When a heritage object is redesigned for modern use, such as a Ming jacket turned into office wear or Yi embroidery placed on a fast-fashion hoodie, some features can change freely and others cannot change without the result no longer reading as that culture. Historians have long described this bargain. Minority courts mastered majority painting to gain legitimacy, folk troupes copied Taoist lineage structures to gain standing, and folk legends were rewritten until they fit orthodoxy. Each case trades away some features in order to keep others. Generative image models make no such trade deliberately. The best automated pipeline on Hanfu-Bench, a benchmark that asks models to turn historical Chinese garments into contemporary wearable designs, succeeds only 42% of the time in human evaluation.
We propose to make the trade explicit. First, we measure, garment component by component, how much each part carries a culture's identity. We do this by editing the part and asking experts whether the garment is still recognizably of its dynasty and type. Second, we generate modern designs with a constraint that preserves the high-identity parts and lets low-identity parts bend to modern structure. The claim is that a small, empirically estimated set of identity weights explains most of the gap between failed and successful transcreation. We also claim that these same parts carry the emotional reading that source-culture viewers attach to the garment, which culture-blind generators lose.
Core Claim
Transcreation is a constrained choice over components, not a style transfer. Let $C$ be a garment's annotated components: the Hanfu-Bench structural tags (sleeve, jin or front panel, collar, bottoms, outerwear) plus motif, color system, and craft technique. Let $w_c$ be the measured identity weight of component $c$. For a modern target structure $s$ (a contemporary garment skeleton), the system chooses for each component a treatment $a_c$ (keep, adapt, or drop) and renders the result $x$ to solve
$$ \max_{a} \sum_{c \in C} w_c \, r_c(x, a_c) \quad \text{subject to} \quad h(x, s) \ge \tau , $$
The symbols are defined as follows. $r_c$ scores how faithfully component $c$ survives in the output, according to a component-level recognition probe. $h(x,s)$ scores conformity to the modern structure, meaning wearability and a silhouette match. $\tau$ is the minimum acceptable modernity. This formalizes the "host conformity versus source preservation" tension that the cultural-capital literature describes only qualitatively.
What is new is the $w_c$. It is a per-culture, per-type identity map that is measured from expert judgments rather than assumed. The planning layer uses it, and it predicts both expert success and preservation of the source culture's emotional reading.
Research Program
The cultural-capital studies play a motivating role, not a methodological one. Pan's account of Qing imperial art education, Qin on folk troupes borrowing Quanzhen Taoist institutional structure, and Tang on the White Snake legend's realignment all show adaptation that keeps a recognizable core while conforming elsewhere. None of them says which features form the core. The heritage-modernization work on Yi embroidery mapped onto fast fashion and San Francisco Chinatown's Minnan ornament on concrete skeletons supplies the operation we automate: keep the technique and replace the silhouette. Xu et al. act as a constraint. They report that naive motif-pasting produces "aesthetic fatigue," so a plan that keeps every motif and simply overlays it is not a solution.
The first phase builds the identity map. Starting from Hanfu-Bench's 496 garment sets with component tags, we create counterfactual edits. Each edit removes or modernizes a single component, either by inpainting with a diffusion model (a generator that learns to turn random noise into images step by step) or by hand from a designer for a calibration subset. Experts and non-experts then answer the benchmark's period and type questions on the edited images. The drop in correct identification gives $w_c$. We fit $w_c$ separately per dynasty and garment type, because we expect the collar to dominate for some periods and the sleeve for others. We also train a vision-language model (an image-and-text model) to predict this drop, so that the map scales beyond the human-labeled subset. Its agreement with held-out expert judgments is reported as a gate. If the prediction is poor, the plan uses only human-estimated weights.
The second phase is the generator. We condition a diffusion model on the modern skeleton through structural guidance, following the ControlNet conditioning used in SCoFT, which fine-tunes against a model's own biased outputs to reduce stereotyping. A planner chooses per-component treatments to maximize the weighted objective, and component masks route each kept or adapted part to its region. This differs from SCoFT, which improves overall cultural accuracy but has no notion of which parts matter or of trading them against a modern target. It also differs from the CycleGAN facade work on Hakka settlements, whose authors note that it treats a facade as a flat image without separating elements. Our method is component-aware by construction.
The third phase brings in regional emotional projection, as an evaluation axis and a hypothesis rather than as a generator input. ArtELingo shows that viewers from different cultures assign different emotions to the same image, for example "awe" versus "fear" for some religious iconography. Emotion-Director improves emotional fidelity in generation but is culture-agnostic. We test whether the emotional response of Chinese viewers to the original garment is preserved in the transcreation. We also test whether that preservation is predicted by the identity-weighted retention score rather than by overall similarity. If both hold, the components that carry identity also carry the regionally specific affect, which links the spatial-affect literature's thesis to a measurable object.
