The Potential of Digital Methods in Researching an Egyptian Temple: ML and HBIM at the Temple of Horus in Edfu

Abstract

The Temple of Horus at Edfu, vast in scale and inscriptional density, has traditionally been studied as a corpus of texts isolated from its architectural and iconographic context. Addressing this divide, the article investigates whether digital methods, including machine learning (ML), can speed up the production of epigraphic facsimile drawings while linking Egyptology, historical building research, and the digital humanities, focusing on the Mesenit (room I) and the courtyard–Wabet ensemble (rooms P-Q). The Wabet was surveyed via Structure-from-Motion photogrammetry to build a Heritage Building Information Model (HBIM) within a Common Data Environment. In the Mesenit, a pilot CycleGAN model was trained on an exploratory unpaired dataset of photographic documentation and expert-produced drawings. Technical model behaviour was assessed using cycle-consistency loss and reconstruction-based image metrics, while the scholarly reliability of the generated drawings requires separate comparison with expert-produced facsimiles. Architectural analysis of courtyard P and room Q yielded dimensions and construction evidence, including wall chamfering and partition-wall sequencing relevant to the building’s construction phases. The study concludes that combining HBIM and ML-assisted image generation in a shared digital framework offers a feasible, developing interdisciplinary approach to documenting the temple, requiring further training and refinement before large-scale application.

1. The Temple of Horus as a Challenge

The Temple of Horus at Edfu is a challenge. It is a challenge for Egyptology as well as for historical building research, owing to its sheer size (137 m long, 76 m broad, and up to 35 m tall; it is the second tallest building from ancient Egypt after the Great Pyramid) and its vast and almost completely preserved corpus of inscriptions. Meanwhile, the historical edition Edfou I to VIII [1,2,3,4,5,6,7,8,9,10,11,12,13,14] displays the hieroglyphic text on 3171 large-format pages, plus schematic line drawings of the wall. However, the built space and epigraphy are intertwined, and scientific interests cannot be separated according to the disciplines involved. The inscriptions provide building researchers with information on the function and use of the rooms, whilst insights into the construction process are of direct relevance to Egyptology. At most, the work packages and methodological tools can and must be divided according to expertise. In this respect, research into the Temple of Horus must be regarded as a single entity, approached through a division of labour. Computer-assisted approaches help to meet this challenge, whereby the digital humanities establish a fundamental entity.

In keeping with the philological tradition of Egyptology, and somewhat ‘seduced’ by the overwhelming abundance of inscriptions, the Temple of Horus has hitherto been regarded more as a repository of writing, a “library of stone” [15], with the texts being studied in isolation from their architectural and iconographic context. This is a phenomenon that characterizes the field of Egyptological epigraphy as a whole, for what is regarded as epigraphic work is, in fact, purely philological work with texts that happen to be inscribed on monuments, and that might just as well be recorded on any other medium [16]. Indeed, the spatial arrangement and distribution of motifs, texts and scenes within the building, with their counterparts along axes of symmetry and the like, has long been discussed under the heading of “grammaire du temple” [17,18]. Yet this term, too, carries a metaphorical resonance that evokes philological associations, just as its derivatives and cognate terms—such as théologie pariétale [19] or poésie pariétale [20]—do. The Würzburg Horus Beḥedety Project (HBP), which has been working in Edfu since 2016, has now begun to place the inscriptions back into their context, starting with the iconographic context of the associated ritual scenes (Figure 1), since all of the texts belong to ritual scenes and their design refers to these scenes.

Figure 1. Two ritual scenes from the north wall of the Wabet, 1st register: on the left, the scene ‘Laying hands on the god’ and on the right, ‘Wiping with the nemes cloth’. Photo: © HBP, University of Würzburg.

2. State of the Art

2.1. Egyptology, Architecture, and Digital Methods

Egypt’s architectural heritage remains in excellent condition in many places, not least the Temple of Horus at Edfu. The significantly limited research into the architectural history of Egyptian temple complexes is, therefore, surprising, but the often-excellent condition is also an explanation for this. Building research may seem to be less important because the temples are there and require little or even no imagination to build an idea of their spatiality. Given Egypt’s overwhelming architectural heritage, the work carried out in this field to date lags much behind philological research into the cultural heritage of ancient Egypt. Furthermore, the monumentality—it has been said at the outset—poses additional challenges. Here, methods from the intersection of the digital humanities and historical architectural research—such as 3D modelling and semantic annotation—could enhance the analysis and interpretation of these structures.

The state of research on Edfu, in terms of methodology, is characteristic of that concerning the great temples of the Ptolemaic–Roman period. There is nothing on Philae and Esna from an architectural–historical perspective; for Kom Ombo, work has been started in recent years applying HBIM, but the full model has not yet been published [21]. Only the Temple of Hathor at Dendera is significantly better documented thanks to Zignani’s work [22,23,24]—though none of it is based on digital approaches. For Edfu, publications on the temple’s architecture are quickly enumerated: Fauerbach [25] produced a seminal study with her examination of the pylon. Otherwise, architectural research on Edfu is limited to a study of the column capitals [26] and a few important architectural observations on the mammisi [27], but due to the frequent absence of measurement scales in [27], this work is noticeably lacking in professional architectural–historical rigour. A comprehensive architectural–historical study of the Edfu Temple, such as that presented by Zignani [22,23] for Dendera, is as yet lacking. Egyptian temples also present architectural–historical challenges because the ‘decoration’—which is not merely an aesthetic embellishment but an integral part of the structure—must be taken into account [28,29].

However, the application of digital tools is only slowly starting. The application of digital methods in Egyptology, as a primarily philological discipline, has so far been largely limited to data generation or to the digital continuation of analogue methods to streamline work: For example, a digital drawing tablet [30] is used instead of pen and paper, or a database replaces the card index. Even the 3D visualizations of ancient Egyptian monuments, as tested at Chicago House [31], are ultimately improved forms of documentation that replace illustrated volumes. Furthermore, what is available is usually static and fragmentary. Dynamic processes within space are not considered. Pantalacci [32] (pp. 487–488) recognized this and developed a vision: digital methods allow us to physically experience the temple space and to capture lighting conditions—whose changes bring dynamism to the space—as factors. The significance of light for Egyptian temples is demonstrated by studies such as that by Zignani [24] on Dendera, who used analogue tools to show the effects of light. The limitations of analogue methods are as evident here as they were inevitable at the time of publication, just as they are in [33] (pp. 570–576 with plates 32–35), when the lighting conditions on New Year’s Day are illustrated at half-hourly intervals using eight different images generated from digital 3D models and then explained in the text. The Near Eastern archaeologist Grabowski [34], for her part, investigated the effects of light in Mesopotamian sanctuaries digitally at an early stage as part of her master’s thesis. The models of the architecture carried out previously are, like the lighting simulation, highly stylized though.

For the purposes of Egyptology and building research, a significantly higher level of detail would be required, one that allows the inscriptions to be read and iconographic and stylistic details to be observed, as well as enabling the structural, material, and craftsmanship details to be discerned. The Temple of Edfu allows us to use the existing structure as a reliable reconstruction of at least the final phase of expansion, but it offers the opportunity to transform the monumental challenge of documenting the existing structure into an architectural edition. This is because an architectural–historical edition project for such a large and complex temple can only succeed through the application of digital methods. A prerequisite for this is the integration of the three-dimensional existing structure model into digital databases and georeferenced information systems. The setup of a concise and comprehensive Common Data Environment (CDE) and the method of Historical Building Information Modelling (HBIM) plays a central role here, as it enables comprehensive digital documentation, enrichment through sources, additional text documents and images, as well as the analysis of complex models, whilst also forming the data basis of the digital architectural and Egyptological information. At the same time, it should be noted that both—CDE and HBIM—are still under development in our project, but are in line with current research trends in historical building studies, as a growing number of projects incorporating HBIM shows. Until a few years ago, the application of HBIM to ancient buildings was rather rare, largely due to the fragmentary nature of the surviving architecture [35]. In a best-practice study, this gap was addressed at a methodological level by Viehmann [36] as part of her doctoral thesis. The direct link between architectural research and HBIM methodology is further demonstrated by Trischberger [37] and Vössing [38]. By combining aspects of cultural heritage protection—on which extensive and well-established studies are available [39]—with the degree of preservation, HBIM provides the ideal working environment for capturing the complex realities of the Edfu Temple complex. Furthermore, these methods and technologies from historical building research and the digital humanities offer further possibilities for visualization and interaction with the data. The resulting datasets can also serve as ‘ground truth’ for training AI models. Although work has started here in the temple of Kom Ombo [21], there is no research on ancient Egyptian temples published in detail yet, which makes the approach presented here a rather novel one.

