The Tyranny of Wholeness

The Tyranny of Wholeness: Restoration Defaults and the Epistemological Status of Fragmentation in Generative AI Cultural Heritage Practices

Abstract

The intervention of generative AI in cultural heritage image processing has moved from the technical frontier to routine production. Millisecond-level mural restoration, mass generation of artifact patterns, and automatic completion of missing areas—capabilities that were experimental subjects of academic papers three years ago—have today become standard tools in museum digital workflows. This paper argues that precisely because AI restoration has matured, a problem previously obscured by the excitement of technical progress deserves to be seriously raised: the “completeness bias” shared by these tools—the notion that complete images are more valuable than fragmentary ones and that restoration is the only legitimate response to fragmentation. Is this a technically neutral engineering goal or the sedimentation of specific cultural values within algorithmic architectures? Using the fragmentary state of Han dynasty pictorial stones (stone reliefs) as the subject of analysis, this paper develops its argument from three dimensions: the systematic suppression of the epistemological status of fragmentation by the restoration paradigm in archaeological epistemology, the implicit encoding of “wholeness” in the denoising architecture of diffusion models, and the theoretical possibility of fragmentation as a repressed cognitive resource within a recontextualization framework. This paper does not argue that AI restoration is wrong, but rather that today, as restoration technology has matured, “non-restoration” needs to be established as an equally legitimate epistemological stance.

Keywords: Completeness Bias; Algorithmic Ethics; Cultural Heritage; Generative AI; Han Pictorial Stones; Fragmentation; Epistemology

1 Introduction: Problems After Restoration

1.1 A Question Already Answered and a Question Not Yet Asked

In late 2024, the mural restoration AI system released by Jiao Licheng’s team at Xidian University could complete automatic filling of damaged murals within five seconds, with “traces of completion being difficult to detect by the naked eye.” In the same year, artifact image restoration tools based on diffusion models evolved from laboratory prototypes to production systems actually deployed in multiple museums. The National Museum of China used AI datasets for the detection and identification of oracle bone inscriptions, and a team from Huazhong University of Science and Technology used diffusion models to simulate the evolution of Chinese characters to assist in the decipherment of oracle bones. On the design side, batik LoRAs, paper-cutting LoRAs, and New Year painting LoRAs on e-commerce platforms output standardized images that are visually complete, neat, and meet the “traditional” expectation within seconds.

The question “Can AI restore cultural heritage images?” has been answered affirmatively at the technical level. This paper has no intention of objecting to this answer. AI restoration has its irreplaceable legitimacy in the production of archaeological knowledge: approaching historical original appearance is the fundamental pursuit of archaeology, and restoration is a necessary path to that pursuit; AI has increased the efficiency and precision of this path by leaps and bounds.

But this paper attempts to raise a different question—a question that has surfaced precisely because of the maturity of restoration technology: When restoration is no longer a technical challenge but a resolved engineering task, does the legitimacy of “restoration” itself as a default choice need to be re-examined?

In other words: under the condition that AI makes “perfect restoration” nearly zero-cost, can “non-restoration” be established as an equally legitimate stance with an independent epistemological status?

This question was difficult to raise seriously before the intervention of AI. When restoring a fragmentary mural required months or even years of professional labor, “preserving fragmentation” was in many cases merely a passive result of insufficient resources, not an active choice requiring theoretical justification. Generative AI has changed this condition. When the marginal cost of restoration drops to near zero, choosing not to restore is no longer a last resort but a decision requiring reason—and providing reasons for this decision requires first questioning the premise that has always been defaulted as correct: whether the complete image is truly superior to the fragmentary image in all senses.

This paper names this premise “completeness bias” and analyzes its operation from three levels: the systematic suppression of the epistemological status of fragmentation by the restoration paradigm in archaeological epistemology (Section 2), the implicit encoding of “wholeness” in the denoising architecture of diffusion models (Section 3), and the theoretical possibility of fragmentation as a repressed cognitive resource within a recontextualization framework (Section 4).

