1. Context and issues

This communication explores a corpus of ancient Indian bas-reliefs using computer vision methods to provide a detailed and comparative iconographic description of selected motifs. The objective is twofold. First, it seeks to highlight the artistic and stylistic developments of these motifs–often still understudied–to reaffirm their significance within Indian art history. Second, it assesses whether lightweight computational methods can yield meaningful results on small heritage corpora without relying on large-scale models, and whether such an approach can be generalised. These two goals structure the study: computational tools support the iconographic analysis, and the case study evaluates the relevance of a frugal workflow adapted to small-scale datasets. As it has been noted, a gap persists between human interpretation and what can be automatically extracted (Van Noord 2022), and no existing method directly addresses this gap for fine-grained iconographic analysis. This work therefore positions itself within methodological discussions on the articulation of distant and close viewing (Moretti 2013; Arnold & Tilton 2019), where combining large-scale analysis with expert interpretation is understood as a means to mitigate the limitations of each approach.

The study draws on 3,066 photographs of Indian bas-reliefs, from which 13,283 motifs were extracted and grouped into six categories: lotus, elephant, lion, goose, horse and bull. These are among the most frequently represented elements across Indian monuments, regardless of period or religious affiliation. The analysis focuses on the earliest phases of monumental stone art in India, from the 3rd century BCE to the early classical period in the 3rd century CE (Huntington & Huntington 2014). It examines how these patterns became part of the architectural vocabulary of the period. Describing them requires enough precision to capture stylistic choices expressed through form, proportion, frequency, and position in the composition. Given the relatively small number of images (compared with the large datasets usually used in computer vision), this communication raises broader questions about the challenges of computational analysis on small-scale datasets in a field largely unexplored within the Digital Humanities, namely Indian Studies (Risam & Gairola 2021).

2. Limitations of current approaches: a need for a frugal alternative

Recent works in Digital Art History often rely on fine-tuning foundation models for object detection and segmentation. Yet, such approaches face structural limitations when applied to heritage corpora. The diversity of forms and styles, the density of sculpted scenes, and the state of conservation of monuments all reduce the effectiveness of current models (Bengamra et al. 2024). Additional constraints include the scarcity of available annotations and the limited size of corpora, which force analysis into broad categories and weaken the iconographic value of results (Ciecko 2020). Even state-of-the-art vision models are trained on datasets with label taxonomies too coarse-grained for the level of iconographic detail required here. In practice, this results in a single label per category, whereas the analysis requires finer distinctions.

Three further issues arise. First, models trained on highly specific datasets are difficult to reuse and fine-tuning them is both energy-intensive and technically demanding (Romein et al. 2020). Second, the investment in annotation and training often seems disproportionate when results remain generic labels that add little value to the analysis and ultimately obscure important stylistic nuances (Wasielewski 2023). Third, these models still lack interpretability, limiting their relevance for detailed artistic analysis (Offert 2018). These constraints motivate a more measured strategy: the approach combines pre-trained zero-shot models in a lightweight workflow that avoids costly annotations and additional training entirely. As a point of reference, processing 250 images requires approximately 40 seconds of inference time, which suggests a substantially lower computational footprint than fine-tuning. The goal is an ecologically responsible pipeline that produces explainable results genuinely informing art-historical interpretation.

3. Proposed methodology: combining clustering and multimodality

To address these challenges, this communication proposes a hybrid approach designed to enrich the semantic description of patterns. Motifs were first segmented automatically using a detection model (YOLOv11 fine-tuned on an annotated set of 400 images). These segmentation results enable the identification of iconographic elements across all images in the corpus. Zero-shot methods are then applied to a selected subset of photographs for more detailed analysis.

The workflow combines unsupervised clustering with a multimodal vision-language model (VLM). Clustering groups patterns into coherent subsets, while the VLM links each pattern and its components to textual descriptions. This combination–which appears to constitute an original methodological contribution to Digital Art History–makes it possible to explore sculpted representations through complementary analytical lenses, while ensuring that expert interpretation remains central throughout.

4. Clustering and visual proximities

The analysis begins with the vectorisation of motifs using the DINOv2 model, which embeds them in a high-dimensional space based on visual similarity (Oquab et al. 2024). These embeddings are then grouped using the HDBSCAN algorithm to identify coherent clusters of visually related motifs without requiring a predefined number of classes (Cafiero & Camps 2020). To make the structure of these clusters more interpretable, embeddings are subsequently projected into a two-dimensional space using UMAP, which facilitates visual inspection and interpretation (Chavooshi 2024).