For the present article, we have selected a chapel, the so-called Mesenit (room I in Figure 2), and an ensemble consisting of a half-open chapel and a courtyard, commonly termed Wabet (Figure 2). The state of the art for both will be briefly presented here.

Figure 2. Map of the Temple of Horus at Edfu by U. Fauerbach (courtesy U. Fauerbach cf. [25], p. 37), where the scale is erroneously enlarged by a factor of 1.25. The letters and numbers in the map are the traditional designations to refer to the rooms in the temple.

2.2. The ‘Mesenit’ (Room I)

Whilst room A has been understood in previous research as the main sanctuary, in which not only the barque but also the secret cult image of the temple was kept (e.g., [15], p. 11, [40], in contrast, however, to [41]), the Mesenit usually appears merely as one of the nine chapels in the series of chapels surrounding the barque sanctuary. However, observations of older temple buildings show that the sanctuary—that is, the room housing the cult image—was always situated at the very rear of the temple, to provide the greatest possible protection against desecration [41,42,43]. This is consistent with the parallel structure of the temple and the residential building/palace, in which the sanctuary corresponds to the bedroom. The same spatial arrangement is also found in the Mesenit: not only is it the furthest from the entrance, but, unlike the barque sanctuary, it is no longer illuminated by the light entering through the main portal (cf., [41], p. 8). From a textual perspective, the outstanding significance of the Mesenit is evident in the following aspects: (1) its mention as the first of the inner chapels in the temple’s architectural description; (2) its designation as the ‘first chapel’; (3) the numerous names known for the Mesenit, e.g., ‘Chapel of Beḥedety’ and ‘magnificent sanctuary of the divine winged scarab’, which indicate great theological significance and explicitly designate it as the sanctuary of Horus-Beḥedety; (4) the name, Msn(.t), which is identical to that of the temple itself and is highly significant in terms of local theology: Edfu as the Upper Egyptian Mesen, that is the counterpart to Mesen in the Nile Delta (Lower Egypt), where the mythical victory of Horus over Seth (the arch-enemy of the Egyptian gods and symbol of chaos) is situated. The theme of repelling enemies and the attainment of Horus’s rightful rule is also inherent in the texts of the Mesenit.

Furthermore, the building description indicates that the Mesenit contained a cult statue of Horus-Behedeti as a crouching falcon (gmḥsw) and one of Hathor, the principal female deity. Traces of the shrine made of black stone can still be found on the floor of the room ([44], p. 16, n. 7). The divine designation gmḥsw is generally associated with aggressive power and the repelling of enemies, but also with royal power and the attributes of the sun god [45].

The first register of the reliefs on the east and west walls depicts the cult staffs of Horus: the staff of Horus-Beḥedety and a harpoon known as sgmḥ. Based on parallels, these are to be classified as portable divine images which were carried, for example, during the New Moon and Beḥedet festival ([46], p. 634, [47], pp. 426–435). This important festival in Edfu is also mentioned at the door of the Mesenit ([1], p. 228, 8–9). Based on a description in the courtyard of Edfu, it is likely that Horus and Hathor proceeded to the Mesenit at the start of the festivities.

Another major festival celebrated at every temple in Egypt, in whose cult procedures the Mesenit was integrated, is the New Year festival ([48], pp. 331–337, [33,49]). The most detailed description of the festival can be found on the staircase walls of Edfu and those of the very similar temple of Dendera similar temple at Dendera, within the context of the depictions of processions. Accordingly, before proceeding to the roof chapel, the king and the priest entered the Mesenit, or, in Dendera, the parallel chamber known as Per-Wer ([33], pp. 346–349). The doors leading to the area surrounding the barque sanctuary also mention the entrance to the axial chapel, for example on New Years’ Day. According to the inscriptions on the stairs, the statue of the crouching bird of prey described in the Mesenit was carried onto the roof for the ‘unification with the sun disc (ẖnm-jtn)’ (cf. [50], pp. 38–39): ‘The gmḥsw appears outside his chapel” ([1], p. 554, 4–5), and further: “The statue of Harakhte (= the name of Horus as the sun god) leads the procession, … whereby he is the gmḥsw…’ ([1], p. 576, 3–4).

2.3. The ‘Wabet’ (Courtyard P and Room Q)

The Wabet and the associated courtyard are an ensemble that features in Upper Egyptian temples from the Hellenistic and Roman Imperial periods. Although, strictly speaking, the Wabet is just the chapel opening to the courtyard, it has become an Egyptological façon de parler to refer to the entire ensemble as Wabet. Coppens [49] provides an overview of all the temples in which he has identified a Wabet. There are 13 such ensembles in total, although 4 of them cannot be identified with certainty due to their state of preservation. The findings clearly focus on Upper Egypt—only one of the nine temples identified with certainty is in Lower Egypt. The name Wabet, commonly used in Egyptology, is not the only one, as other terms can also be found in Egyptian sources. Coppens has also written further essays on the subject, of which two in particular are worth mentioning: one draws a comparison between the courtyards of Late Period Theban funerary architecture and the courtyard belonging to Wabet [51], whilst the other examines the play of light through the courtyard in temples [52]. Coppens favours s.t ḥb tpy, ‘place/seat of the first festival’, as the relevant name, because the courtyard at Dendara is also called wsḫ.t n.t s.t ḥb tpy, ‘courtyard of the seat of the first festival’. However, this may also be an indication of the courtyard’s pivotal role between the interior spaces of the temple—the Wabet on the one hand and the staircase to the roof on the other—where the kiosk for the New Year festival is located, which is called s.t ḥb tpy in most cases ([33], p. 359–366). According to Coppens, the origins of this ensemble are to be found in the sun altars that some temples of the New Kingdom possessed. Furthermore, the Wabet itself bears a formal resemblance to the ephemeral chapels of the Sed festival. This festival served to ritually renew the royal reign [53], the king would take his seat on a throne placed on a platform accessible via a few steps, beneath a canopy.

The Edfu Wabet shares with the Mesenit that it was the stage for rites during the New Year festival [48]. Disagreement between Alliot [48] and Fairman [54] centres on when and how often the procession of the divine images from their shrines, via Wabet and the courtyard—where vesting rites and exposure to sunlight took place—was conducted to the kiosk on the temple roof. More recently, it has been suggested that, in addition to festivals, the Wabet may also have served in everyday life [33].

Dendera is a particularly important source of comparison for Edfu, but differs significantly in its architectural execution, as the transition between the courtyard and the staircase, via Room Q (‘Treasury’), is much more spacious than at Edfu: Whilst in Edfu, Room U is very low and one must pass through it hunched over—perhaps not necessarily in antiquity, assuming that the average height of the population was smaller in these times, but it nevertheless remains a cramped chamber given the shrines that the priests had to carry along here—in Dendera, Room Q has almost the character of a hall, which modern people can also enter standing upright. The building description of Edfu describes the relationship between the Wabet ([4], p. 6, 2; [7], p. 16, 1) as the courtyard being hierarchically superior to the Wabet and, therefore, theologically more important than the Wabet itself, second only to the site of the New Year festival on the roof as the central place in the temple during this occasion.

Coppens identifies the Wabet of Edfu as the first classical Wabet. However, given the small number of only nine ensembles that can be reliably identified as Wabet ensembles, the dataset is too limited to determine what constitutes a classical Wabet; it may be due to our own bias that we regard the Wabet of Edfu as classical or archetypal, because we have become accustomed to perceiving the Temple of Horus at Edfu as the archetype of an Egyptian temple, even though it is itself a specific product [15] with its unique own form. Despite Coppens’ assessment that, in Edfu, we see for the first time the characteristic, distinctly pronounced form of a Wabet, he largely excludes it—as well as that of Dendera—because they have already been studied in detail. Due to Coppens’ self-imposed limitations, the Wabet of Edfu remains an under-researched area [55].

