Method for evaluating jade color carving material fidelity based on V-O-I protocol

Abstract

This application provides a method for evaluating the fidelity of jadeite carving based on the V-O-I protocol, relating to the field of image processing. The method includes: acquiring a material data package, which includes: a raw stone image, a contour mask, a color-carving mask, a crack mask, a backlit perspective view, and a light-transmitting mask; converting the material data package into knowledge layer constraints based on process rules; mapping the knowledge layer constraints to control handles; generating a design image using at least one generative AI model under the control handles; evaluating the design image based on four quantitative indicators: color utilization rate, contour deviation, consistency of texture, and consistency of intent, obtaining evaluation results; and combining failure mode analysis to perform attribution assessment of the evaluation results. The technical solution of this application provides the first quantifiable and comparable quality benchmark for the application of AIGC in traditional crafts, enabling designers to objectively evaluate and select AI solutions, significantly reducing trial-and-error costs and material risks.

Description

Technical Field

[0001] This application relates to the field of image processing, and in particular to a method for evaluating the fidelity of jadeite carving based on the VOI protocol. Background Technology

[0002] With the rapid development of generative artificial intelligence (AI) technology, its application in arts and crafts design is gradually increasing. However, traditional crafts such as jade carving, which relies heavily on the properties of materials (such as texture, clarity, color, and cracks), have extremely high requirements. Existing AIGC models often ignore the constraints of physical materials during generative design, making the generated solutions difficult to implement in actual crafts. Current research focuses mainly on aesthetic evaluation or cultural imagery expression, lacking systematic evaluation methods for material-dependent tasks. For example, existing technologies often use general image metrics such as FID scores or CLIP similarity to evaluate the quality of the generated material, but these methods cannot reflect the fidelity of the material, especially whether the natural characteristics of the raw material are respected in subtractive processing. Summary of the Invention

[0003] The purpose of this invention is to address the lack of an effective method and system for quantitatively evaluating the compliance of the physical characteristics of original materials with generative AI in traditional subtractive manufacturing processes, and to provide a method for evaluating the material fidelity of jadeite carving based on the VOI protocol.

[0004] The above-mentioned objective of this application is achieved through the following technical solution: S1: Obtain material data package, which includes: raw stone image, contour mask, color mask, crack mask, backlight perspective view and light transmission mask; S2: Convert material data packages into knowledge layer constraints based on process rules; S3: Map knowledge layer constraints to control handles; S4: Use at least one generative AI model to generate design images under the control handle; S5: Based on the utilization rate of color, outline deviation, germplasm consistency, and intention Figure 1 The design images were evaluated using four quantitative indicators of consistency, and the evaluation results were obtained. S6: Combine failure mode analysis to conduct attribution assessment of the evaluation results.

[0005] Optionally, step S1 includes: Each image in the material data package has undergone spatial alignment, resolution unification, and color gamut calibration, and is accompanied by metadata files; The construction of the standard material data package specifically includes: A front view image of the raw jadeite was acquired under standard light source conditions, and color temperature and white balance were corrected to obtain the corrected front view image. The corrected front view image is segmented to generate a contour mask. The image segmentation process includes preliminary contour extraction using the Canny edge detection algorithm, supplemented by manual correction; In the CIELAB color space, threshold segmentation and region growing are performed on the front view image to identify and generate a color mask. Based on high-contrast images, crack region masks are generated by manual annotation by domain experts; The generated contour mask, shading region mask, and crack region mask are spatially aligned with the front view image, their resolution is unified to the preset size, and then packaged into a structured data package.

[0006] Optionally, step S2 includes: The process rules include: Clever use rule: The area corresponding to the clever use area mask must be assigned a core entity object related to the preset design theme in the candidate design scheme; Defect avoidance rules: The area corresponding to the crack area mask should be avoided in the candidate design scheme as a critical load-bearing structure or a deeply engraved area; The rule of parallel texture composition requires that the main shape and lines of the candidate design should be consistent with the implicit texture trend inside the raw material.

[0007] Optionally, step S3 includes: The control handle includes a basic prompt layer, a style prompt layer, and a material constraint layer.