Baselines
The naive reference point is a strong text-to-image model prompted with the modern target and the garment's period and type name ("modern office jacket, Ming dynasty style"). It has no component information and shows what generic style knowledge provides.
The strongest existing method is the best cascaded transcreation pipeline from Hanfu-Bench, which reports a 42% human-evaluated success rate. Its raters and target designs differ from ours, so we rerun that pipeline on our test garments and modern skeletons and have the same raters judge it under the same protocol. Beating it means a statistically reliable gain in expert success rate on identical inputs, with no loss on the wearability score $h$.
A second strong competitor is a frontier vision-language model prompted with the full component annotation and an explicit instruction to preserve each part. Section 2 of the comparison below explains why this is the main threat. It receives the same inputs as our planner, including the identity weights given as text, and the same number of generations per garment.
The key ablations isolate the identity map. The first replaces $w_c$ with uniform weights. The second replaces it with weights from a general VLM's guess of importance, without counterfactual measurement. The third uses a random plan. All three keep the same generator, skeleton conditioning, masks, compute, and sampling budget. If uniform weights perform as well as measured weights, the core claim fails.
Why not just…?
Why not just prompt a frontier multimodal model with a detailed component list? This is the alternative most likely to work, and we treat it as a real threat. The file's evidence is suggestive but indirect. Cultural Bias Matters reports a specialized model (80.16% F1 on English, 77.79% on Chinese) outperforming GPT-4o and Gemini on culturally grounded metaphor understanding, and Hanfu-Bench finds closed models about 10% behind experts at recognizing garments. Neither result concerns generation that is given component-level instructions. The weight-informed prompting baseline settles the question. If it matches our planner, the contribution reduces to the identity map itself, which is still useful, and the generator is unnecessary.
Why not just cut out the original motifs and composite them onto the modern garment? This would score very high on per-component retention, which is exactly why it is a necessary control. Xu et al. argue that this kind of pasting reads as disjointed. We test that claim directly by having raters judge integration and success separately. If compositing wins on success too, the "fusion" premise of the aesthetic-synthesis direction is wrong for this domain, and we would report that.
Why not just fine-tune a LoRA (a small add-on weight module) on Hanfu images and let style transfer handle it? SCoFT reports that plain LoRA on small cultural sets memorizes or collapses. More fundamentally, a style adapter has no mechanism for keeping the collar while dropping the long sleeve. We include a LoRA-only arm with the same data and compute. Our prediction is that it keeps texture and color but fails on component-level retention of high-weight structural parts.
Proposed Experiments
The main experiment uses held-out Hanfu-Bench garment sets, each paired with three modern skeletons (jacket, dress, trousers-with-top). Every method generates the same number of candidates under a fixed sampling budget. Domain experts and non-expert Chinese viewers judge transcreation success with the Hanfu-Bench protocol. We also measure wearability, component retention via the recognition probe, and integration versus pasting.
The first decisive check is the uniform-weight ablation. Measured identity weights must beat uniform weights on expert success at matched modernity. Otherwise the central hypothesis is rejected.
The second decisive check is a frontier plot. We sweep $\tau$ and trace success against wearability for each method. The claim is that the identity-weighted planner dominates the curve, staying recognizable at higher modernity than the baselines.
The third check is affective. Chinese and English viewers assign one of ArtELingo's nine emotion categories to originals and transcreations. We test whether the identity-weighted retention score predicts which transcreations keep the source-culture emotion distribution better than CLIP-style overall similarity does. We also test whether this predictive power is weaker for the English group, which would show that the preserved affect is regionally specific rather than generic.
As a transfer probe, we apply the same counterfactual-edit protocol to the 200-image Hakka facade set. We ask whether a small identity map (roof ridge, door frame, wall material) improves component-aware facade re-skinning over the original CycleGAN results. No quantitative score for those results was found on Lacuna, so we rerun that baseline ourselves.
References
- Hanfu-Bench: A Multimodal Benchmark on Cross-Temporal Cultural Understanding and Transcreation
- 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
- Study on the Influence of Yi Handmade Embroidery and Fast Fashion Clothing Combined Development on the Achievements of Poverty Reduction in Liangshan Area
- Dual Cultural Influence on the Architectural Style Evolution of San Francisco Chinatown: A Comprehensive Examination
- China-chic: From Chinese Elements to International Trend
- SCoFT: Self-Contrastive Fine-Tuning for Equitable Image Generation
- Urban Cultural Inheritance: Generative Adversarial Networks (GANs) Assisted Street Facade Design in Virtual Reality (VR) Environments Based on Hakka Settlements in Hong Kong
- ArtELingo: A Million Emotion Annotations of WikiArt with Emphasis on Diversity over Language and Culture
- Emotion-Director: Bridging Affective Shortcut in Emotion-Oriented Image Generation
- Cultural Bias Matters: A Cross-Cultural Benchmark Dataset and Sentiment-Enriched Model for Understanding Multimodal Metaphors