2.4. Current Documentation Strategies

The still-missing comprehensive analysis of the extensive corpus of inscriptions and scenes in the Temple of Edfu would require several monographs; a start has been made towards this in the HBP. The basis of a comprehensive epigraphic study are the facsimile drawings ([16], p. 132) of the entire wall decoration produced to a uniform standard that is the result of reflection on various epigraphic methods in use in Egyptology (cf. [16,56]). While the existing edition of the temple of Edfu [1,2,3,4,5,6,7,8,9,10,11,12,13,14] separated ritual scenes from texts (compare Figure 3 and Figure 4), the new facsimile drawings reunify these aspects. Facsimile drawings are required, because a photograph can provide too much distracting, or even misleading visual information, impeding the legibility. In contrast, a drawing can help to grasp a relief’s meaning by deliberately reducing the information to what is important for the eye, and that is the drawing’s foremost function: to serve Egyptological research questions (compare Figure 1 and Figure 5). An example from epigraphy for providing too much information in drawing would be Thiem’s publication on the speos of Gebel es-Silsileh [57]. The drawings show the tiniest damage on the wall, and it is not always easy to distinguish between marking damage or chiselled relief (Figure 6).

Figure 3. Episode from the ritual of founding the temple (2nd hypostyle, southern wall, 1st register): the king and the goddess Seshat stretching the cords. Photo: Stadler (2015).

The HBP’s defined uniform standard covers, amongst other things, line width, shadow lines, and the recording of painting and damage. The ritual scenes and inscriptions are fully documented through photography and drawings. A streamlined version of the Chicago House Method [58] is being applied. Essentially, this involves fewer review cycles. They also seem less necessary due to technological advances, because the high-resolution digital photographs, which can be zoomed in on, are clearly visible on the graphic tablet without any loss of quality, whereas a photographic print viewed through tracing paper—which is the traditional technique used by Chicago House—always appears somewhat opaque. In this respect, a fairly reliable epigraphic transcription is produced on the graphics tablet already in the first step [16,30]. Figure 5 and Figure 7 illustrate the result of the Würzburg Standard. Therefore, this way of producing facsimile drawings is akin to the Karnak Method [59] yet distinct in its use of shadow lines.

Figure 4. Drawing of the same ritual scenes as in Figure 3. Historic drawing, detail from [60], pl. 40d. No hieroglyphs are included; the numbers refer to the hieroglyphs in the corresponding text volume. Furthermore, some details are wrong or missing (e.g., on hands, crowns).

Figure 4. Drawing of the same ritual scenes as in Figure 3. Historic drawing, detail from [60], pl. 40d. No hieroglyphs are included; the numbers refer to the hieroglyphs in the corresponding text volume. Furthermore, some details are wrong or missing (e.g., on hands, crowns).

Figure 5. Facsimile drawing of the same ritual scenes as in Figure 1. Drawing by K. Hepp, © HBP, University of Würzburg.

Figure 5. Facsimile drawing of the same ritual scenes as in Figure 1. Drawing by K. Hepp, © HBP, University of Würzburg.

Figure 6. Example of a facsimile drawing of a scene in the speos at Gebel es-Silsileh by Thiem ([57], Figure 6); clarity is compromised by the numerous damage marks, which are drawn in the same line thickness as the actual relief.

Figure 6. Example of a facsimile drawing of a scene in the speos at Gebel es-Silsileh by Thiem ([57], Figure 6); clarity is compromised by the numerous damage marks, which are drawn in the same line thickness as the actual relief.

Despite the reduction in collation and correction cycles compared to the Chicago House Method, the production of facsimile drawings remains a labour-intensive and, therefore, very costly process. The initial drawing alone takes around one hour’s work for an area of 20 cm × 20 cm = 400 cm². For the Wabet, with a wall area of just under 200 m², this would therefore require almost 5000 working hours. A draftsperson would have to work full-time for 2.8 years to complete this. Due to the complexity of hieroglyphic writing in the Ptolemaic–Roman period, draftspersons must possess, besides sound epigraphic skills, a good knowledge of ancient Egyptian hieroglyphs, ancient Egyptian iconography, and the scholarly literature—in other words, they should be Egyptologists. The production of these so-called facsimile drawings is, therefore, extremely costly, and given the scale of the Edfu corpus, it is an undertaking that could take just as long to complete as the 180 years it took to build the temple (237–57 BCE).

Figure 7. Facsimile drawing of a ritual scene in the interior of the barque sanctuary (room A; west wall, 3rd register, scene 1) recording traces of the ancient polychromy; drawing by S. Dirksen, © HBP, University of Würzburg.

Figure 7. Facsimile drawing of a ritual scene in the interior of the barque sanctuary (room A; west wall, 3rd register, scene 1) recording traces of the ancient polychromy; drawing by S. Dirksen, © HBP, University of Würzburg.

3. Research Goals

Largely neglected is research into the architecture of the Edfu temple. Exceptions are the detailed examination of the pylon [25] and a study of the column capitals as a characteristic element of the building’s ornamentation [26]. Both monographs are pioneering works on the architecture of the Temple of Horus, yet they highlight the limitations of traditional architectural research methods in the face of the building’s monumental scale.

We aim to demonstrate the potential of this collaboration by examining the two selected spatial units in the Temple of Horus Mesenit and Wabet. These case studies were chosen because, on the one hand, the Mesenit has a central ritual function, contrary to the common belief that the room in front of it, room A, was the Holy of Holies. Using the Mesenit as an example, we can also demonstrate how the individual steps of the process (from SfM via orthophoto to facsimile drawing) go hand in hand. Secondly, the courtyard and the Wabet, as an ensemble, offer both an integration into essential cultic processes and—perhaps as a result of this—a variety of architectural forms: as a transition from a closed room to an open courtyard, with steps leading to a semi-open area separated by columns and screen walls, whereby the courtyard also provides access to the staircase leading to the temple roof.

Digital techniques seem to open up new possibilities in this area. Thus, the questions that arise are as follows: Can digital methods and the application of AI facilitate the work of producing drawings, particularly for a temple that is so challenging due to its dimensions, thereby freeing up academic work capacity that can be channelled into quality management (collation and proofreading), speeding up the work and thus saving costs? Do digital methods allow the established and still-necessary epigraphic publication in 2D to be complemented by 3D models in such a way that text and imagery can be experienced spatially, thereby merging Egyptological and architectural–historical research?

4. Materials and Methods

4.1. HBIM

4.1.1. The Egyptian Situation and the Egyptological Needs

The Temple of Horus at Edfu is regarded as the archetype of an Egyptian temple, although no other Egyptian sanctuary is identical to it as far as can be judged from their states of preservation. The Temple of Hathor at Dendera, founded following the completion of the Horus Temple, exhibits a whole series of close parallels with Edfu, e.g., it displays the same layout with the barque sanctuary at the front and a chamber behind it which, in this case, can be identified even more clearly than at Edfu as the sanctuary of Hathor, thanks to its inscriptions and a cult niche. But the temple was never completed and therefore lacks both a pylon and an enclosure wall. It also demonstrates the further development of, for example, the Wabet and its connection to the staircase leading to the roof. The Temple of Horus also deviates from the structure described as the ideal in the ancient Egyptian treatise of the so-called Book of the Temple [61,62]. Various factors, influencing one another, play a role here. As none of the temples are truly contemporary, the defining principle of ‘extending the existing’—as described by E. Hornung for Egypt ([63], p. 88)—comes into play: although being quite similar in its structure, the younger Temple of Hathor at Dendera is considerably larger than the Temple of Horus at Edfu; here, the transition from the Wabet to the staircase leading to the roof is more generously proportioned and presumably more convenient to traverse for ritual purposes than the narrow passageway at Edfu. Alongside architectural development, local traditions shaped by the specific topographical conditions of each site are decisive. These traditions encompass both architectural forms and cult practices. For the Wabet of Edfu, which is regarded as typical, a structural survey is entirely lacking; consequently, considerations regarding its integration into ritual procedures are of a rather theoretical, academic nature. Whilst the ground plan does show an access to the staircase up to the roof, the narrowness of this passageway is striking when viewed in context: a bearer of a cult image would have to stoop and struggle through a narrow passage.