[0008] Optionally, step S5 includes: The formula for calculating the utilization rate of the color is: in To generate a mask for the highlighted areas in the image, This is a color mask for the raw material; the closer CU is to 1, the more fully and accurately the color is utilized. The formula for calculating the profile deviation is: in For symmetric Hausdorff distance, These are the generated profile and the raw material profile, respectively. The OD is the diagonal length; the smaller the OD, the higher the fidelity of the geometric profile. The formula for calculating germplasm uniformity is: in The deep texture feature vector is extracted through a pre-trained convolutional neural network; The closer it is to 1, the more consistent the visual quality of the generated result is with the real jade material; The meaning Figure 1The consistency is calculated as follows: The obtained design intent text T is encoded to obtain a text embedding vector. ; Design images Encode to obtain image embedding vector The encoder can be a multimodal alignment network or an image-text retrieval network; The design intent text T includes: subject name, key entities, and culturally significant terms; Calculating meaning using cosine similarity Figure 1 Desire:

[0009] Expert scores were obtained using a structured expert rating method. At least three domain experts will score the scores on a 0-1 scale according to a pre-defined scoring criterion, and the results will be calculated. The coefficient is used to assess the reliability of the rating; meaning Figure 1 IS is consistent with With expert rating The weighted calculation was used; when expert scores were unavailable, the following method was employed. As IS output; when a graphic encoder is missing, use As IS output.

[0010] Optionally, step S6 includes: The failure modes include: boundary overflow, form disregard, theme deviation, semantic misalignment, process paradox, and texture fictitiousness.

[0011] An electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to perform a method for evaluating the fidelity of jadeite carving based on the VOI protocol.

[0012] A computer-readable storage medium storing instructions that, when executed, perform a method for evaluating the fidelity of jadeite carving based on the VOI protocol.

[0013] The beneficial effects of the technical solution provided in this application are: Proposes and quantifies material fidelity evaluation dimensions: Addressing the core issue of "material distortion" in subtractive processing by AIGC, it innovatively decomposes the process principle of "tailoring techniques to materials" into calculable dimensions such as "color utilization rate," "contour deviation," "material consistency," and "intent." Figure 1 The four indicators of "consistency" have established a system that transitions from qualitative experience to quantitative analysis.

[0014] Constructing a process-oriented VOI evaluation protocol and data package: Proposing a three-layer evaluation framework of "vision (V)-object (O)-intention (I)" and designing a "standardized material data package" containing multi-level masks, transforming the tacit knowledge of craftsmen (such as crack avoidance and grain alignment) into machine-executable constraint rules, thus realizing the standardization and reproducibility of the evaluation process.

[0015] This study reveals the "failure fingerprint" and the model's capability boundaries: Through a "5×3×3" comparative experiment, it not only quantifies model performance but also systematically summarizes six failure modes, including "boundary overflow" and "texture fictitiousness," mapping them to the three major cognitive gaps of "perception-knowledge-generation," providing precise diagnostic basis for model optimization. This method provides the first quantifiable and comparable quality benchmark for the application of AIGC in traditional manufacturing processes, enabling designers to objectively evaluate and select AI solutions, significantly reducing trial-and-error costs and material risks. Attached Figure Description

[0016] The present application will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of an embodiment of this application; Figure 2 This is a schematic diagram of the material data package in the embodiments of this application; Figure 3 This is a diagram showing the condition control hierarchy and representative results in the embodiments of this application; Figure 4 This is a box plot of quantitative indicators in the embodiments of this application; Figure 5 This is a radar chart of the model performance in the embodiments of this application; Figure 6 This is a failure mode ratio diagram in the embodiments of this application; Figure 7 These are before-and-after comparison examples of applying techniques according to the specific materials in this application. Figure 8 This is a schematic diagram of the electronic device structure in the embodiments of this application. Detailed Implementation

[0017] To provide a clearer understanding of the technical features, objectives, and effects of this application, the specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0018] The embodiments of this application provide a method for evaluating the fidelity of jadeite carving based on the VOI protocol.

[0019] Please refer to Figure 1 , Figure 1This is a flowchart of a method for evaluating the fidelity of jadeite carving based on the VOI protocol, as described in an embodiment of this application, including: S1: Obtain material data package, which includes: raw stone image, contour mask, color mask, crack mask, backlight perspective view and light transmission mask; As one embodiment, a standardized material data package is constructed: a standard image of the target raw material is acquired, and based on the standard image, a structured data set containing at least a contour mask and a feature region mask is generated; the contour mask is used to characterize the original outer contour geometric information of the raw material; the feature region mask is used to characterize regions on the raw material that have specific process value, including at least one of color-changing regions and crack regions.

[0020] S2: Convert material data packages into knowledge layer constraints based on process rules; As one embodiment, a multi-level evaluation protocol is defined: a VOI evaluation protocol is established based on a visual feature layer, an object semantic layer, and a process intent layer; the visual feature layer corresponds to the original material features in the material data package; the object semantic layer corresponds to the rules for mapping material features to specific design themes and elements; and the process intent layer corresponds to the cultural narrative and process feasibility constraints that the final design scheme must follow.

