Traditional crafts such as Yi embroidery, Zhuang brocade and Wuqiang woodblock prints keep failing as consumer products. Buyers say they value the heritage and still do not buy. The design literature has a confident explanation. The problem, it says, is "superficial mixing": a traditional motif pasted onto a modern object like a sticker, rather than woven into the object's structure. But no one has ever measured this. Every account is a qualitative case study or an anecdote about local sales. Meanwhile, the computational tools that could help point in opposite directions. Market-driven design generators score new designs by their similarity to past best-sellers, which by construction punishes anything unfamiliar. Cultural knowledge graphs describe how a design is composed but know nothing about markets.
We propose to use a controllable image generator as an experimental instrument rather than as a product. The generator makes pairs of product designs that differ in exactly one structural property. Examples are whether a motif follows the garment's seams or sits centered like a patch, whether a figure is literal or geometrically abstracted, and whether its palette is original or remapped. We then measure which design different consumer groups would buy. This turns the "superficial mixing causes fatigue" claim into a testable causal question. It also yields a preference predictor built on interpretable structural features. We ask whether that predictor holds up on crafts it has never seen, where history-based popularity scores should fail. Scalability enters as a constraint: the structural annotation must be automatic and the generation cheap enough to run thousands of controlled edits. Otherwise the instrument cannot be deployed beyond a hand-labeled pilot.
Core Claim
We hypothesize that, among heritage-derived consumer products, purchase preference depends more on how the motif is structurally integrated into the product than on which motif is used or how much of it appears. Integration here means placement relative to the product's own form, abstraction level, edge blending with the substrate, and negative space. A preference model built on these structural features should also transfer across crafts better than two alternatives: nearest-neighbor popularity estimates, and generic learned aesthetic scores. Both hypotheses can fail, and the experiments are designed so that either failure is informative.
Where This Sits in Prior Work
Three groups of work motivate the gap. The first is qualitative. China-chic argues that cosmetic use of cultural elements produced consumer fatigue. Studies of Yi embroidery with fast fashion, textile heritage patterns and Wuqiang New Year print packaging document heritage goods failing when they stay decorative, ceremonially cut or "rustic." The "cattle" idiom design study specifically blames literal figurative iconography over geometric abstraction. The middle-aged apparel study shows that producers misjudge what a demographic wants. None of these separates the candidate causes, and we do exactly that.
Two papers supply methods. The Painting Language Knowledge Graph (PL-KG) defines seven structural dimensions of an image: composition, shape and form, perspective and space, light and shadow, color, brushwork and texture, and edge relations. It places Chinese and Western conventions side by side. However, it is a fine-art annotation hierarchy with no market link, and its authors acknowledge that manual annotation does not scale. We adopt a product-adapted subset of its dimensions as our intervention axes. We make annotation automatic by reusing the hierarchical description approach of ArtQuant. ArtQuant is a multimodal language model trained to describe perception, cognition and emotion in artworks before scoring them.
Two papers constrain the design. Huang et al.'s trend-driven fashion system couples a fast conditional StyleGAN2 to a popularity module that averages the historical click-through and sales figures of a design's nearest neighbors. This is the natural market scorer, and precisely the one we expect to misread novel heritage designs. Zhang et al. show that preference optimization can align a vision model's embeddings with human aesthetic judgments in retrieval. Their human-preference benchmark, HPIR, uses 150 queries with 30 judgments each, and we adopt that annotation protocol. Their aesthetic notion is generic photographic appeal, which may treat intentional flat perspective or calligraphic distortion as defects. That is a hypothesis we test rather than assume.
The closest overall system would join Huang's popularity loop to PL-KG-conditioned generation. We deliberately do not build an optimizer first. Optimizing against an unvalidated market signal would bake in whatever bias that signal carries.