The new techniques promise to provide a significant remedy here, as valid data can be obtained much more quickly than in long and cumbersome campaigns measuring such a complex structure. Lighting conditions under various circumstances and at different times of the day and year can be simulated, and field-of-view analyses carried out, both of which can contribute to reconstructing cultic procedures and gaining an emic understanding of the rooms. Yet, a 3D model that merely shows the building volumes and cubature of the architecture would be insufficient for the needs of Egyptology. A 3D model must also be capable of displaying the reliefs and inscriptions that form an integral part of the built space and are essential for the correct interpretation of a room’s functions. The display must be in a way that allows for collation, not only to review one’s own readings and interpretations, but also to enable other researchers to do the same in the future and to make arguments transparent.

4.1.2. Three-Dimensional Documentation and HBIM

Over the past few decades, digital tools for surveying and researching historical buildings have undergone continuous development. In particular, techniques such as three-dimensional building surveying—namely laser scanning and 3D photogrammetry—are now part of the standard repertoire of historical building research. This is also the case here. The rooms in the present case study in Edfu have been recorded and referenced using Structure-from-Motion (SfM). This is currently the primary technique used in archaeology and historical building research for recording object-related findings, particularly in architecture, excavation finds, etc. ([64], pp. 51–53). One of the great advantages of this technique is the three-dimensional representation of the recorded object. It is, therefore, even more surprising that numerous researchers in their subsequent analysis revert to two-dimensionality. There are a variety of reasons for this, the discussion of which is not the aim of this article. It should be noted, however, that there is a certain lack of good and applicable tools for further processing in three dimensions. Yet it is precisely the conversion of the real building into a model that opens the possibility of gathering and linking knowledge about it. This serves the fundamental principle of historical building research: to use the building itself as the most important source of information. At the same time, the digitized building stock and its modelled representation can be viewed, analysed and researched.

One methodology that would make this possible whilst also serving as a unifying element in interdisciplinary work is the aforementioned HBIM approach. This is largely derived from Building Information Modelling (BIM), which is fundamentally based on the principle that building components in digital 3D models can be enriched with additional information. BIM was developed for the construction of new buildings to collect data from the various trades involved and process it in a networked manner ([65], p. 48). All project participants can view and modify the model. In this way, the tool helps to optimize communication surrounding the construction of a building and to identify potential problems even before the execution phase.

The benefits of BIM for historical building research are also evident, which is why the development of HBIM has been driven forward for some time ([66], pp. 311–312). However, both research into its potential and the development itself are still in their infancy. As research in architectural history and Egyptology continues to rely primarily on two-dimensional drawings for working, communicating and analysing—and this is often done without the surrounding context—HBIM could be significant here on several levels: the survey, its abstract model and any form of supporting documentation are all combined within a single model. The basis is the creation of a 3D model derived from the actual building stock, usually a three-dimensional building survey. As with BIM in new construction, the building components are ideally modelled individually, missing components are reconstructed, and the entire model or individual components are provided with the necessary information ([67], pp. 6694–6696).

The generation of an HBIM comprises a four-step process chain:

• Data collection: carrying out a building survey using digital tools.
• Filtering data: reducing the building survey to relevant information.
• Modelling data: creating each component using simple geometries.
• Processing data: enriching the model with further information ([68], p. 6).

It is particularly the final step—namely, enriching the model with information—that distinguishes HBIM from conventional methods of architectural history research. The information in question is wide-ranging. It encompasses not only basic details such as dimensions, materials and geographical context, but also, for example, building plans, damage surveys, photographs, information on construction methods or technical terms, as well as historical documents ([68], p. 7). In the case of an Egyptian temple, this also includes the wall reliefs as photographs and/or drawing, restoration notes (e.g., preservation of colour or gilding) and their philological translation and comments on the elements and figures depicted.

Therefore, we view the application of HBIMs in the context of ancient Egyptian buildings as a repository of knowledge. The CDE, which the HBIM-model will be a part of, can primarily serve to organize representations of the building stock, reconstructions, and all related information and research data. It facilitates the sharing of information, as this is entered directly into the model, thereby allowing the model itself to function as a heuristic tool. A further advantage, which may even be decisive for the usefulness of the methodology, is that the information in the model can be stored precisely at the component or group of components to which it also belongs on the real object ([69], p. 23). Working with HBIM makes locating and linking information within the model significantly more intuitive. At the same time, it opens up new analytical approaches to the building. This enables analysis at multiple disciplinary and interdisciplinary levels. For example, which components belong to which decorative forms? How are the components treated? What surface areas and what spaces do different forms of decoration occupy? And last but not least: How can all the various and sometimes highly heterogeneous pieces of information be linked together? HBIM is thus far more than just a visualization tool for studying individual components: it enables comprehensive conclusions to be drawn about the building.

Particularly in our context, with its multidisciplinary approaches, such interconnections within a digital model can help to better understand buildings and their contextual significance, identify their individual construction phases, present their research history transparently and, ultimately, document decision-making processes—historically important decisions as well as methodological decisions—using models such as those based in AI. The planning and implementation of HBIM follow defined standards. Depending on the use case and information density, the EU BIM Task Group—a panel of experts operating at European level—has established various levels of detail. These principles, based on BIM standards, have been translated as follows for the handling of built heritage ([68], p. 4):

• Level 0: These are purely CAD drawings based on data collected on site and distributed on paper.
• Level 1: The data in the field—mostly in 2.5D, i.e., with x and y coordinates as well as elevations—is measured and processed in conventional CAD environments.
• Level 2: A comprehensive digital survey is transferred to a 3D modelling programme using 3D scanners or photogrammetry. On this basis, a model of the buildings or individual building components is created, the individual objects of which are in turn converted into an HBIM using standardized exchange formats such as IFC (Industry Foundation Class). IFC serves here as a container for both the model geometry and all other attached information.

Whilst the first two levels of detail have been the standard in building research for several decades, the approach described in Level 2 stems from a desire to create comprehensively interconnected building models for historical analysis as well. For several years now, international experts in heritage conservation, building research and archaeology have been working to model and study historic buildings using BIM software such as Autdesk Revit [66,67,68]. While German researchers were initially rather cautious about this topic, recent developments in the past few years have shown a growing interest. HBIM is represented in various specialist conferences [70] and new publications [36,71] arising from postgraduate theses. Furthermore, an interdisciplinary HBIM working group has formed from the research data infrastructure consortium NFDI4Objects [72], to which some of the authors of this article also belong. A much-discussed problem here is the lack of a standard for HBIM—beyond the general definitions listed above. Thus, it has not yet been established what requirements HBIM software, and the resulting models must meet ([67], p. 6691). Early attempts to capture historic buildings using BIM software failed because the models could not be constructed with sufficient detail ([67], p. 6694).

The usefulness of a model stands or falls on the level of detail that can be represented and the amount of associated information. It is also labour-intensive to model historic buildings with a high degree of detail ([67], p. 6696): due to deformations, different construction phases or simply their research history, they are usually more complex than standardized new BIM buildings. Furthermore, BIM tools and libraries are generally not designed for historic buildings ([67], p. 6705). Consequently, the HBIM working group is making efforts to develop libraries specifically for historic buildings. Above all, a common vocabulary is needed to improve the desired communication ([68], p. 6). One way to better address the needs of HBIM lies in the use and further development of Bonsai—the BIM module of the open-source 3D modelling software Blender v4.2. Compared to proprietary solutions, this software offers not only the advantage that the modelling tools are significantly more flexible, but also that the exchange between the tools and BIM or HBIM is based on the IFC standard. Whilst these considerations are largely of a technical nature, the underlying building itself often presents the greatest hurdle to the application of HBIM. However, this is not the case in the present case study of the Temple of Horus at Edfu—quite the contrary. Due to the serial nature of its architecture, its excellent state of preservation and its multi-layered levels of meaning, it offers a virtually exemplary case study for the application of this promising methodology.

