Commercial brands increasingly use image generators to turn traditional crafts, such as ethnic embroidery or folk prints, into products people will actually buy. Two failures recur. When a generator is optimized only for popularity, it strips out the symbolic structure that makes the craft meaningful, and discerning buyers notice. When a craft is preserved unchanged, it stays unsellable. We propose to represent "authenticity" differently. Instead of treating it as a single score to trade off against sales appeal, we treat it as a map written by the craft's own practitioners. The map marks each design element as fixed, adjustable within a stated range, or free to modernize. It is enforced as a hard constraint while images are generated, and market appeal is optimized only inside the space it leaves open.
The payoff is partly technical and partly institutional. Technically, we expect a hard, element-level constraint to produce more designs that both artisans approve and consumers like than a blended score does. Institutionally, the evolutionary-game literature on collaborative governance finds that whether community members keep participating depends mostly on what participation costs them, not on subsidies. An artisan who writes a map once, rather than reviewing every generated design, pays much less to take part. So the note's central measurable claim is a yield-per-effort claim: how many artisan-approved, consumer-preferred designs each hour of artisan time produces.
Core Claim
For commercial adaptation of a craft, authenticity is best represented as a partition of design elements (motif structure, symbolic colors, placement on the garment) into load-bearing and translatable parts, and that partition should be authored by the craft community. We claim three things. First, a generator that enforces this partition as a constraint dominates scalar multi-objective alignment on the joint rate of artisan approval and consumer preference. Second, at an equal budget of artisan time, it also beats generate-then-filter, where artisans simply veto outputs. Third, it only beats these alternatives when the map comes from practitioners, not from brand designers or from statistics learned from data.
Why This Framing, and What It Builds On
The cultural-capital literature motivates the partition hypothesis. Historical adaptations that kept their legitimacy held some elements invariant while translating others. A Brief Discussion on the Art Education of the Qing Imperial Family documents Qing rulers adopting the Han literati's Dong Qichang calligraphy while deliberately keeping Manchu finger painting as an identity marker. The Flow of Institutional Charisma shows marginal performing troupes borrowing the organizational form of Quanzhen Taoism, not its content. A Study on the Adaptation of Legend of the White Snake is the cautionary counterpart. There, the load-bearing element, the protagonist's autonomy, was itself rewritten to fit the dominant order. That case is why the proposal insists the map be authored by practitioners rather than by the party with market power. All three works are qualitative and make no predictions. We do not claim to test their historical theses. We borrow their structural observation as a design hypothesis.
The commercialization literature supplies the target failure. Challenges and Opportunities of Sichuan Opera Face Changing in Haidilao reports average acceptance of 3.95/5, falling below 3.5/5 in coastal regions with their own strong performance traditions. This suggests that consumers with high cultural literacy punish shallow adaptation. The evidence is weak (N=167, skewed toward urban youth), so we treat it as a hypothesis to replicate, not a fact. Exploration on the Intangible Cultural Heritage Technique of She Ethnic Group shows by hand, in a 24-hour studio course, that the She "Three Reds" pattern system can be moved into modern garments while its symbolic logic survives. Our map is essentially that course's "cognition" stage written down in machine-checkable form. Research on the Path of Revitalizing and Inheriting Rural Intangible Cultural Heritage calls for government–enterprise–villager synergy but concedes it has no quantitative data. We supply one concrete mechanism and measure it.
The generation methods constrain and supply the build. Generating Chinese intangible cultural heritage images with structure and color awareness ("Colorful Heritage") adds structure tokens (axial and radial symmetry) and a symbolic color vocabulary to Stable Diffusion XL, a text-to-image diffusion model that generates images by gradually denoising random noise. It fixes compositional chaos and color dilution for 2D art. However, it applies the same grammar to every output and offers no way for a community to declare which elements may change. We reuse its structure and color vocabulary as the vocabulary in which maps are written. A Trend-Driven Fashion Design System for Rapid Response Marketing in E-commerce optimizes designs for popularity using nearest-neighbor retrieval over sales and click-through data. It has no authenticity constraint, and it represents the popularity-only extreme. DyMO schedules several objectives across denoising steps without retraining, and Pareto Multi-Objective Alignment for Language Models finds Pareto trade-offs cheaply. Both treat every objective as a tradeable score. Our proposal differs precisely here: fixed elements are not traded at any price. DeepFashion supplies the garment landmarks needed to specify placement constraints, for example "the phoenix motif stays at the collar."