1.2 The Triple Inhabitation of Completeness Bias

Before entering the analysis, it is necessary to clarify the conceptual boundaries of “completeness bias.” What this paper refers to is not the personal preference of a specific discipline or a specific practitioner, but a cross-disciplinary, institutionalized default orientation embedded in technical architectures. It operates simultaneously on three mutually reinforcing levels:

Epistemological Level. In the dominant paradigms of archaeology and cultural heritage research, fragmentary images are defaulted as “information deficit”—a negative event deviating from the “should-be state.” The research approach of digital restoration presupposes a clear value hierarchy: complete > fragmentary, post-restoration > pre-restoration. Fragmentation has no independent epistemological status of its own; it is merely an intermediate state to be overcome.

Methodological Level. In cultural and creative design practice, completeness bias operates in the form of material selection logic. Designers extract clearly identifiable pattern units, complete shapes, and parameterizable compositional structures from cultural heritage images. Damaged parts are filtered at the earliest stage of the workflow—broken edges do not enter design drafts, weathered areas are not selected as material, and incorrectly joined stones are not included in image libraries. This is not a choice made after explicit discussion but a filter that is not even noticed.

Technical Architecture Level. In the algorithmic architecture of generative AI, completeness bias is encoded as the implicit objective function of the model. This point will be argued in Section 3—it constitutes the core analysis of this paper at the level of algorithmic ethics.

There is a mutually reinforcing cycle among the three levels: the restoration paradigm of archaeology provides an epistemological endorsement of “value only in completeness” for cultural and creative design; the material selection logic of cultural and creative design provides a dataset dominated by complete images for AI training; and the default “completion” behavior of the AI model in turn confirms the presupposition that “completeness is superior to fragmentation”—each link solidifies the premises of the others, while the premise itself has never been questioned directly.

2 The Presupposition of Completeness in Archaeological Epistemology

2.1 Legitimacy and Limitations of the Restoration Paradigm

The complete legitimacy of the restoration paradigm within its own epistemological framework must first be acknowledged. The core question pursued by archaeology is “what the image originally looked like,” and digital restoration is a rational path to that pursuit. The establishment of the thematic classification framework for Han pictorial stones by Jiang Yingju and Yang Aiguo [1], the systematic interpretation of image symbolic meaning and compositional patterns by Xin Lixiang [2], Wu Hung’s revelation of the functional logic of the Wu Liang Shrine images in funeral rituals [3], and Zheng Yan’s analysis of the narrative structure of tomb space [4]—the common foundation of these studies is the reductive inference of the complete state of the image, with fragmentation seen as an obstacle to historical truth. Within the problem consciousness of archaeology, this orientation is beyond reproach.

The problem arises in the stage of paradigm transplantation. When the restoration paradigm is transplanted from archaeology to design without reflection, a judgment that serves as a methodological premise in archaeology (“we need to approach the original appearance to understand history”) is transformed into a value judgment that is not self-evident in design (“complete images have more design value than fragmentary ones”). Archaeology asks “what was it originally,” while design asks “what operational space can be generated in the current context”—the latter does not require completeness as a premise.

2.2 Fragmentation: Information Deficit or a Different Kind of Information?

Han pictorial stones provide an ideal case for analyzing this problem. These Han dynasty image remains, spanning Jiangsu, Shandong, Henan, and Shaanxi, were produced mainly from the mid-Western Han to the late Eastern Han, spanning nearly three hundred years. A basic physical fact has always been in a background position in existing studies: the fragmentary state of pictorial stones is not an accidental exception but the physical norm after two thousand years of weathering and fracturing; cases of complete preservation are the minority.

The restoration paradigm interprets this physical norm as “information deficit”—originally existing visual information lost due to physical processes. This interpretation is correct within the archaeological framework. But it simultaneously obscures a symmetrical fact: while physical fracturing eliminates certain information, it also produces information that did not originally exist.

The direction of a fracture line records the interaction between the internal crystal structure of the stone and the direction of external force—this is knowledge about matter, which the complete pictorial stone does not provide. Weathering gradients record the differential degradation caused by temperature, humidity, and microbes in different mineral component areas of the stone surface over two thousand years—this is knowledge about time, which restoration precisely eliminates. Misalignment of joins records the spatial judgment of a certain group of people in a certain historical period attempting to reassemble fractured stones—this is knowledge about the history of reception, which completeness restoration corrects as an “error.”