4.2. From Photographs to Line Drawings: Machine Learning for Egyptology

4.2.1. Epigraphic Requirements and Scope of AI-Assisted Line-Drawing Generation

Egyptological research on temples relies on extensive photographic documentation, which serves as the foundation for scaled schematic renderings (see above). These renderings are crucial for verifying and interpreting text passages, particularly in cases where problematic readings arise, and for art-historical and palaeographic analyses. Beyond their epigraphic relevance, scaled schematic renderings form a basis for historical building analysis. However, the creation of these scaled schematic renderings has thus far been a time-consuming and labour-intensive process (see Section 2.4 above). This makes the partial automation of selected drawing steps a potentially useful field of experimentation. The present study investigates whether an image-to-image translation model can support this process by generating preliminary line drawings from photographic documentation. For this exploratory experiment, CycleGAN [73,74] was selected because it can learn mappings between two image domains without requiring strictly paired training data. This characteristic makes the approach suitable for exploring the relationship between the photographic and manually drawn documentation available for Edfu. The requirements of digital epigraphy differ fundamentally from those of many general image-generation tasks. CycleGAN has so far mainly been applied in general image-to-image translation tasks, such as style transfer, image restoration, semantic-to-image synthesis, and controlled image generation, where the target output is usually evaluated primarily in terms of visual plausibility. Digital epigraphy [75,76] requires not only visually convincing results but also metrically accurate, convention-based line drawings in which line weight, omission, colour coding, architectural context, and palaeographic detail carry scholarly meaning. An output may therefore appear visually convincing while nevertheless introducing errors that affect scholarly interpretation.

However, the model’s reliance on unpaired data introduces specific challenges in this epigraphic context: CycleGAN must implicitly learn to distinguish between intentional marks, such as hieroglyphic outlines, and visual noise such as surface cracks or weathering, without explicit paired supervision. More generally, the model must learn the difference between the heterogeneous visual information contained in the photographs and the selective conventions represented in manually produced drawings. Surface irregularities, shadows, cracks, weathering and other photographic features should not automatically be translated into drawing lines; conversely, relevant details must not be omitted, displaced or reshaped during translation. The present experiment should consequently be understood as a feasibility study rather than as an attempt to generate publication-ready epigraphic facsimiles autonomously. The CycleGAN outputs are treated as preliminary AI-generated line drawings that require comparison with expert-produced facsimiles and, ultimately, expert verification and correction. The central question at this stage is whether the model can learn a meaningful translation between the photographic and the drawing domains and which types of scholarly relevant errors occur in the generated output. Accordingly, technical measures of model behaviour, such as cycle-consistency loss and reconstruction-based image metrics, are not interpreted as evidence of epigraphic accuracy in themselves. They provide information about the stability of the learned transformation, whereas the scholarly reliability of the generated drawings can only be assessed to a limited extent at the present image resolution of 512 × 512 pixels. Fine epigraphic and palaeographic details are represented too coarsely for a comprehensive Egyptological evaluation. Larger structural errors, such as omissions, additions, displacements, and shape alterations, can nevertheless be assessed through comparison with expert-produced reference drawings. The following sections therefore first describe the preparation of the image corpus and the CycleGAN experiment before distinguishing between technical model evaluation and epigraphic validation.

4.2.2. Data Source and Data Collection

The data basis for the model training consists of some 30,000 photographs and some 2000 drawings from previous on-site campaigns in Edfu. The photographic and drawn documentation does not merely capture surface decoration but records a wide range of architectural features such as block courses, plinth areas or connecting walls. These features are of particular relevance for architectural analysis and interpretation. The complete corpus was not used for the present experiment. Instead, the CycleGAN pilot was based on a small exploratory subset selected from this substantially larger body of material. The purpose of this subset was not to provide statistically representative coverage of the temple, but to test whether a meaningful image-to-image translation between photographic documentation and manually produced line drawings could be learned under controlled experimental conditions.

Preparing the material for machine learning applications therefore requires additional data curation, including file normalization, metadata enrichment, quality control, and the establishment of a directory structure appropriate for the selected model architecture. The manually produced drawings are characterized primarily by thin black lines on a light background and follow established project-specific conventions for the documentation of reliefs and inscriptions. Some drawings additionally record traces of ancient polychromy through colour coding (Figure 7). These drawings should therefore not be understood as simple visual reductions of the photographs. They constitute scholarly representations in which selection, omission, line differentiation, and colour may carry disciplinary meaning. This distinction is important for machine learning, since the target domain reflects not only visual characteristics but also decisions made according to epigraphic conventions. The preliminary experimental subset cannot represent the full diversity of the available corpus. Differences in preservation, lighting conditions, camera perspective, surface texture, accessibility, decorative density, and drawing conventions all potentially influence model performance. The subset used in the present pilot therefore has to be understood as exploratory rather than representative. Any conclusions drawn from it are consequently restricted to the technical feasibility of the tested workflow and cannot yet be generalized to the complete documentation corpus of the temple. Expert-produced facsimiles provide an essential point of reference for assessing the generated drawings. However, they are themselves scholarly interpretations rather than neutral pixel-level ground truth. Decisions concerning damaged or ambiguous features, the degree of abstraction, and the representation of uncertain lines may differ between experts. For the purpose of domain-specific evaluation, expert drawings therefore serve as domain-specific reference data, against which the preservation, omission, displacement, or alteration of relevant features in AI-generated outputs can be assessed. The CycleGAN dataset was organized into four directories: trainA (100 photographs), testA (11 photographs), trainB (64 manually produced drawings), and testB (7 drawings). No separate validation subset was used. The test directories were used for technical inference after training. For the subsequent domain-specific evaluation, selected line drawings generated from photographs are compared with corresponding expert-produced facsimiles where such reference drawings are available. This domain-specific assessment is distinct from the technical CycleGAN evaluation and is intended to identify omissions, additions, displacements, and alterations in the generated output.

4.2.3. Data Preprocessing Pipeline

The experimental subset described above required technical standardization before it could be processed by the CycleGAN implementation. Photographic material and manually produced drawings were prepared as two separate image domains. Because CycleGAN performs unpaired image-to-image translation, corresponding photograph–drawing pairs were not required during model training. Source material occurred in different file formats and resolutions. Files were therefore converted into raster image formats suitable for the training pipeline where necessary. During preparation of the experimental subset, technically unsuitable files, including duplicated, incorrectly oriented, severely blurred, incompletely cropped, or otherwise unusable images, were excluded. This step was intended to ensure technically processable input rather than to homogenize the scholarly content of the photographs or drawings. For the CycleGAN experiment, all input images were resized to 512 × 512 pixels. The implementation used preprocess = resize, with load_size = 512 and crop_size = 512. Since the selected preprocessing mode performs resizing rather than cropping, no additional random crop was applied. The resulting images therefore had standardized input dimensions, but they were not geometrically or metrically rectified. Perspective distortions, differences in scale, and other properties of the original photographic documentation could remain in the model input.

This distinction is important for the interpretation of the generated output. Resizing the source material to 512 × 512 pixels reduces fine epigraphic and palaeographic detail and does not establish metric correspondence between photographs and generated drawings. The outputs of the present experiment should therefore be understood as technically standardized, preliminary line-drawing proposals rather than metrically accurate or publication-ready facsimiles. Metric rectification, higher-resolution processing, and systematic alignment with orthorectified reference data would require additional processing steps beyond the scope of the present pilot. The main steps of data preparation and their integration into subsequent CycleGAN training, inference, and evaluation workflow are summarized in Figure 8.

Figure 8. Workflow of data preparation, CycleGAN training, inference, and technical and domain-specific evaluation.

5. Results

5.1. Three-Dimensional Documentation and HBIM—A Spatial Analysis of the Wabet

The spatial complex addressed in this article is the so-called Wabet in the Temple of Horus at Edfu, a sequence of two rooms arranged one behind the other (Figure 2, rooms P, Q). This unit appears to occupy an important place within the overall structure of the naos rooms (A–Q)—architecturally, typologically and ritually. The ritual and typological significance of this ensemble of rooms was noted at the outset; likewise, its architectural arrangement within the overall structure of the naos, and indeed that of the entire temple, is striking. Thus, it is precisely this unit that breaks up the strongly axially symmetrical structure of the temple’s composition. It seems obvious that this may be related to the function of these rooms, but also to their immediate location by the temple’s eastern staircase. Nevertheless, the primary focus should be on the purely architectural development of this spatial unit and, in line with the focus chosen here, it will be demonstrated how this novel methodological approach can help to link the various levels of content—architectural, typological and ritual classification—with one another.

During a one-week campaign in October 2025, the Wabet was captured using Structure-from-Motion and spatially referenced. The resulting point cloud serves as the basis for the implementation of the HBIM; the following dimensions are also derived from the dense point cloud. The model was reconstructed in Agisoft Metashape using 1889 photographs (Figure 9). Additional data on the camera in use, metadata, model calculation, asf. can be found in the Supplementary Material.

