputs. Rather than treating the AI matches as definitive, we treat them as   hypotheses  that must be corroborated with contextual evidence. For each proposed match (i.e. an identified photograph that was purportedly on display), we consult a range of historical sources: archival metadata and provenance records, wartime newspapers and exhibition reviews, as well as secondary scholarship on the events or figures depicted. This cross-referencing serves multiple purposes. It verifies the match’s plausibility (does the photo’s known date, location, and content align with the exhibition’s theme and timeline?) and it situates the image within the broader historical narrative. In cases where our pipeline flags a surprising correspondence—such as a photo appearing in an exhibition dated earlier than the photo’s supposed creation date—this discrepancy is investigated rather than assumed to be an error. Such anomalies can lead to new questions for historians (e.g. a misdated archive entry or previously unknown reuse of an image).

Performance and Technical Outcomes:  Our integrated pipeline yielded substantial improvements over baseline approaches, validating the benefits of its AI innovations. Using a test set of six archival photographs of exhibitions (containing 78 identifiable images on display), we compared identification success under various settings. Without any upscaling, only a small fraction of the displayed images could be identified. Classic upscaling (bilinear/bicubic) improved this somewhat, but the transformer matcher still struggled with tiny, blurry inputs. By contrast, applying diffusion-based super-resolution followed by HRP produced a striking boost in retrieval performance. In quantitative terms, our best configuration (LDM super-resolution   with  HRP) correctly identified nearly   70%  of the exhibition images within the top-10 suggestions – a notable increase from the ~60% achieved by the next-best setup (conventional upscaling with no HRP). This ~10 percentage-point gain in recall confirms that the hallucinatory high-frequency details added by diffusion upscaling, while detrimental to naive global matching, become   valuable cues  for our partial-region matching approach. Moreover, because our system ranks candidate matches rather than issuing a binary verdict, historians can attain near-perfect precision by focusing on the top-ranked suggestions. In practice, this means the pipeline not only finds more correct matches, but also presents its results in a way that researchers can easily verify and trust. Equally important, we adhered to transparent research practices: all AI-enhanced images and matches are clearly marked and documented, allowing us to draw a firm line between raw empirical evidence and the plausible reconstructions inferred from it. This commitment to transparency (echoing principles like the London Charter in digital heritage) ensures that our reconstructions, while innovative, remain anchored in verifiable data and responsible interpretive methods.

Case Studies and Historical Insights:  We applied this digital forensic pipeline to reconstruct significant portions of two high-profile photographic exhibitions from wartime China:   The Memorial of Our International Friend Norman Bethune (1940) and the   Second Jinchaji Border Region Art Festival  (1941). From a handful of surviving photographs of these events, our method recovered the specific images that were on display and approximated their layout in the exhibition space. This enabled us, for the first time, to analyze these exhibitions with a level of granularity previously unattainable. We could map out how dozens of photos were arranged on walls and panels, identify thematic groupings, and even conjecture the sequence a visitor might have followed through the exhibit. For example, the   Jinchaji Art Festival  exhibition, reconstructed through our pipeline, revealed a deliberate juxtaposition of photographs depicting female militia training, agrarian support efforts, and frontline battles – suggesting a narrative of   gendered mobilization  and collective effort in the war. By identifying each image and its placement, we discerned how organizers used visual storytelling to galvanize public sentiment: they interwove scenes of heroism, everyday resilience, and international alliance to inspire viewers emotionally and politically. These findings illustrate that wartime exhibitions were not static arrays of propaganda, but carefully orchestrated “image-scapes” designed to engage viewers as active participants in the war effort. Our spatial reconstructions allow us to hypothesize, for instance, where moments of collective pause or heightened emotional impact might occur (such as a particularly dramatic photograph placed at a focal point on the wall). By comparing the reconstructed content with contemporary reports (e.g. newspaper descriptions of the exhibitions), we found a high degree of consistency, lending further credence to our methods. In some cases, the computational results even led us to revise historical assumptions – for example, clarifying which photos were actually present at an event, thereby correcting misattributions in the archival record.

Figure 2. Cross-referencing Matches in Databases

Significance and Interdisciplinary Impact:  This work showcases a new paradigm for digital humanities scholarship in the face of archival scarcity. We transform fragmentary image archives from being insurmountable dead-ends into solvable puzzles, using digital methods to reassemble dispersed visual evidence. Notably, we move beyond mere image enhancement for its own sake; instead, the aim is to   recover the spatial, thematic, and ideological architecture  of past exhibitions – in other words, to reconstruct the stories that these images told together, in context, to their original audiences. The result is a richer historical interpretation that remains grounded in empirical data while embracing informed conjecture where evidence is thin. Our study thereby responds to calls within the humanities for innovative methodologies to address “archival silences” and partially erased cultural histories. By integrating art-historical analysis, archival research, and machine learning techniques, we were able to generate insights unattainable by any single discipline in isolation. This collaboration not only yielded concrete historical knowledge (e.g. identifying exhibition content and curatorial intent) but also serves as a   proof-of-concept  for interdisciplinary research design. The methodological transparency and critical scrutiny we applied set an ethical standard for such work, ensuring that digital interventions remain accountable to historical truth and scholarly interpretation. Furthermore, our approach is scalable and adaptable: it opens up possibilities for reconstructing other lost visual experiences—whether propaganda exhibitions in different regions, derelict art installations, or even   other media  like deteriorated films, murals, or monuments using similar forensic techniques.