[0021] S3: Map knowledge layer constraints to control handles; S4: Use at least one generative AI model to generate design images under the control handle; S5: Based on the utilization rate of color, outline deviation, germplasm consistency, and intention Figure 1 The design images were evaluated using four quantitative indicators of consistency, and the evaluation results were obtained. S6: Combine failure mode analysis to conduct attribution assessment of the evaluation results.

[0022] This application provides an embodiment as follows: To scientifically evaluate the material fidelity of generative AI in the "skillful carving" task, it is necessary to transform the tacit principle of "adapting techniques to the material" in jade carving practice into a set of calculable and reproducible standardized protocols. This research constructs an evaluation method with VOI (Visual-Object-Intention) as its core logic, aiming to establish a complete data link from objective material evidence to structured process semantics, and then to generative guidelines that can be accurately executed by the AIGC model. The core of this method lies in transforming the evaluation process from a "black box" behavior that relies on fuzzy prompts into a transparent scientific experiment based on standardized data input and structured parameter control.

[0023] Step S1 includes: Each image in the material data package has undergone spatial alignment, resolution unification, and color gamut calibration, and is accompanied by metadata files; The construction of the standard material data package specifically includes: A front view image of the raw jadeite was acquired under standard light source conditions, and color temperature and white balance were corrected to obtain the corrected front view image. The corrected front view image is segmented to generate a contour mask. The image segmentation process includes preliminary contour extraction using the Canny edge detection algorithm, supplemented by manual correction; In the CIELAB color space, threshold segmentation and region growing are performed on the front view image to identify and generate a color mask. Based on high-contrast images, crack region masks are generated by manual annotation by domain experts; The generated contour mask, shading region mask, and crack region mask are spatially aligned with the front view image, their resolution is unified to the preset size, and then packaged into a structured data package.

[0024] As one example, the overall process of the V-O-I evaluation method is as follows: Figure 2 As shown, its design strictly aligns with the principles of knowledge-driven design, deconstructing the complex evaluation task into three clear functional layers: The perception layer transforms the physical jade raw material into a machine-readable "material data package," serving as the sole "factual benchmark" for all subsequent evaluation and generation tasks. The knowledge layer extracts expert experience and craft rules regarding "skillful use of color and texture" in jade carving into a series of executable logical constraints. The generation layer maps the logical constraints of the knowledge layer into specific, controllable "generation control handles" within the AIGC model, and systematically evaluates the model's material fidelity through progressive experimental levels. This process ensures the fairness and reproducibility of the evaluation, providing a methodological foundation for quantitatively analyzing the performance boundaries of AIGC in complex craft tasks.

[0025] The data package includes a metadata file (meta.yml) that records the raw material dimensions, imaging equipment, light source parameters, and image processing workflow to ensure reproducibility. Since the original high-resolution photographs are used under license from a partner, this study only discloses the mask and parameter files generated by the authors. Third parties can fully reproduce the evaluation index calculation and model evaluation process based on the structure described in this section and the algorithm description in the appendix.

[0026] As one example, the data package used in this study is based on the front view image of the authorized jadeite raw material sample "Seeking Methods," which was generated by the author using Photoshop to perform lighting correction, color gamut equalization, and multi-layer mask segmentation. This data package serves as the underlying input of the V-O-I evaluation system, standardizing and defining the geometric shape, color distribution, fracture structure, and light transmission characteristics of the raw material. Its specific composition is as follows (see...). Figure 2 The following images are used to calculate the profile: Raw image: A front view of the raw material taken under standard light source (D65, 5500 K), corrected for color temperature and white balance, serving as the registration reference for the mask; Silhouette mask: Obtained based on Canny edge detection and manual correction, accurately defining the two-dimensional geometric boundary of the raw material, serving as the reference true value for calculating the outline deviation (OD); Qiaose mask: Generated in the CIELAB color space using threshold segmentation and region growing algorithms, marking the core qiaose area, serving as the main basis for qiaose utilization (CU); Crack mask: Manually drawn by domain experts on a high-contrast image, marking internal cracks and areas to be avoided, used to assess process feasibility; Backlight perspective view: Obtained by taking a raw material image under backlight conditions and normalizing the brightness, reflecting the thickness and structural distribution; Translucency mask: Calculated based on the grayscale distribution of the backlight image, using it to calculate the germplasm consistency index (SC, SD). All images are calibrated to a uniform resolution (512 × 512 px) and color gamut, and spatially aligned using the same affine matrix. Figure 2 Material data package diagram (taking the "method of finding" case as an example).

[0027] Step S2 includes: The process rules include: Clever use rule: The area corresponding to the clever use area mask must be assigned a core entity object related to the preset design theme in the candidate design scheme; Defect avoidance rules: The area corresponding to the crack area mask should be avoided in the candidate design scheme as a critical load-bearing structure or a deeply engraved area; The rule of parallel texture composition requires that the main shape and lines of the candidate design should be consistent with the implicit texture trend inside the raw material.