Figure 9. Cross-section of the SfM model of the Wabet of Edfu; SfM model by C. Brünenberg.

To provide context for the hypotheses set out below regarding construction organization, measurement systems and the building process, it should be recalled that the Wabet is conceived as a two-room complex, beginning with Courtyard P. Spatially, this courtyard serves as a connecting link between the stairwell (U, V), room Q and the vestibule of the barque sanctuary (N), and is equipped with access points to these very rooms (see Figure 2). This connecting function is revealed in the orientation of the room, which initially appears unclear. In contrast to its axially symmetrical counterpart, the ‘House of Min’ (room O), which has a clear westward orientation due to its single opening, the situation is different in courtyard P. Here, the north–south orientation, which is dominant both structurally and typologically, prevails. The courtyard itself is approximately square, with clear dimensions of 4.47 by 4.61 m, the measurement having been taken at the level of the upper edge of the first course of ashlars. The height of the room can no longer be determined from the existing remains, as this room is one of the few in the entire complex that is no longer preserved in its entirety. However, the number of surrounding courses of ashlar and the spatial arrangement in relation to room Q suggest that the height was at least 7.00 m. A plinth course approximately 6 cm high is followed by uniformly worked courses of ashlar of varying heights ranging between 44 and 50 cm. Up to 11 of these courses are still preserved in the current structure. In contrast to the door jamb leading to the stairwell, the door jamb for the passage into the vestibule of the barque sanctuary protrudes slightly from the wall line, thereby accentuating this alignment.

A free-standing staircase bridges the height difference of approx. 0.88 m between courtyard P and room Q. Room Q is roofed and measures 4.41 by 4.28 m with a clear height of 5.28 m. This room, too, is constructed from uniformly laid courses of ashlars 44–50 cm high (Figure 10).

Figure 10. Detail from the SfM model of the Wabet. The masonry structure is hierarchically subordinate to the wall reliefs. However, both levels are important for architectural analysis and for understanding the relationship between design and decoration; SfM model by C. Brünenberg.

A total of 12 courses can be counted, with the lowest course corresponding to the third stone course of courtyard P. This method of construction—that is, the joint masonry of two interconnected rooms—is also evident in the partition wall between the rooms. Thus, in the lower courses of courtyard P, an interlocking of the stone courses also appears to be suggested in the partition wall; however, on closer inspection, this cannot be confirmed. Rather, the partition wall appears to have been inserted (Figure 11, left), yet no structural interlocking can be observed at any point along the partition wall.

Figure 11. In the well-preserved Wabets at Edfu and Dendera, the partition wall between the courtyard and the room above is not built into the structure in either case; photo (left): C. Brünenberg 2025, photo (right): M. Stadler (2025).

This is not to suggest that these represent two distinct phases. The coordination between the construction sequence and the decoration is too great for that. Rather, this may be an indication that, in the case of the Wabet at Edfu, it was deemed important to first establish the basic layout of both rooms and then separate them in a second construction phase. The question of whether this might be related to the function of the Wabet can be answered by looking at the Wabet in the Temple of Dendera, where a virtually identical situation is found (Figure 11, right). In both cases, the typologically related unit comprising a courtyard and a raised adjoining room was first formed without structural constraints, and only later separated with the subsequent installation of the partition. In the case of Edfu, this structural separation appears to extend right up to the uppermost courses of stone, whilst in Dendera, the partition wall is re-anchored into the surrounding wall at the level of the architrave—a structurally sensible solution.

Finally, a detail that is initially inconspicuous but stylistically and structurally significant is the chamfering of the upper edge along the topmost course of stone (Figure 12)—a phenomenon that occurs without exception in all the interior spaces of the Temple of Edfu. It is precisely this, however, that may provide the necessary clues for deciphering the construction sequence (see Section 4.1).

Figure 12. The shadow joint is present in all interior rooms of the Temple of Edfu; here is a close-up photo of the upper edge of the wall in Room J. The chamfering (blue arrows), made after the wall reliefs were carved and painted, is clearly visible; photo: C. Brünenberg.

5.2. Machine Learning for Egyptology

Recording and analysing the spatial structure of the Wabet is, as we see, a manageable undertaking in terms of both time and data. In their course, an SfM model is produced from which orthophotos may be generated that are essential for the second issue that we want to investigate here, redrawing the wall reliefs. This proves to be significantly more complex. To investigate whether this labour-intensive and very costly process in the future, an AI-based approach was tested to automatically generate the drawings using photographs and orthophotos from the SfM model.

5.2.1. First Model Training: CycleGAN Experiment

A first CycleGAN training was conducted on the exploratory unpaired training dataset comprising 100 photographs in Domain A and 64 manually produced drawings in Domain B. The model used a ResNet generator with six residual blocks (resnet_6blocks) and was trained with a batch size of 1 at an image resolution of 512 × 512 pixels. The learning rate was set to 0.0002. Training comprised 100 epochs, with a constant learning rate for the first 50 epochs followed by linear decay over the remaining 50 epochs. The cycle-consistency weights were ƛA = 10 and ƛB = 10, and the identity-loss factor was 0.5. At epoch 100, the recorded cycle-consistency loss values were cycle_A = 0.488 and cycle_B = 0.329. The weighting value Lambda = 10 follows the recommendation from Zhu et al. [74] for consistent reconstruction. These values describe cycle-reconstruction behaviour and should not be interpreted as direct measures of the accuracy o the intermediate generated images. In particular, low cycle-consistency loss does not demonstrate that a generated line drawing correctly reproduces all epigraphically relevant features of the corresponding photograph. Domain B is visually simpler and more homogenous and showed more consistent cycle reconstruction, whereas Domain A contains substantially greater visual complexity and variation (Figure 13).

Figure 13. Example 1—detail of the soubassement in the Mesenit: output from the first CycleGAN training, from top to bottom: real A (photo: © HBP, University of Würzburg), fake B, and reconstructed photograph, rec_A.

5.2.2. Evaluation with Signal-to-Noise Ratio (SNR)

The first results were evaluated with SNR as a quantitative measure of reconstruction quality. Following Gonzalez and Woods [77], SNR quantifies the ratio of meaningful image signal to noise, providing a quantitative measure of similarity between original and reconstructed images. SNR was therefore used as a descriptive measure of cycle-reconstruction behaviour. The original evaluation comprises 25 SNR observations for each domain. The resulting descriptive statistics are summarized in Table 1.

Table 1. SNR summary.

Domain	SNR Observations (n)	Mean SNR (db)	Std SNR (dB)	Max SNR (db)	Min SNR (db)
Domain A	25	20.15	3.82	26.56	11.12
Domain B	25	24.33	2.12	27.69	20.54

The recorded SNR values differ between the two domains. Domain B shows a higher mean SNR of 24.33 dB and a lower standard deviation of 2.12 dB, indicating greater similarity between the original and cycle-reconstructed images and lower variability across the recorded observations. Domain A yields a mean SNR of 20.15 dB and a higher standard deviation of 3.82 dB, which is compatible with the greater visual complexity and variability of the photographic source material.

5.2.3. SNR Trajectory Across 100 Epochs

The SNR trajectory for Domain A was recorded at 25 checkpoints between epochs 4 and 100, at intervals of four epochs (Figure 14 and Figure 15). The values fluctuate substantially during the early stages of training, followed by a marked increase and a later phase in which SNR remains broadly between approximately 20 and 26 dB without a consistent upward trend. The lowest recorded value was 11.12 dB at epoch 8, while the highest value of 26.56 dB was reached at epoch 36. Across the 25 checkpoints, the mean SNR was 20.15 dB. The trajectory suggests stabilization of cycle-reconstruction behaviour rather than continuous improvement. Since SNR measures reconstruction similarity rather than the correctness of the generated line drawings, the curve does not establish whether epigraphic accuracy improved over the same period or whether extending training beyond 100 epochs would improve the scholarly quality of the outputs.

Figure 14. Descriptive comparison of mean, standard deviation, minimum, and maximum SNR values across the 25 recorded observations for Domains A and B.

Figure 15. Line chart of the SNR trajectory for Domain A over 100 epochs, with the training phases indicated.