The concept of “age-value” (Alterswert) proposed by Alois Riegl in The Modern Cult of Monuments provides a classic theoretical fulcrum for this analysis [5]. Riegl distinguished “historical value” (interpretation depending on professional knowledge) from “age-value” (temporal experience depending on intuitive perception), pointing out that the latter precisely inhabits the traces of wear on the material surface. A stone surface covered with weathering etches, even if the viewer completely understands nothing of Han funeral culture, allows for the perception of “antiquity” and the “passage of time” from its material state. Restoration eliminates not only visual “deficits” but also this low-threshold temporal information accessible through material intuition.

Ingold’s discussion of the “life history of things” pushes this perspective from aesthetics to anthropology [6]: traces on material surfaces are not static “properties” but processual records of the continuous interaction between the thing and its environment. Restoration interrupts this record, replacing it with a conjecture about “original appearance”—no matter how precise the conjecture, it belongs to a completely different epistemological category from the real traces of two thousand years of material interaction.

This does not mean that restoration is wrong in all cases. It means: Fragmentation is not a zero-information state; restoration is not a zero-cost operation. Every restoration necessarily eliminates one type of information while restoring another; this trade-off should be explicitly recognized and discussed, rather than being obscured by the default orientation that “completeness is naturally better than fragmentation.”

3 Completeness Encoding in Diffusion Models

3.1 The Denoising Process: An Implicit Value Judgment

If completeness bias in archaeology operates in the form of a disciplinary paradigm, then in the algorithmic architecture of generative AI, it is encoded in a more hidden and difficult-to-question form: the model’s training objective function.

The core working principle of the Diffusion Model can be summarized as follows: in the training phase, noise is gradually added to the image until it becomes pure noise; in the generation phase, the reverse process is learned—gradually “denoising” from pure noise until a “clean” image is obtained [7]. This architectural design contains a deep metaphor: noise is what needs to be removed, and clarity/completeness is what needs to be restored. The model’s loss function explicitly sets the optimization goal to “minimizing the gap between predicted noise and actual noise”—in other words, the model is trained as a machine to “remove everything that deviates from the training data distribution.”

When this architecture is applied to cultural heritage image processing, a critical question arises: Does fragmentation belong to “signal” or “noise” in the model’s internal representation? Fracture lines, weathering etches, and missing areas—these features are low-frequency or even zero-frequency events in the distribution of training data (since the training set is dominated by complete images), and the model’s denoising process will naturally treat them as deviations to be “removed” or “completed.” Even if no one explicitly requests the model to restore a fragmentary pictorial stone, the model’s default behavior tends toward smoothing fracture lines, sharpening blurred areas, and filling missing areas—because that is the direction its training goal points toward.

This is not a bug, but a feature—but it is a feature carrying a specific value judgment. The model has no ability to distinguish “damage that needs to be repaired” from “historical traces that should not be eliminated,” because in the mathematical form of its loss function, these two are no different—they are both things that “deviate from the training data distribution.”

3.2 The Politics of Silence in Training Data

The second embedding point of completeness bias in AI systems is the composition of training data.

Large-scale pre-trained image generation models (such as the LAION-5B dataset for Stable Diffusion) take images crawlable on the internet as their data source. The vast majority of these images are complete, clear, and meet contemporary visual norms—fragmentary cultural heritage images account for an extremely low proportion of the dataset. This means that when the model learns “what an image should look like,” the distribution it learns is completeness-centric. Fragmentation is not something the model actively refuses to learn, but something it has never seen sufficiently—it is not in the model’s “vocabulary.”

Foka and Griffin, when discussing the relationship between AI and cultural heritage, pointed out the core of this problem: the material and historical information carried by cultural heritage data faces the risk of being homogenized in the AI generation pipeline [8]. The word “homogenization” precisely describes this process: the uniqueness of cultural heritage images—including their fragmentation—is compressed into a statistical position in the training data distribution after entering the AI pipeline, and the center of this distribution is “complete, clear, and without deficit.”