[0028] As one example, after obtaining the standardized "material data package," the role of the knowledge layer is to "translate" it into machine-understandable design constraints. This process draws on the ideas of knowledge-driven design and ontology engineering[ ], aiming to formalize the experience rules of jade carvers in "adapting their craft to the material" into executable logical instructions. These rules serve as a "checklist" for judging whether the AIGC-generated scheme conforms to domain knowledge during evaluation. **Rules for utilizing natural colors (V→O):** The natural color mask area (V layer feature) should be assigned to the core theme (O layer entity), not the background. **Rules for avoiding flaws (V→O):** When generating a three-dimensional shape, the crack mask area should avoid deep carving or being used as a key load-bearing point. **Rules for aligning with the grain (V→O):** The lines of the main design element should be consistent with the direction of the internal texture of the jade material.

[0029] Step S3 includes: The control handle includes a basic prompt layer, a style prompt layer, and a material constraint layer.

[0030] As one implementation, after obtaining the formal rules of the knowledge layer, the task of the generative layer is to convert these rules into control handles that can be invoked in the AIGC model. Three levels of control handles are defined—the Base layer, the Style layer, and the Material Constraint layer. Figure 3 Each handle is mapped to model parameters (such as stylize, ref_image, mask_strength, CU, RCF, etc.), thereby guiding the model's output in stages during the generation process. These handles are the specific technical implementation of "knowledge layer rules" in the AIGC model. They map abstract design intentions to specific and controllable parameters in the model, and their design philosophy is consistent with the human-computer collaboration concept in the step-by-step generation framework.

[0031] Step S5 includes: The formula for calculating the utilization rate of the color is: in To generate a mask for the highlighted areas in the image, This is a color mask for the raw material; the closer CU is to 1, the more fully and accurately the color is utilized. The formula for calculating the profile deviation is: in For symmetric Hausdorff distance, These are the generated profile and the raw material profile, respectively. The OD is the diagonal length; the smaller the OD, the higher the fidelity of the geometric profile. The formula for calculating germplasm uniformity is: in The deep texture feature vector is extracted through a pre-trained convolutional neural network; The closer it is to 1, the more consistent the visual quality of the generated result is with the real jade material; The meaning Figure 1 The consistency is calculated as follows: The obtained design intent text T is encoded to obtain a text embedding vector. ; Design images Encode to obtain image embedding vector The encoder can be a multimodal alignment network or an image-text retrieval network; The design intent text T includes: subject name, key entities, and culturally significant terms; Calculating meaning using cosine similarity Figure 1 Desire:

[0032] Expert scores were obtained using a structured expert rating method. At least three domain experts will score the scores on a 0-1 scale according to a pre-defined scoring criterion, and the results will be calculated. The coefficient is used to assess the reliability of the rating; meaning Figure 1 IS is consistent with With expert rating The weighted calculation was used; when expert scores were unavailable, the following method was employed. As IS output; when a graphic encoder is missing, use As IS output.

[0033] This application provides an embodiment as follows: To systematically evaluate the material fidelity performance of generative AI in the "color and texture carving" task, this study constructs a comprehensive evaluation protocol based on a V-O-I three-layer process, deconstructing "material fidelity" into four computable core quantitative indicators (CU, OD, SC, IS) and six typical failure modes. This protocol realizes a quantifiable, comparable, and reproducible evaluation system from perceived features to generated results.

[0034] This application provides an embodiment as follows, IS: meaning Figure 1 Intention Similarity (the higher the better) Concept definition: To what extent the generated solution conforms to the pre-set design intent and cultural narrative (such as the thematic logic of "Seeking the Law").

[0035] Calculation method: The structured expert scoring method was adopted, and three experts in the field of jade carving were invited to score each generated sample on a scale of 0-1 according to the scoring criteria.

[0036] Rating reliability The calculations showed that all results were greater than 0.85, indicating good subjective consistency. Ultimately, the IS (Information Standard) used the average of the three experts' scores as the evaluation result.

[0037] Step S6 includes: The failure modes include: boundary overflow, form disregard, theme deviation, semantic misalignment, process paradox, and texture fictitiousness.

[0038] This application provides an embodiment as follows: In addition to quantitative evaluation, to reveal the potential limitations of the AIGC model in traditional process scenarios, this study further analyzes the generated results using failure modes. Combining controlled experiments and expert review, six typical failure types are summarized and corresponding to the three major gaps of "perception-knowledge-generation": Through this standardized process and evaluation protocol from the perception layer (V) to the knowledge layer (O) and then to the generation layer (I), this study achieves a reproducible, comparable, and diagnosable quantitative evaluation of the material fidelity of different AIGC models in the "color carving" task, providing a unified quality benchmark and failure attribution framework for the digitization of traditional processes.