On the governance side, An Evolutionary Game Study of Multi-Agent Collaborative Disaster Relief Mechanisms finds that the final equilibrium depends on participants' participation costs, while subsidies change only the speed of convergence. Evolutionary game analysis of FinTech transformation finds that cooperation requires keeping supervision costs low. Neither model concerns culture. We take from them the hypothesis that artisan review cost is the variable that decides whether community governance of AI design persists. We therefore measure it directly rather than building a new game model.
Research Program
We work with one craft in depth: She embroidery from Luoyuan and Xiapu counties. It is the case with the richest documented symbolic grammar in the file. A second craft drawn from ICHStyleBench's embroidery styles checks whether results transfer.
The first stage is authoring the map. Artisan panels, working with a facilitator, annotate a set of motifs. For each structural element (symmetry axis, motif components, border rhythm), each color in the symbolic vocabulary, and each landmark placement, they mark it fixed, adjustable (with a stated range, such as a hue band or scale limits), or free. As a control, brand designers annotate the same motifs separately. We measure agreement among artisans. If artisans do not converge on a shared map, the concept fails at this stage.
The second stage is the generator. We start from Colorful Heritage's structure-token and color-embedding setup on Stable Diffusion XL and add constraint enforcement during sampling. At each denoising step, detectors for the structure and color vocabulary estimate whether fixed elements are present and whether adjustable ones are within range. Guidance gradients push violations back, and a final rejection step removes any output that still breaks a fixed element. Appeal is then optimized only in the free and adjustable dimensions, using DyMO-style scheduling with a preference model and a retrieval-based popularity estimate in the style of Huang et al.
The third stage is evaluation, with three independent groups of judges. Artisans judge approval. Consumer panels, stratified by region and cultural familiarity, judge purchase preference. All artisan time is logged in minutes.
Baselines
The naive reference point is off-the-shelf Stable Diffusion XL, prompted with the craft name and a modern garment description. It shows what brands can do today at no cost. We run it with the same prompts, image count, and sampling steps as every other system.
The strongest existing method for authenticity is Colorful Heritage. On ICHStyleBench it reports content preservation of 84.6 against 78.9 for B-LoRA, style alignment of 81.2, a CLIP score of 0.502, and a 34.5% user preference rate Generating Chinese intangible cultural heritage images with structure and color awareness. We will rerun it on the same garment prompts and training data. Beating it means a higher joint artisan-approval-and-consumer-preference rate while staying within two points of its content-preservation score on ICHStyleBench.
The strongest multi-objective competitor is DyMO scheduling with authenticity entered as a scored objective, a classifier trained on the same artisan annotations, next to appeal. For reference, DyMO raises PickScore from 20.73 to 23.07 on Stable Diffusion v1.5 DyMO. We give it the same annotation data and sweep objective weights to trace its full trade-off curve. Our method beats it only if its point lies above that curve, not merely above one weighting.
The popularity-only extreme is the Huang et al. approach of retrieval-scored appeal with no authenticity term. Its shoe dataset is not available for garments, so we reimplement the retrieval scoring over a garment catalog. No result on heritage garments is reported on Lacuna for this system.
Generate-then-filter uses the naive or Colorful Heritage generator, and artisans veto outputs one by one. Its artisan-time budget is set equal to the time spent authoring maps plus the final spot-check review in our method.
There are two ablations. The first isolates who authors the map: identical generator, with maps written by designers, or with a random fixed/free split matched in proportion. The second isolates the adjustable-range category by collapsing maps to fixed/free only. On artisan approval or on consumer preference in high-familiarity regions there is no reported result found on Lacuna for any of these baselines. Those numbers will be new.