The concept of “poor image” proposed by Steyerl in her critical writing on image circulation gains new applicability here [9]: fragmentary cultural heritage images occupy the position of “poor images” in the data politics of AI—not because their resolution is low, but because their visual features deviate from the center of the data distribution, and thus their “right to speak” in the model’s internal representation is extremely low. When the model is required to process a fragmentary pictorial stone image, its response is not to “understands” the fragmentation, but to “overwrite” it with the prior of the data distribution center (complete images).

3.3 The Rhetoric of Negative Prompts

A mechanism widely used in practice but rarely reflected upon further exposes the operation of completeness bias: negative prompts.

When using diffusion models to generate images, designers usually write “broken, damaged, cracked, weathered, incomplete” in negative prompts to exclude “defects” from the generated results. The default classification of these words—placing them in the “don’t want” position—is itself a rhetorical act: it defines fragmentation as an attribute that should be actively rejected. Interestingly, even if designers do not explicitly reject these attributes in negative prompts, the model’s training data distribution will make the generated results naturally incline toward completeness—the negative prompt merely transforms an already existing bias from implicit to explicit.

And when someone attempts to reverse this logic—writing “broken, fragmented, weathered” in positive prompts to actively generate fragmentation effects—what they obtain is not “real fragmentation,” but “fragmentation as imagined by the model”: a statistical product that fits the distribution of the “fragmentation” label in the training data. Its visual features are uniform, predictable, and lack the non-uniformity and randomness of real physical processes—in other words, it is a domesticated representation of “fragmentation” within the frame of “completeness aesthetics.” The model can generate images that “look old,” but it cannot generate images that “truly experienced two thousand years,” because the information source of the latter is the interaction between matter and time, not between data and algorithms.

3.4 From Technical Problem to Ethical Problem

What the above analysis reveals is not a technical defect of AI systems—in an engineering sense, the denoising architecture of diffusion models is efficient and self-consistent. It reveals an ethical problem: when an algorithmic architecture carrying specific cultural value judgments is deployed at scale for cultural heritage processing, it accelerates the production of one type of knowledge at a systemic level (conjecture-based knowledge about “original appearance”) while accelerating the elimination of another type of knowledge at a systemic level (materiality-based knowledge about “temporal process”)—and this trade-off is never explicitly placed on the decision-making table.

It needs to be emphasized that this paper’s position is not “AI should not restore cultural heritage.” The value of AI restoration in archaeological knowledge production has been fully demonstrated; that is not the issue to be debated. The position of this paper is: Today, as AI restoration has become the default option, “non-restoration” needs to have its legitimacy re-established as a choice—not because restoration is wrong, but because under the condition that restoration has become so easy, “non-restoration” has instead become a stance requiring reason, and that reason has not been systematically provided to date.

The following two sections attempt to provide that reason.

4 Epistemological Status of Fragmentation: Fourfold Resource Attribute

If fragmentation is not just a loss of information but the acquisition of another kind, what does it carry? Using fragments of Han pictorial stones as subjects of analysis, this section identifies a fourfold resource attribute of fragmentation, each pointing to cognitive possibilities that complete images do not possess.

4.1 Narrative Openness

A complete image of a “chariot and horse procession” has its own narrative closure. The carriage, attendants, road, and direction of travel together constitute an internally self-consistent scene, within which the viewer’s cognition is guided, digested, and also limited. When only a fragment of a chariot remains due to physical fracture, the fracture line cuts off the continuation of the image in its original composition; the remaining visual information no longer points toward a fixed narrative destination.

Umberto Eco, in The Open Work, distinguished between “closed texts” and “open texts” [10]: the former guide the reader toward a preset interpretive path, while the latter retain space at the structural level for the coexistence of multiple interpretations. Fragments present a non-intentional open structure—it should be noted that this is different in generative mechanism from the openness intentionally constructed by creators discussed by Eco, but is isomorphic in formal structure. The openness of a fragment is not designed; thus, it does not carry any presets about “how it should be interpreted”—this is precisely its unique resource attribute.