[0039] Table 1 shows six types of failure modes and their corresponding "perception-knowledge-generation" gaps.

[0040] In one embodiment, to verify whether the proposed V-O-I evaluation protocol can effectively characterize material fidelity and to further test the inference that "the general AIGC model has systematic distortion in the 'color carving' task," a set of control experiments was built around typical process scenarios. The experiment focuses on comparing the performance differences of different models at different control levels under strict control of materials, prompts, and parameters, and observing their capability boundaries in material compliance and semantic consistency.

[0041] In one implementation, the electronic device automatically determines the failure mode based on quantitative indicators and the topological relationship of the mask: F1 boundary overflow: When the contour fidelity is below a threshold (e.g., ... (or the proportion of the generated contour that exceeds the raw material contour area exceeds a threshold) Time determination; F2 form ignored: when OD> Furthermore, the angle between the direction of the generated main skeleton and the direction of the raw material main axis is greater than [missing information]. Time determination; F3 Texture Imitation: When SC< Furthermore, the difference between the high-frequency texture energy of the generated image and the texture energy of the original image exceeds [a certain threshold]. Time determination; F4 Topic Deviation: When IS_machine < (or IS<) When to determine; F5 Process Paradox: When the generated scheme falls into the "critical load-bearing structure / deep carving area" in the crack mask region or the minimum structural thickness is below the process threshold. Time determination; F6 Semantic Misalignment: When the masked area does not contain the core entity object, or the proportion of the core entity object falling in the non-masked area exceeds a threshold. Time-based determination.

[0042] Furthermore, the electronic device outputs a failure report, including: failure category, trigger rule number, trigger threshold, corresponding location (pixel / mask area), and confidence level.

[0043] Experimental subjects and parameter settings: The experiment adopted a factorial design of "5 models × 3 control levels × 3 repeated generation" to ensure the statistical significance and robustness of the results.

[0044] Experimental Models: This study selected five mainstream AIGC models that are representative of the design and creative fields as evaluation subjects: Midjourney: A typical closed-source commercial model, widely used for its image stylization capabilities and strong aesthetic transfer ability. This study mainly observes whether it can maintain adherence to contours and color areas in material-constrained processing tasks.

[0045] Lovart, as an agent-type tool attempting to serve the design process, can decompose natural language requirements into several structured tasks and invoke various models to complete the generation. This research focuses on its understanding of complex process instructions under highly constrained material scenarios, as well as its semantic control capabilities.

[0046] Google (Imagen series): Represents the large-scale text-to-image model route. Its advantages lie in semantic parsing and overall image construction capabilities. However, whether its generation logic can naturally adapt to the material requirements of "subtractive processing" is one of the key issues explored in this study.

[0047] Seedream 4.0: One of the fastest-growing commercial models in China in recent years, especially active in product rendering and style generation. Its output often has obvious image filter characteristics. This study focuses on whether it can stably follow the shape and color region boundaries in tasks that are highly dependent on material constraints.

[0048] GPT-5 (OpenAI Generated Graph Series): Represents the latest version of a large-scale cross-modal model. This study does not aim to evaluate its internal architecture from a technical perspective, but rather to observe its adaptability as a "general-purpose generator" in material-dependent process tasks, with particular attention to its semantic consistency and material representation capabilities.

[0049] Control Hierarchy: The experiment strictly followed the three progressive generation control hierarchies defined in Chapter 2 (see Table 1) to test the model's response capability when receiving constraints of different granularities. Level 1 (Basic): Provides only core theme keywords and basic style keywords. Level 2 (Style): Based on Level 1, add style reference images.

[0050] Level 3 (Material): Based on Level 2, it adds raw material images and outlines, as well as color masks, to achieve the highest precision conditional control.

[0051] Experimental Samples and Procedures: Sample: The experiment used a specific tricolor jadeite raw material (number J001) as the unified evaluation object, and its "material data package" had been constructed in advance.

[0052] Process: For each model, at each control level, 3 images are independently generated using the same random seed to reduce bias caused by randomness. A total of 5 models × 3 levels × 3 repetitions = 45 images are obtained for evaluation.

[0053] Review process and reliability test: In the review stage, this study adopted a structured expert scoring method, which transformed the four evaluation dimensions that could originally be calculated by algorithms into quantitative scores with a continuous scale of 0–1.

[0054] The judging panel consisted of three experts with over ten years of experience in jade carving design, craftsmanship evaluation, or related teaching, possessing extensive practical experience and aesthetic judgment in "skillful use of natural colors in jade carving." Before the official judging began, a briefing session was held focusing on "understanding the evaluation agreement" and "calibrating scoring examples." Through repeated comparisons of typical samples, the three experts reached a consensus on the meaning of the four indicators, the scoring criteria, and the use of the 0-1 continuous scale, ensuring the reliability of the subsequent judging.