Research Program
Stimulus construction. We take four craft sources named in the literature: Yi embroidery, Zhuang brocade (Lan & Tan), Wuqiang prints and Thai Pahkahmah weaving (Burapajatana et al.). We apply them to three product substrates: garments drawn from DeepFashion images, sneakers, and gift packaging. DeepFashion's clothing landmarks (DeepFashion) give seam and panel geometry for "placement follows structure" edits.
A diffusion model is a generator that learns to turn random noise into images step by step. We use a diffusion image editor with region masks and control images. From one base design it produces a matched variant that changes a single dimension. The dimensions are placement (structure-following vs. centered patch), abstraction (literal vs. geometric), color (heritage palette vs. palette remapped to the substrate's trend colors), negative space (dense vs. sparse), and edge relations (hard border vs. blended transition). Motif identity and coverage area are also varied independently, so that "which motif" and "how much" compete directly with "how integrated."
Automatic structural annotation. We fine-tune an ArtQuant-style annotator to output PL-KG-derived labels for each design. Its agreement with three expert designers is checked on a 500-image audit set. Every generated pair passes this annotator. A pair is kept only if the target dimension changed and the others did not, which prevents the edit from silently altering several properties at once. This step addresses the scalability concern, because it replaces the manual labeling PL-KG's authors flag as the bottleneck.
Preference measurement. Following the HPIR protocol of 30 judgments per item, we collect pairwise purchase choices from three pre-registered groups. The groups are domestic consumers aged 18–30, domestic consumers aged 45–60, and international (non-Chinese, non-Thai) consumers. Each respondent also gives a willingness-to-pay slider response. The first target is roughly 1,500 minimal pairs, which comes to about 45,000 judgments. Following the small-initial-dataset logic of materials acceleration platforms, a pilot of about 100 pairs sets effect sizes before the full collection.
Preference model. We fit a group-specific Bradley–Terry model, a standard model for pairwise choice:
$$P(a \succ b \mid g) = \sigma\big(w_g^{\top}(\phi(a) - \phi(b))\big).$$
Here $a$ and $b$ are two designs, $g$ is the consumer group, $\phi(\cdot)$ is the vector of automatically annotated structural features plus motif-identity and coverage indicators, $w_g$ is the group's weight vector, and $\sigma$ is the logistic function.
Because each pair differs in one controlled dimension, the corresponding entry of $w_g$ estimates that dimension's causal effect on choice for group $g$. The central quantity is the share of explained preference attributable to integration dimensions versus motif identity and coverage.
Baselines
The naive reference point is chance (50% pairwise agreement), together with a "motif-only" model. The motif-only model is the same Bradley–Terry fit restricted to motif-identity and coverage features, and it is the model the "superficial mixing" story says should be weak.
The first strong competitor is Huang et al.'s retrieval popularity estimator (Huang et al.). It embeds a design, finds its top-K neighbors in a product database, and soft-votes their historical engagement. It is the strongest deployed market signal in the file. Their 100k-shoe database may not be released, so we reimplement it over a public commercial catalog with engagement proxies. We tune K on the pilot split only. Huang et al. report trend accuracy of 0.79 for cStyleGAN2 against 0.82 for an autoregressive baseline, but that measures trend matching, not purchase prediction. We found no reported result on Lacuna for popularity estimation on heritage-derived products.
The second competitor is a preference-aligned vision embedding in the style of Zhang et al.. This is CLIP fine-tuned with direct preference optimization on our training pairs, scored by embedding similarity to an "appealing product" anchor. It is the strongest method in the file for learning human aesthetic preference end to end. Its reported result is a CLIP aesthetic preference rate rising from 62.1% to 71.7% on HPIR. It receives the same training pairs and a hyperparameter budget equal to ours.
The third competitor is ArtQuant's aesthetic score, used zero-shot and fine-tuned. The file reports only training cost (7.2 GPU hours on BAID versus 48 for SAAN), not accuracy numbers.
The key ablation is a pixel-level preference model of the same capacity trained on the same pairs without structural features. It isolates whether the ontology adds anything beyond learned pixels.