Each cluster brings together motifs sharing visual traits, revealing stylistic, morphological, or compositional affinities that might otherwise remain unnoticed. Although DINOv2 was trained on generic categories, our approach identifies visual similarities without enforcing predefined labels. Still, any categorisation remains hypothetical and must be confirmed by expert knowledge (Ballesteros Zapata 2022). Close viewing–supported by interactive navigation through the cluster visualisation–allows researchers to assess motifs in each cluster, identify visual regularities, and refine or correct proposed categories as needed. Complementary statistical analysis tests whether observed regularities are representative of an entire cluster or limited to a few examples, strengthening the robustness of the results by distinguishing genuine patterns from local anomalies (figure 1).

5. Multimodality and semantic enrichment

The second part of the analysis uses SigLIP2, a multimodal model chosen for its strong performances (Tschannen et al. 2025). Its role is to associate extracted patterns with textual descriptions, by embedding them in a shared space where their semantic proximity is computed as cosine similarity between their respective embeddings. For each pattern, the model ranks descriptions drawn from a structured iconographic taxonomy, enabling progressively finer characterisation of each motif. The taxonomy is adapted from IconClass (Posthumus & Brandhorst 2005), used as a structural backbone and extended to accommodate Indian visual traditions, with categories fed directly to SigLIP2 without any model training.

The shared embedding is treated not as a stand-alone interpretative medium but as a comparative aid that remains under expert control through a posteriori validation of the associations produced. Unlike clustering, which groups motifs by visual similarity, multimodality aims to relate visual forms to conceptual and iconographic categories, supporting semantic enrichment and enabling the production of nuanced descriptions suited to broader comparative and diachronic analyses (figure 2).

6. Results and methodological contributions

The results highlight both the potential and the limits of the proposed approach. Clustering groups motifs into 11 to 68 main clusters depending on the motif, each further subdivided into sub-clusters to capture smaller iconographic groupings. The proportion of outliers varies from 5% to 31% for the most iconographically diverse motifs. Formal quantitative coherence metrics (e.g. silhouette scores, Davies-Bouldin index) yield uninformative results given the distribution of these embeddings, which reflects the complexity of the corpus rather than a methodological limitation. For instance, clusters frequently group motifs by their position within larger scenes–even though surrounding context had been removed–revealing shared compositional conventions previously difficult to demonstrate empirically. On the VLM side, SigLIP2 achieves high zero-shot accuracy for visually salient descriptors: 93% for wings attributes across animal categories, 86% for elephant’s trunk position, 60% for overall posture, and 48% for orientation. The strong results on the first two categories demonstrate its effectiveness for iconographic analysis without any fine-tuning, while the lower scores on orientation identify precisely where expert correction is most needed. This suggests that similar iconographic configurations generate visual similarity between patterns, revealing shared stylistic or compositional conventions that previously remained difficult to demonstrate.

This variability reflects the hybrid and iterative nature of the method: computation accelerates processing, while domain expertise ensures accuracy and interpretative depth. The main methodological contribution lies in demonstrating that fine-grained iconographic analysis can be carried out on small-scale datasets by combining lightweight computer vision methods with expert-driven evaluation, without requiring additional model training.

7. Conclusion

This communication rethinks automated analyses of heritage images by prioritising interpretative precision over broad generic categories. The aim is to produce detailed descriptions suited to iconographic inquiry and aligned with specific research questions. This approach makes it possible to describe in detail stylistic developments and representational choices that are difficult to capture through qualitative methods alone, while maintaining the interpretative depth required in art history. It also shows that robust results can be supported by lightweight computational methods, provided they rely on their integration within a critically informed analytical framework.

While distant viewing remains effective for processing entire corpora and detecting recurring elements, it can prove insufficient for fine-grained analyses. Conversely, close viewing alone restricts large-scale comparison and obscures broader stylistic patterns. The approach advocated here combines both perspectives to produce rich and scalable results from small corpora–by computer vision standards, yet larger than those traditionally studied in the Humanities. Beyond our case study, this methodology can be transposed to other corpora with appropriate adjustments to the iconographic taxonomy. Small-scale datasets, such as those of ancient Indian patterns, should therefore not be regarded as marginal but as opportunities to refine methodological expectations in computer vision, where the articulation of automation and disciplinary expertise can produce deeper and more situated knowledge than either approach might yield alone.

Figures

Figure 1 – Sample clusters of lion motifs using DINOv2 and HDBSCAN

Figure 2 – SigLIP2 semantic ranking for an elephant motif