[0055] Evaluation Metrics and Process: The review process was double-blind, meaning experts were unaware of the model from which the samples came from or the control level at which they were generated. Each expert independently assigned four scores to all 45 images, assigning values ​​between 0 and 1 (0 indicating complete non-compliance with material or semantic logic, and 1 indicating high consistency) based on the defined meanings of the metrics. Specifically, these included: Cubic Color Utilization (CU), Contour Preservation (1... OD), germplasm uniformity (SC) and intention Figure 1 Consistency (IS). This approach retains the flexibility of qualitative judgment while ensuring consistency of measurement among different experts.

[0056] Reliability test: To confirm the consistency of expert ratings, Krippendorff's α coefficient was calculated for the ratings of the four indicators. The results showed that the α coefficient for all indicators was above 0.8, with the IS indicator reaching an α value of 0.87. This result indicates that the differences in ratings among experts are small, and the overall rating system has good reliability, serving as a reliable basis for subsequent statistical analysis.

[0057] Statistical and Analytical Methods: To understand model performance from multiple perspectives, this study employed a combined analytical strategy. First, descriptive statistics were performed to calculate basic characteristics such as the mean and standard deviation of various indicators to obtain an initial impression of the performance distribution. Then, inferential statistical methods such as analysis of variance were used to test whether the differences between different models and control levels were statistically significant. Finally, graphical presentation was employed, including box plots displaying the data distribution characteristics. Figure 4 ), a radar chart describing the overall performance profile ( Figure 5 ), and a stack percentage chart showing the proportion of various failure modes ( Figure 6 These analyses together form a multi-dimensional framework for observing model performance and lay the foundation for subsequent qualitative attribution.

[0058] Results and Analysis: This chapter presents the material fidelity performance of five AIGC models in the "Skillful Carving" task using a dual approach of quantitative indicators and failure modes, verifying the three major gaps of "perception-knowledge-generation" proposed in Chapter 2.

[0059] Overall comparison of quantitative indicators: After calculating four core indicators (CU / OD / SC / IS) for 45 generated samples, a global view of model performance can be obtained. Overall, the performance of each model in terms of material fidelity is relatively low, which quantitatively supports the judgment that "general AIGC has systematic limitations in high-constraint process tasks".

[0060] Key indicator performance: such as box plots and radar charts ( Figure 4 , Figure 5 As shown in the figure, there are significant differences among the models, but no single model has yet achieved a comprehensive lead across all four metrics. IS (meaning...) Figure 1 Consistency and CU (color utilization) are the most effective ways to distinguish model capabilities; most models perform better in these two aspects. The mean score is below 0.4, indicating difficulty in simultaneously understanding cultural narratives and accurately utilizing color. Lovart (Agent type) is relatively superior in IS, indicating an advantage in task decomposition and complex instruction adherence; however, it is still insufficient in OD (outline deviation) and SC (genetic consistency), suggesting that its respect for physical constraints and material texture is still limited.

[0061] Impact of Control Level: As the control level progresses from "Basic" to "Style" to "Material," the average performance of all models improves slightly, especially in contour fidelity (OD reduction). However, the improvement is limited and cannot fundamentally change low-level performance. This indicates that simply increasing control signals without bridging the cognitive gaps in perception / knowledge / generation of the model is insufficient to achieve a qualitative leap.

[0062] Failure Mode and Effects Attribution Analysis: Qualitative analysis of all generated results yielded six typical failure types (see...). Figure 6 proportional distribution Figure 7 (Case study panel), and mapped to the three major gaps of "perception-knowledge-generation": Perception gaps: Boundary overflow and shape disregard occur most frequently, accounting for over 45% in total. AIGC struggles to treat masks and shapes as hard constraints.

[0063] Knowledge gaps: Thematic deviation and semantic misalignment reflect the model's shortcomings in cultural semantic understanding and layout. For example, it failed to establish a narrative mapping of "yellow jade → desert / hardship, emerald green → oasis / hope", resulting in a disconnect from the theme of "Seeking the Dharma".

[0064] Generation gaps: The paradox of process and the fiction of texture reflect the conflict between the logic of generation and the physical world: the former cannot be achieved in subtractive manufacturing or is insufficient in strength; the latter presents an imagined texture that does not conform to the optical properties of real germplasm.