Recent research by Colitti, Formia, and Gasparotto has further distinguished “absence” in digital cultural heritage, pointing out that absence can under certain conditions become an entry point guiding the viewer to participate in meaning construction [11]. In the operational design of recontextualization, narrative openness means that fragments have higher contextual adaptation flexibility than complete images: when a narrative-closed complete image enters a contemporary scene, it is easily categorized as “decoration” or “cultural reference,” and cognitive processing is immediately completed; when a narrative-unclosed fragment enters a contemporary scene, the relational space formed between the void at the fracture and the semantic context of the scene itself may trigger more persistent interpretive activity.

4.2 Non-forgeability of Age-Value

The second resource attribute of fragmentation points to the temporal dimension. As already discussed in Section 2, Riegl’s concept of “age-value” [5] reveals that wear traces on material surfaces constitute an independent category of value. This attribute gains new urgency in the context of generative AI: digital tools can create a “sense of history” through filters in seconds, but there are discernible differences between artificial aging effects and real physical wear—real weathering gradients have non-uniform, non-periodic random characteristics, and fracture lines follow the physical laws of internal crystal structure in stone rather than design templates.

Benjamin’s concept of “aura” forms a cross-generational echo here [12]: mechanical reproduction (or AI generation) can infinitely approach the visual appearance of the original, but cannot reproduce the material history that makes the original “this one” rather than “any one.” Section 3 argued that even if AI is used to actively generate “fragmentation effects,” what is obtained is a “statistical approximation of fragmentation” sampled from the training data distribution—it may visually look “old,” but it does not carry any real material-temporal information. The non-forgeability of age-value thus constitutes the material foundation of fragmentation as a resource: it is something that digital counterfeits cannot replace.

4.3 Cognitive Activation: Incongruity and Defamiliarization

When a fragment bearing two thousand years of physical traces appears in a visual environment belonging to contemporary daily life, what happens at the cognitive level of the viewer?

The “incongruity-resolution model” proposed by Carroll [13] provides a first layer of explanation: the gap between cognitive expectation and actual perception triggers the active search of the cognitive system—the viewer attempts to establish a connection between two incompatible tenses, a search process that itself constitutes the production of meaning.

Shklovsky’s theory of “defamiliarization” (ostranenie) provides a complementary second layer of explanation [14]. Unlike the incongruity model emphasizing “resolution of the gap,” defamiliarization emphasizes “delay of the perceptual process”—making the familiar strange to recover people’s perceptivity toward things. The material quality of the fragment (rough stone surface, irregular fracture edges, weathering gradients) contrasts with the smooth surfaces of the contemporary environment, slowing down the speed of visual scanning and forcing the viewer to shift from “browsing” to “gazing.”

Both mechanisms operate simultaneously in the recontextualization of fragments: the incongruity model explains why the viewer attempts to find a connection between two tenses (resolution), while defamiliarization theory explains why the viewer stays longer on the physical quality of the fragment (delay). When a complete image is implanted into a contemporary scene, due to its narrative closure and visual completeness, it is more easily categorized and “digested” by the cognitive system—the incongruity effect is absorbed faster, and the dwell time is shorter.

4.4 Material Testimony

The fourth resource attribute of fragmentation—and the one most easily obscured by aesthetic discussion—is its testimonial function.

Fracture lines, weathering gradients, transport wear, and join misalignments—these traces are not just aesthetic objects of being “beautiful” or “not beautiful”; they are material records of physical processes over two thousand years. Weathering gradients record the interaction between geological properties of stone and environmental conditions; the direction of fracture lines records a specific mechanical event; join misalignments record the spatial judgments of a certain group of people at a certain historical period toward fragments.

This information is not knowledge about “what the image originally portrayed” (the question pursued by the restoration paradigm), but knowledge about “what the image has experienced over two thousand years”—a processual, material, irreversible type of knowledge. Restoration operations eliminate the latter while restoring the former. In most cases, this trade-off is recognized; but under the influence of completeness bias, it is rarely treated as a trade-off requiring discussion—because the former is defaulted to be more valuable.