Beating a baseline means higher pairwise agreement with held-out human choices, with 95% bootstrap intervals over respondents, under a leave-one-craft-out split.
Why not just…?
Why not just prompt Stable Diffusion with "minimalist sneaker with Yi embroidery" and pick the best outputs? Prompting gives no control over which structural property changes, so the outputs cannot answer the causal question. PL-KG's authors also argue that such models reproduce surface texture rather than compositional grammar. We still include a prompted set as a stimulus condition. If prompted designs are preferred as often as our structure-following edits, that is evidence the integration dimensions matter less than claimed.
Why not just use Huang et al.'s nearest-neighbor popularity score? It is the cheapest and most deployable option, and it may well work on designs close to the catalog. Our concern is that it cannot score what history has not seen. We test this directly by stratifying accuracy by each design's distance to its nearest catalog neighbor. If its accuracy on heritage designs matches its accuracy on mainstream ones, the case for a structural predictor weakens substantially, and we will say so.
Why not just train an end-to-end preference model on pixels with DPO? This is the most serious threat. Given enough pairs, a pixel model may predict within-craft choices as well as ours. The ontology earns its place only if it wins on unseen crafts, or if its attributions agree with the controlled interventions where the pixel model's saliency does not. If the pixel model ties under leave-one-craft-out, we will recast the ontology's contribution as interpretability and editability rather than prediction.
Why not just keep manual designer-artisan collaboration? Governance studies (Wang & Liu; Hu) report that one-off adaptations stay trapped in niche, high-cost markets. Our goal is not to replace designers. We aim to give them measured guidance about which structural choices matter for which buyers. Designer-made adaptations enter the study as an additional held-out test set.
Decisive Checks
The first check is whether integration dimensions explain more held-out preference variance than motif identity plus coverage, for at least two of three consumer groups. If they do not, the central claim fails.
The second check is whether the kNN popularity score's agreement with human choices drops as distance to the catalog grows, while the structural model's agreement stays flat.
The third check is whether the automatic annotator reaches expert-level agreement on the audit set. We measure this by comparing annotator-expert Cohen's kappa with inter-expert kappa. If it does not, the pipeline does not scale and we report the manual-label ceiling instead.
The fourth check is whether the structural model beats the pixel ablation under leave-one-craft-out.
Group-specific weights will additionally test the claim of generational image rigidity. If older buyers reward palette remapping as much as younger ones, producers' conservative assumption is wrong in a measurable way.
References
- China-chic: From Chinese Elements to International Trend
- Study on the Influence of Yi Handmade Embroidery and Fast Fashion Clothing Combined Development on the Achievements of Poverty Reduction in Liangshan Area
- Visual Transformation of Textile Intangible Cultural Heritage Patterns in the Design of Cultural and Creative Products
- Packaging Design of Wuqiang New Year Pictures
- Modern Creative Expression of Chinese Traditional Graphic “Cattle” — Taking the Visual Design of Idioms as an Example
- A Multidimensional Analysis of Middle-Aged Consumers’ Apparel Demand from a Market Adaptation Perspective
- Knowledge Graph for Intelligent Generation of Artistic Image Creation: Constructing a New Annotation Hierarchy
- Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment
- A Trend-Driven Fashion Design System for Rapid Response Marketing in E-commerce
- Aligning Vision Models with Human Aesthetics in Retrieval: Benchmarks and Algorithms
- Exploring the Inheritance and Development of Guangxi Zhuang Brocade
- Development of Local Cultural Woven Textile Product to Contemporary Fashion Product
- DeepFashion: Powering Robust Clothes Recognition and Retrieval with Rich Annotations
- Designing materials acceleration platforms for heterogeneous CO2 photo(thermal)catalysis
- Research on the Path of Revitalizing and Inheriting Rural Intangible Cultural Heritage from the Perspective of Cultural Governance
- Research on the Development and Visual Image of Opera Culture Creative Products