[0065] The technical "fingerprints" of different models: A more valuable finding is that models from different technical approaches exhibit unique "failure fingerprints," which provide clues for understanding their underlying mechanisms. Agent-based models, exemplified by Lovart, exhibit a relatively high rate of semantic misalignment, indicating that even when they understand thematic elements, they struggle to achieve precise semantic mapping in spatial layout. Closed-source business models, such as Midjourney, show the most significant thematic deviation. It is speculated that in pursuit of broad universality and aesthetic appeal, these models employ "fuzzification" in semantic understanding, making them highly susceptible to "creative deviation" when faced with highly specific cultural instructions.

[0066] Some models exhibit a relatively higher failure rate in F1 "boundary overflow" and F3 "texture fictitiousness," indicating that they possess strong pixel-level shaping capabilities in generating local details, but are also more prone to "over-exploiting," generating textures or shapes beyond the material's potential. Other models are more susceptible to F4 "subject deviation" or F6 "semantic misalignment," reflecting limitations in their semantic alignment capabilities. These different technical approaches thus present relatively stable "failure fingerprints," providing indirect evidence for understanding their internal generation logic.

[0067] In summary, combining quantitative results with typical failure samples, this study systematically presents the limitations of current generative models in the "skillful carving" task through three paths: perception, knowledge, and generation. These findings not only describe the performance boundaries of the models but also reflect the paradigmatic tension between their underlying "additive logic" and the "subtractive logic" of jade carving.

[0068] From "Technological Tool" to "Knowledge Carrier": This study shows that the failure of general AIGC in the "skillful carving" scenario does not originate from the quality of generation, but from the fundamental conflict between its additive generation paradigm and the subtractive philosophy of "tailoring craftsmanship to the material." AIGC is closer to a "free creator" in an open context than a "collaborative craftsman" who respects material constraints. Therefore, simply increasing the model size and data volume cannot overcome the dual constraints of materials and processes. The next step should be to promote the transformation of AIGC from a "technical tool" to a "knowledge carrier," deeply embedding domain knowledge into the entire process of perception-reasoning-generation: constructing "material vocabulary" (material, cracks, light transmission) and "cultural grammar" (crack avoidance, grain alignment, imagery), enabling the model to "understand" the material and adhere to the process. This path is consistent with the process knowledge service paradigm of "knowledge deconstruction-reconstruction-restoration-innovation."

[0069] Establishing a "Quality Benchmark" for the Digitization of Traditional Crafts: The evaluation protocol proposed in this study is not only used to "diagnose" the current state of generative models, but can also serve as a quality benchmark in the digitization process of traditional crafts. Faced with diverse model choices, jade carvers and designers can use this protocol to objectively compare the models' performance in terms of material fidelity, reducing creative losses or material risks caused by inappropriate tool selection. For model developers, this protocol is also directional: from the "three gaps" and "failure fingerprints," the shortcomings of the model at the perception, knowledge, or generation levels can be directly located. For example, the "perception gap" suggests the need to strengthen the model's understanding of the optical properties and spatial constraints of materials; the "knowledge gap" points to the deep coupling between the semantic system and process rules; and the "generation gap" can serve as a key clue for optimizing the generator structure.

[0070] Constructing a “creative dialogue” model of human-machine collaboration: In the foreseeable future, AIGC will find it difficult to undertake complex subtractive creation tasks on its own, but it has obvious advantages in generating multiple schemes and rapid iteration under limited conditions. Based on the V–O–I three-dimensional material-knowledge-intention modeling in this study, a human-machine collaboration model characterized by “creative dialogue” is proposed: the I layer (human) is responsible for setting cultural narratives, intentions and aesthetic directions; the O layer (knowledge) provides process rules, semantic logic and constraints; the V layer (material) presents the real material boundaries in the form of masks or measurements; the G layer (model) generates candidates under these constraints. This model is highly consistent with the “knowledge service driven” process restoration and innovation path in related studies [9], and also with the concept of “living inheritance” in the protection of intangible cultural heritage—activating the creativity of the process with the help of technology, rather than letting the process stay in a fixed pattern.

[0071] This paper takes the highly material-sensitive scenario of "skillful carving using the natural colors of jade" as an example to propose and validate a computationally achievable evaluation protocol for material fidelity. Through a "5×3×3" controlled experiment, it is confirmed from both quantitative indicators and failure modes that current general AIGC models have structural deficiencies in the three dimensions of perception, knowledge, and generation. The contradiction between its "additive generation" and the "subtractive logic" of jade carving is the most fundamental reason for material distortion. The V-O-I evaluation paradigm provides a methodological channel for making the tacit experience of jade carvers explicit and computationally achievable, and also provides an important foundation for building a knowledge-driven generative framework oriented towards craftsmanship in the future. Although this research focuses on a single case, "skillful carving using the natural colors of jade," its method has potential transferability across materials and processes, and is expected to support the digital innovation of the "subtractive logic" in more traditional craft scenarios.