Yuk Hui’s discussion of “cosmotechnics” [15] provides a broader theoretical framework here. Hui points out that the understanding of “technology” in different cultural traditions is embedded in different cosmological premises. “Empty space” (liubai) and “fragmentary mountains and waters” in traditional Chinese craft are not the result of technical deficiency but are manifestations of a set of cosmologies concerning “imperfection as perfection” in material practice. When generative AI covers cultural heritage images with its specific concept of “completeness,” what is eliminated is not only physical traces but potentially a different set of cultural wisdoms for dealing with “incompleteness.”

5 Recontextualization: An Alternative Framework

5.1 Appropriation and the Second Context Transition

If fragmentation carries cognitive resources that complete images do not, then an operational question arises: how are these resources activated?

Appropriation theory provides the framework. Taking an existing object from its original context and placing it in a new one to obtain a re-definition in a new system of meaning—from Duchamp’s “readymades” to the image reproduction of Pop Art, the core proposition established by appropriation theory is: meaning is not an inherent property of the object but is bestowed by the context [16].

The specificity of Han pictorial stone fragments lies in having experienced two context transitions. The first is the non-intentional physical fracture—the fragment is detached from the funeral ritual context of the tomb chamber. But the fragment did not thereby enter a state of complete absence of context: through archaeological publication, museum display, and academic research, it obtained an institutional context of archaeology and art history. The second context transition is the intentional design operation—extracting the fragment from the “archaeological/museum context” and implanting it into the “contemporary daily context” (subway stations, cafes, bookstores). The distinction between the two transitions is critical: the design operation is not a restoration of the “loss” caused by the first transition, but an opening of a third path of existence while acknowledging all prior context levels of the fragment.

In terms of Rogers’s typology of appropriation [17], this belongs to transculturation: not for the purpose of consumption or derogation, but introducing material into a new system of expression while fully acknowledging the original context.

5.2 From Aesthetic Stance to Methodological Tool

The aesthetic stance of “fragmentation as resource” has already seen deep accumulation in multiple traditions. The Japanese aesthetic of wabi-sabi treats imperfection, impermanence, and incompleteness as core dimensions of beauty [18]. Kintsugi (gold-joining) highlights cracks with gold lines rather than hiding them. Kentridge’s charcoal animations record the layering of time with incompletely erased strokes [19]. Cai Guo-Qiang’s gunpowder paintings take the traces left by uncontrollable physical processes as their core language.

But in the field of design practice for cultural heritage, “fragmentation as resource” has remained at the level of aesthetic declaration without being transformed into an actionable methodology. In the current creative practices surrounding Han pictorial stones—the series of products from the Xuzhou Museum of Han Pictorial Stones, various cartoonized redesigns, cultural gene extraction models based on Analytic Hierarchy Process (AHP) [20][21]—complete images are used as material without exception. Fragmentation is filtered at the earliest stage of the design workflow and has never entered methodological discussion as an independent dimension of analysis.

This paper argues that transforming “fragmentation as resource” from an aesthetic stance into a methodological tool requires at least three steps: typology (decomposing the vague “broken” into categories with different visual affordances—such as missing fractures, weathering blur, join misalignments, and edge damage, each carrying different narrative openness, temporal texture, and cognitive activation potential); parameterization (establishing a quantifiable annotation system so that the state of fragmentation can be recorded, compared, and called upon); and matching (establishing a structured matching logic between fragments and contemporary contexts, transforming recontextualization operations from reliance on intuition to a traceable decision process). The detailed development of these three steps is beyond the scope of this paper as a theoretical essay, but they indicate the direction of the transformation from stance to method.

6 Conclusion: The Timing of the Question

The core thesis of this paper can be summarized as: Completeness is not a natural attribute of cultural heritage images, but a cultural choice about “what is worth preserving.” The emergence of generative AI did not create this choice, but it reduced the cost of operating this choice to zero—and it is precisely this zero cost that makes questioning it possible, and makes questioning it necessary.