Why not just…?
Why not just have artisans veto outputs from a good generator? This is the real threat. If the generator's hit rate is already high, filtering may be cheaper than authoring a map. The yield-per-artisan-hour comparison at a matched budget settles it directly. If filtering matches our method, we will report that the map adds nothing beyond a filter. We expect it to lose because authoring cost is paid once and then amortized over thousands of designs, and because the governance results in An Evolutionary Game Study of Multi-Agent Collaborative Disaster Relief Mechanisms suggest that per-item review cost is what erodes participation over time.
Why not just add an authenticity score as another reward? A blended objective lets a large gain in appeal buy away a sacred element, and the Haidilao data suggests that is exactly the trade high-literacy consumers punish. Pareto Multi-Objective Alignment for Language Models shows that scalarized objectives oscillate and can collapse weaker objectives. Our DyMO weight sweep tests this directly. If some weighting reaches our joint rate, we will concede that the constraint is not needed.
Why not just use Colorful Heritage as is? Its grammar is uniform. It enforces symmetry and palette everywhere, which should leave little room for the modernization that functional garments need (Visual Transformation of Textile Intangible Cultural Heritage Patterns and Yi Handmade Embroidery and Fast Fashion attribute market failure to exactly this). The direct comparison shows whether selective freedom raises consumer preference without losing artisan approval.
Proposed Experiments and Decisive Checks
On ICHStyleBench (724 images, 10 styles) we report content preservation, style alignment, and CLIP score, to confirm that constraint enforcement does not damage the base generator. We also compute preference-model scores and ImageReward on our garment prompts. The main study uses roughly 2,000 generated garment designs per system. Artisans rate each design approve or reject, and consumer panels in at least one high-familiarity coastal region and one lower-familiarity region make pairwise purchase-preference choices.
The first decisive check is the joint rate per artisan-hour against generate-then-filter at a matched budget. The second is whether artisan-authored maps beat designer-authored maps. If they do not, the governance half of the claim fails. The third is whether violations of fixed elements lower preference more in the high-familiarity region than in the other, which would replicate the Haidilao pattern with controlled stimuli. The fourth, run first because it gates the rest, is whether artisans agree on fixed elements well above chance. If they do not, the idea of a shared community map does not hold for this craft.
References
- A Brief Discussion on the Art Education of the Qing Imperial Family
- The Flow of Institutional Charisma: Quanzhen Taoism and Local Performing Arts in Republic Shandong and Henan
- A Study on the Adaptation of Legend of the White Snake in Chinese Feudal Society
- Challenges and Opportunities of Sichuan Opera Face Changing in Haidilao in Cultural Communication
- Exploration on the Intangible Cultural Heritage Technique of She Ethnic Group in the Teaching of Fashion Design Under the Background of Cross-boundary Integration
- Research on the Path of Revitalizing and Inheriting Rural Intangible Cultural Heritage from the Perspective of Cultural Governance
- Generating Chinese intangible cultural heritage images with structure and color awareness
- A Trend-Driven Fashion Design System for Rapid Response Marketing in E-commerce
- DyMO: Training-Free Diffusion Model Alignment with Dynamic Multi-Objective Scheduling
- Pareto Multi-objective Alignment for Language Models
- DeepFashion: Powering Robust Clothes Recognition and Retrieval with Rich Annotations
- An Evolutionary Game Study of Multi-Agent Collaborative Disaster Relief Mechanisms for Agricultural Natural Disasters in China
- Evolutionary game analysis of FinTech transformation: A social co-governance pattern of peer-to-peer lending market in China
- Visual Transformation of Textile Intangible Cultural Heritage Patterns in the Design of Cultural and Creative Products
- Study on the Influence of Yi Handmade Embroidery and Fast Fashion Clothing Combined Development on the Achievements of Poverty Reduction in Liangshan Area