[0072] This application also discloses an electronic device. (See reference...) Figure 8 , Figure 8 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0073] The communication bus 502 is used to enable communication between these components.

[0074] The user interface 503 may include a display screen, and optionally, the user interface 503 may also include a standard wired interface or a wireless interface.

[0075] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0076] This application also discloses a computer-readable storage medium storing multiple instructions adapted for loading by a processor to execute the above-described method for evaluating the fidelity of jadeite carving based on the VOI protocol.

[0077] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure.

[0078] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A method for evaluating the fidelity of jadeite carving based on the VOI protocol, characterized in that, The method includes the following steps: S1: Obtain material data package, which includes: raw stone image, contour mask, color mask, crack mask, backlight perspective view and light transmission mask; S2: Convert material data packages into knowledge layer constraints based on process rules; S3: Map knowledge layer constraints to control handles; S4: Use at least one generative AI model to generate design images under the control handle; S5: The design image is evaluated based on four quantitative indicators: color utilization rate, outline deviation, genetic consistency, and intention consistency, and the evaluation results are obtained. S6: Combine failure mode analysis to conduct attribution assessment of the evaluation results.

2. The method for evaluating the fidelity of jadeite carving based on the VOI protocol as described in claim 1, characterized in that, Step S1 includes: Each image in the material data package has undergone spatial alignment, resolution unification, and color gamut calibration, and is accompanied by metadata files; The construction of the standard material data package specifically includes: A front view image of the raw jadeite was acquired under standard light source conditions, and color temperature and white balance were corrected to obtain the corrected front view image. The corrected front view image is segmented to generate a contour mask. The image segmentation process includes preliminary contour extraction using the Canny edge detection algorithm, supplemented by manual correction; In the CIELAB color space, threshold segmentation and region growing are performed on the front view image to identify and generate a color mask. Based on high-contrast images, crack region masks are generated by manual annotation by domain experts; The generated contour mask, shading region mask, and crack region mask are spatially aligned with the front view image, their resolution is unified to the preset size, and then packaged into a structured data package.

3. The method for evaluating the fidelity of jadeite carving based on the VOI protocol as described in claim 2, characterized in that, Step S2 includes: The process rules include: Clever use rule: The area corresponding to the clever use area mask must be assigned a core entity object related to the preset design theme in the candidate design scheme; Defect avoidance rules: The area corresponding to the crack area mask should be avoided in the candidate design scheme as a critical load-bearing structure or a deeply engraved area; The rule of parallel texture composition requires that the main shape and lines of the candidate design should be consistent with the implicit texture trend inside the raw material.

4. The method for evaluating the fidelity of jadeite carving based on the VOI protocol as described in claim 1, characterized in that, Step S3 includes: The control handle includes a basic prompt layer, a style prompt layer, and a material constraint layer.

5. The method for evaluating the fidelity of jadeite carving based on the VOI protocol as described in claim 1, characterized in that, Step S5 includes: The formula for calculating the utilization rate of the color is: in To generate a mask for the highlighted areas in the image, This is a color mask for the raw material; the closer CU is to 1, the more fully and accurately the color is utilized. The formula for calculating the profile deviation is: in For symmetric Hausdorff distance, These are the generated profile and the raw material profile, respectively. The OD is the diagonal length; the smaller the OD, the higher the fidelity of the geometric profile. The formula for calculating germplasm uniformity is: in The deep texture feature vector is extracted through a pre-trained convolutional neural network; The closer it is to 1, the more consistent the visual quality of the generated result is with the real jade material; The calculation method for intent consistency is as follows: The obtained design intent text T is encoded to obtain a text embedding vector. ; Design images Encode to obtain image embedding vector The encoder can be a multimodal alignment network or an image-text retrieval network; The design intent text T includes: subject name, key entities, and culturally significant terms; Calculating intent consistency using cosine similarity: Expert scores were obtained using a structured expert rating method. At least three domain experts will score the scores on a 0-1 scale according to a pre-defined scoring criterion, and the results will be calculated. The coefficient is used to assess the reliability of the rating; Intent Consistency IS is With expert rating The weighted calculation was used; when expert scores were unavailable, the following method was employed. As IS output; when a graphic encoder is missing, use As IS output.

6. The method for evaluating the fidelity of jadeite carving based on the VOI protocol as described in claim 1, characterized in that, Step S6 includes: The failure modes include: boundary overflow, form disregard, theme deviation, semantic misalignment, process paradox, and texture fictitiousness.

7. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to perform the jadeite carving material fidelity evaluation method based on the VOI protocol as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a computer, perform the jadeite color and carving material fidelity evaluation method based on the VOI protocol as described in any one of claims 1-6.

Images