The ability of AI to restore cultural heritage images is real, valuable, and irreplaceable in archaeological knowledge production. This paper has no intention of denying that. But it attempts to point out: today, as restoration technology has matured, we are standing at a point where it is worth pausing. The core question of the previous stage was “Can AI restore?”—that question has been answered. The core question of this stage should be “Should AI restore in all cases?”—and answering that question requires first acknowledging the independent epistemological status of fragmentation, acknowledging that restoration is not a zero-cost operation, and acknowledging that “non-restoration” can be a deliberate choice supported by argument rather than a passive result of insufficient resources.

The Han pictorial stone fragment is just a case, but the problem it reveals has broader applicability. Rust on bronzes, peeling of murals, moth-eaten textiles, weathering of buildings—the physical reality of cultural heritage is a process of continuous wear. When AI can “restore” any fragment to any imagined “completeness,” what is erased is not only physical traces but also the temporal depth, narrative openness, cognitive activation potential, and—perhaps most importantly—a cultural wisdom that is different from “completeness as perfection” carried by fragmentation.

Preserving these things is not a technical regression but an epistemological progression.

References

[1] Jiang Yingju, Yang Aiguo. Han Pictorial Stones and Han Pictorial Bricks [M]. Beijing: Cultural Relics Publishing House, 2001.
[2] Xin Lixiang. Comprehensive Study of Han Pictorial Stones [M]. Beijing: Cultural Relics Publishing House, 2000.
[3] Wu Hung. The Wu Liang Shrine: Ideology of Ancient Chinese Pictorial Art [M]. Liu Yang, Canhe (Cenhe), Translators. Beijing: Sanlian Bookstore, 2006.
[4] Zheng Yan. Mask of the Deceased: Research on Tomb Art from Han to Tang [M]. Beijing: Peking University Press, 2013.
[5] RIEGL A. The Modern Cult of Monuments: Its Character and Its Origin[J]. Oppositions, 1982(25): 21-51.
[6] INGOLD T. Making: Anthropology, Archaeology, Art and Architecture[M]. London: Routledge, 2013.
[7] ROMBACH R, BLATTMANN A, LORENZ D, et al. High-resolution image synthesis with latent diffusion models[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans: IEEE, 2022: 10684-10695.
[8] FOKA A, GRIFFIN G. AI, Cultural Heritage, and Bias: Some Key Queries That Arise from the Use of GenAI[J]. Heritage, 2024, 7(11): 6158-6176.
[9] STEYERL H. In defense of the poor image[J]. e-flux journal, 2009(10): 1-9.
[10] Eco, U. The Open Work[M]. Anna Cancogni, trans. Cambridge, MA: Harvard University Press, 1989.
[11] COLITTI S, LENGUA M, FORMIA E, et al. Designing with Absence: Advanced Design Approaches to Missing Data in Digital Cultural Heritage[J]. Heritage, 2025, 8(12): 536.
[12] BENJAMIN W. The work of art in the age of mechanical reproduction[M]// Illuminations. ZOHN H, trans. New York: Schocken Books, 1969: 217-251.
[13] CARROLL N. Humour: a very short introduction[M]. Oxford: Oxford University Press, 2014.
[14] SHKLOVSKY V. Art as device[M]//Theory of Prose. SHER B, trans. Champaign: Dalkey Archive Press, 1990: 1-14.
[15] HUI Y. The Question Concerning Technology in China[M]. Falmouth: Urbanomic, 2016.
[16] DUCHAMP M. The creative act[C]//Convention of the American Federation of Arts. Houston, 1957.
[17] ROGERS R A. From cultural exchange to transculturation: a review and reconceptualization of cultural appropriation[J]. Communication Theory, 2006, 16(4): 474-503.
[18] KOREN L. Wabi-Sabi for Artists, Designers, Poets & Philosophers[M]. Berkeley: Stone Bridge Press, 1994.
[19] KRAUSS R. Under blue cup[M]. Cambridge, MA: MIT Press, 2011.
[20] Li Huaqing. Research on Cultural and Creative Design of Han Pictorial Stones in Wu’s Shrine Based on Semiotics [J]. Design, 2025.
[21] Wang Ruiqing, Qiu Chunting. Extraction and Design Transformation of Physical Culture Genes of North Shaanxi Han Pictorial Stones Based on AHP Analytic Hierarchy Process [J]. Packaging Engineering, 2025.