5.2.4. Egyptological Evaluation of AI-Generated Drawings

When evaluating the results, it must be borne in mind that the AI training results are only available in reduced resolution. At the present resolution of 512 × 512 pixels, fine epigraphic and palaeographic details can only be assessed to a limited extent. Taking this into account, the results of the AI-generated drawings must nevertheless be regarded as mixed so far. Although they demonstrate that CycleGAN is indeed capable of recognizing structures and generating images in both directions (from photograph to drawing, and from drawing to a fake photograph) that bear similarities to Egyptian reliefs, the lines remain too imprecise for use as scholarly facsimile drawings without substantial expert verification and correction. Furthermore, the generated outputs do not consistently distinguish between damage and relief. Due to the low resolution used in the pilot experiment, the automatically generated drawings do not yet meet the requirements of publication-ready epigraphic facsimiles.

When considering the outputs in more detail, it becomes apparent that particularly in examples with clearly visible scene elements or relatively large hieroglyphs, larger-scale proportions and scene boundaries are recognizable in the generated drawings, while some hieroglyphic signs retain sufficiently distinct outlines to permit reading. In the example shown in Figure 16 and 25 of 42 hieroglyphic signs or sign groups (59%) were assessed as sufficiently recognizable for reading.

Figure 16. Example 2—text line from the lower southeastern wall of the Mesenit: output from the first CycleGAN training: (top), real A; (bottom), fake B.

In scenes containing smaller hieroglyphs, this is more difficult at the current resolution (cf. Figure 17); here, it is mainly the outlines of the divine and human figures that can be assessed. They still show inaccuracies or missing lines in some areas, as well as grey scales for shading, which are undesirable in the line drawings we are aiming for. However, the AI-generated drawings would allow the identification of the scene type and may provide a preliminary basis for subsequent manual drawing.

Figure 17. Example 2—Scene in the second register of the southwestern wall of the Mesenit: output from the first CycleGAN training: (left), real A; (right), fake B.

In Figure 18, it is again larger sections of the scene—i.e., greater detail in the source photograph—that yield better results. Here, too, the problem of the grey areas would need to be resolved, as they are currently preventing the outline from being clearly distinguished from the background.

Figure 18. Example 3—detail from the scene in the second register of the eastern wall of the Mesenit: output from the first CycleGAN training: (left), real A; (right), fake B.

However, the present experiment is based on a limited exploratory dataset and should therefore be regarded as an initial feasibility study. Further development will require systematic curation of the training data, including selection, quality control, metadata enrichment and the definition of suitable expert reference material. At the present stage, the AI-assisted workflow does not replace Egyptological expertise. Rather, expert assessment and correction remain essential components of the process. There is still a long way to go before the Egyptological workforce can be set free to undertake Egyptological work proper with the support of AI, doing the basic work of drawing, so that the Egyptologists can merely check for quality and accuracy in small but important details.

6. Discussion

6.1. Potential and Limitations

The long-term goal of this research is to integrate AI-generated epigraphic drawings into a comprehensive digital framework that bridges Egyptology and architectural history. This integration is enabled by the Common Data Environment (CDE) of the HBIM, which serves as the central repository for heterogenous data streams: photographic documentation, point cloud data, three-dimensional architectural geometry, and AI-generated line drawings alike. By consolidating these data types within a single, semantically structured environment, the CDE ensures that each layer of information remains interoperable and mutually referential rather than existing as an isolated dataset. One promising avenue within this framework will be the direct embedding of the AI results into the HBIM, thereby enabling the spatial analysis of texts and scenes within their architectural context. For example, the distribution of ritual scenes in the Wabet could be correlated with the room’s lighting conditions, structural features, or access points, providing new insights into the temple’s design and function. To achieve this, future work will focus on standardizing the output format of the AI-generated drawings to ensure compatibility with HBIM software Bonsai Studio (v0.9.0) (e.g., exporting drawings as scalable vector graphics with embedded metadata).

A further area of significant future potential lies in the application of AI to the processing of three-dimensional point cloud data within the Scan-to-BIM pipeline. While the present study has tested CycleGAN as a potential means of supporting labour-intensive stages of epigraphic documentation, with the aim of reducing manual effort in the production of epigraphic line drawings, the same principle of AI-assisted automation could also be applied upstream in the documentation workflow. The automated segmentation, classification, and semantic enrichment of raw point clouds—for instance, the differentiation of architectural elements such as wall faces, column shafts, and ceiling blocks—represents a critical challenge in the transition from dense survey data to a semantically structured HBIM. A key open question in this context is how the semantic annotations assigned during the AI-processing stage can be preserved and propagated as the data passes through successive stages of the pipeline, from raw scan to geometric mesh to HBIM object. Addressing this challenge will require the development of standardized metadata schemas capable of carrying semantic labels across software environments, as well as validation protocols that allow domain specialists to verify and correct AI-generated attributions without sacrificing the efficiency gains that automation provides. Across all these applications, AI should be understood primarily as a tool for reducing manual effort—a means of delegating laborious, repetitive tasks so that disciplinary expertise can be directed towards interpretive and analytical questions rather than technical production. This reorientation of the research process has direct consequences for the collaboration between the three disciplines involved. By combining CycleGAN (as well as other models), HBIM, and traditional scholarship, this research proposes a new model for interdisciplinary temple analysis, one that leverages digital tools to uncover connections between text, image and architecture.

It is within this integrated digital framework that the following hypotheses are situated. They will be formulated based on the 3D analysis of the courtyard and the rooms of the Wabet. These hypotheses are provisional in nature and cannot yet be conclusively verified based on the current evidence. However, they are grounded in the preceding architectural observations made on site and analysing the three-dimensional model. It can be understood as a heuristic tool for reconstructing the construction and decoration process, thus integrating the AI-based image reconstruction into the CDE of the HBIM-Model. From a purely architectural perspective, the construction process can generally be divided into five ideal, typical steps: (1) the setting out of the building, (2) the laying of the foundations, (3) the transfer of the setting-out marks to the foundation walls, (4) the laying of the ground layout, and (5) the actual stone laying, including the openings. This abstracted sequence serves in the following as an analytical framework within which different variants of the interplay between construction and decoration processes can be discussed.

A key research question in the interdisciplinary collaboration between Egyptology, historical building research and the digital humanities concern the correlation between architectural planning and decorative planning, as well as their execution. In line with the principles of reverse engineering, a nuanced assessment of this can consequently only be made regarding the final stage—the construction and decoration of the walls—whilst the first four stages require further, extensive research. Nevertheless, these steps should be briefly summarized with the utmost caution, given the hypothetical nature of their conclusions. Surprisingly little is known to date about the basic surveying and staking out of a building in ancient Egyptian times ([78], pp. 52–57, 82–84). Nevertheless, it can be assumed that in Ptolemaic times at least the standard repertoire of building surveying was known and used. Thus, the general layout of the temple—the spatial arrangement, wall axes, staircases, etc.—was presumably surveyed using plumb lines or a comparable technique ([78], p. 82). That this was not only put into practical use but was also of importance as part of ritual practice [79,80] is evident from the depiction of the string line being stretched in preparation for the temple survey (cf. Figure 3 and Figure 4). How the next two steps—the laying of the foundations and the transfer of the stakeout to the foundation walls—were carried out remains unclear at this point. However, the plinth course, which runs continuously throughout the entire temple area and is up to 10 cm thick, suggests that this course was used to establish the temple’s ground plan. This was followed by the fine work of both a structural and decorative nature. It was only after the walls were built upon these plinth courses that door and window openings were finally inserted; indeed, the decoration is found exclusively above the plinth course.

6.2. Considerations on the Construction Sequence and Future Research

Based on both the architectural findings and the decorative tradition, together with their preliminary analysis ([28], pp. 17–24), several scenarios can be distinguished within this fifth construction phase, each implying different temporal relationships between construction and decoration. The findings allow for all scenarios, albeit with a certain emphasis on the details and the logic of the building site:

• A first scenario (5a) envisages the largely simultaneous execution of walling, openings and decoration. This would have the advantage that the necessary scaffolding could be used for all stages at the same time. However, this variant can most likely be ruled out. The evidence of the inscriptions on the columns contradicts such an assumption. The inscriptions are to be read from top to bottom, suggesting that the wall surfaces were already fully completed before the final drafting of the texts took place. In addition, there is further evidence in some Egyptian temples, including Edfu, that the walls were decorated from top to bottom, whether through unfinished sections of the decoration or through the arrangement of royal cartouches ([78], pp. 2, 21, 25; cf. the opposite case, pp. 36–37). However, this is not a rule without exceptions. The analyses of the cartouches in the three registers of the inner side of the northern section of the enclosure wall shows that the uppermost register was engraved after the two registers below [81]. Furthermore, it can be assumed that incisions were made, which also presuppose a finished wall surface.
• A second scenario (5b) assumes that the stonework, openings and ceiling installation would have been fully completed before the decoration began. The advantage of this model would lie in the clear separation between structural and artistic trades, which could also have operated largely independently of weather conditions. At the same time, the parallel use of infrastructure such as scaffolding would be conceivable. However, the sometimes extremely fine detailing of the upper wall finishes arguing against this scenario. If the solid ceiling slabs had already been installed, their fitting would have entailed considerable additional effort and carried an increased risk of damage to the wall finishes.
• In contrast, there is a third scenario (5c) in which the stonework, including the openings, was completed up to the top edge of the walls, whilst the room initially remained without a ceiling. This approach would have offered several practical advantages: on the one hand, the existing scaffolding could have been reused here for the subsequent decorative work. Secondly, the incisions and relief carving could have been carried out in sufficient daylight, which would have been particularly important for finer work. For structural reasons, in this model, the installation of the ceiling would only have taken place after the main decoration had been completed. The final painting of the wall and ceiling surfaces would then have happened whilst the scaffolding was still in place. Regarding the Temple of Dendera, Zignani also notes that the ceilings were not installed until after the reliefs had been carved—a conclusion he draws from the plaster that has dripped down onto the reliefs. As there is a layer of paint over this in some places, the plaster dripped down in antiquity, shortly after the reliefs were completed ([23], pp. 187–191).

In assessing scenarios 5b and 5c, the analysis of the upper edge of the block, specifically the chamfer described, is of particular importance. The resulting joint can now be interpreted in two ways, although it is reasonable to assume that the temple’s designers combined both variants to their advantage. Firstly, this creates a shadow joint in each interior space, which can be regarded as a stylistic element and makes the massive ceiling slab appear to float. At the same time, the chamfer serves an essential structural purpose. By removing the right-angled, sharp edge, it prevents the finely worked upper stone coping from chipping uncontrollably when the very heavy ceiling slabs are laid. In this context, the potential of the digital HBIMs extends beyond documentation: digital load calculations and the simulation of assembly sequences within the CDE offer a systematic means of testing the plausibility of the competing scenarios against structural constraints (Figure 19). Crucial for the architectural–historical interpretation, in turn, is the question of the order of work: if the shadow joint was created before the decorative detailing—whose chaîne opératoire follows a fundamentally similar pattern to that of the architectural planning assumed here ([28], pp. 30–34)—this would tend to support scenario 5b, in which the ceiling was already installed. If, on the other hand, the decoration was executed first and the shadow joint only worked on afterwards, this would suggest scenario 5c. On closer inspection of the chamfering, it can be observed that it was carried out only after the decoration had been completed, but before the ceiling panels were installed. This can be established by the fact that clear chipping marks are visible in the relief of the decoration, yet there is no damage whatsoever to the ceiling panel—neither to the stone blocks themselves nor to the paintwork (Figure 12). A detailed examination of these transitional areas therefore appears particularly promising for the further reconstruction of the construction sequence.

Figure 19. Diagram of the preliminary CDE for Wabet’s HBIM.

Overall, the hypotheses outlined here illustrate that construction and decoration processes should not be understood as strictly separate phases, but rather were part of a complex, partly overlapping context of planning and execution. A nuanced analysis of individual architectural elements can provide crucial insights into these processes, without, however, allowing for a definitive reconstruction at this stage. Until now, it has been assumed that both planning and execution were rigorous and outlined on the basis of the construction work. However, the fact that the processes involved in the construction of the temple are more complex is evident, for example, in the plinth zone of the walling of courtyard P. Here, changes in the plinth height can be observed, particularly on the eastern side of the courtyard, which suggest an adjustment to the design. Based on the current state of knowledge, it must be assumed that these are not subsequent corrections made during the construction phase, but rather modifications within the planning phase itself. The term ‘adjustment’ is, therefore, only of limited suitability in this context, as it implies a reactive correction, whereas the evidence on the building suggests a dynamic planning process that had not yet been fully finalized. This, too, is entirely in keeping with the standards of the time. The continuation of the decoration in the stairwell’s entrance hall could also be interpreted in this light and points to a close interplay between the architectural layout and the overall decorative concept.

Taken together, these observations underscore the value of the integrated approach pursued here: it is only by holding architectural evidence, epigraphic documentation, and digital modelling in a shared analytical space, as the CDE of the HBIM enables, that the full complexity of the temple’s construction and decoration history can begin to be reconstructed.

7. Conclusions

This article uses a room (Mesenit) and an ensemble of two rooms (the Wabet) in the Temple of Edfu to highlight new ways of documenting, reanalysing, and reinterpreting the entire building using digital technologies. This can be achieved by combining the methodological approaches of Egyptology, historical architectural research, and the digital humanities.

Two approaches were initially presented—HBIM and AI-based image generation—each pursuing different objectives. AI-assisted image generation, provided that sufficient reliability can be achieved, may reduce parts of the manual workload in the future. It could support the more rapid production of preliminary digital presentations of the rich decorative sequences of the Temple of Edfu. By converting the three-dimensional building survey and all associated information into a structured HBIM, all participating disciplines would in turn be provided with a powerful tool to better understand the temple as both a complex architectural structure and a ritual site, while simultaneously storing research results in a structured and interconnected manner. Thus, the connection between both approaches can be seen in the underlying CDE, in which the SfM model, AI-based image generation, and the reconstruction model serve as information carriers. At the same time, the focus areas of Egyptology and historical building research—namely, the analysis and interpretation of construction-related processes as well as the work of the sculptors who carved the reliefs and inscriptions—are integrated at this level and combine to produce shared, comparable, and traceable research results.

Therefore, the creation of a long-term CDE is necessary for such complex research tasks as the Temple of Edfu in our view. The conversion of the captured point clouds into HBIM-compatible models beyond a single spatial sequence is already feasible today and can be implemented at any time, given the necessary time and funding. AI-based research would also be conceivable for this approach, such as the use of AI in Scan2BIM processes during point cloud processing. A central challenge in this context remains the preservation of semantic enrichment across successive stages of the pipeline: as point cloud data passes from raw scan to geometric mesh to HBIM object, the semantic annotations generated during AI-assisted processing must be retained and propagated in a structured and verifiable manner. Developing standardized metadata schemas and discipline-specific validation protocols will be essential to address this challenge without forfeiting the efficiency gains that automation provides.

The AI component described here remains at an early stage. The results demonstrate that CycleGAN can learn a basic transformation between the photographic and drawing domains, but they do not yet establish the scholarly accuracy of the generated drawings or a reduction in the overall time required for documentation. Further development will require higher-resolution processing, systematic dataset curation, independent validation, and continued expert assessment and correction. However, the data—and this is the real crux of the matter—is available in large quantities. Yet it must be acknowledged that Egyptology constitutes a niche field in terms of AI applications and large-scale training data. Unlike domains such as natural language processing or general image recognition, no dedicated Egyptological foundation model currently exists that has been trained specifically on the visual and epigraphic conventions of ancient Egyptian art and script. The disciplinary specificity of the material, from the strict canon of figural representation to the formal conventions of hieroglyphic inscription, means that general-purpose models cannot be straightforwardly transferred to this domain without substantial fine-tuning. Although both Egyptology and architectural research are rather specialized fields in terms of AI applications and big data, they need not remain passive beneficiaries of development in other domains. On the contrary, they can and should serve as a model for basic research into high-quality AI applications. Dealing with the complex material from Egypt in which writing and image are deeply interconnected in a monumental scale and context pose challenges whose solutions may serve for other fields as models. The development of domain-specific, high-fidelity AI models for the cultural heritage sector aligns directly with broader policy objectives at the European level, where frameworks such as the EU’s approach to trustworthy AI and its commitments to cultural diversity explicitly call for investment in applications that serve minority languages, underrepresented cultural traditions, and specialized scholarly domains [82].