RESEARCH · 03 HERITAGE LOGIC

Latent Heritage:
Industrial Muse

An instrumented AI pipeline for transferring Chinese craft logic into new design directions.

StatusOngoing international collaboration
InstitutionsDonghua University × HAW Hamburg
My roleLoRA training · ComfyUI workflows · Generation runs · Method development
CollaboratorsProf. Peter Kabel — HAW Hamburg
Aimee Zhang — Donghua University
Industrial Muse generated design informed by craft logic
Generated design direction.
Traditional bamboo craft source object
Craft source.

Research thesis

Transfer functional logic—not decorative motif.

Research question

What new design possibilities emerge when AI learns how Chinese crafts work—not merely how they look?

Latent Heritage studies around thirty Chinese crafts—from bamboo weaving to Miao silverwork—as operational grammars: folding, weaving, locking and suspension. These principles enter an AI image system through four controlled channels—material, structure, form and language—to produce unfamiliar designs without decorative clichés. Every source image, dataset and generation run remains traceable, grounding experimentation in provenance and respect for craft communities.

01 · Knowledge foundation

A governed corpus turns cultural provenance into part of the research method.

2,170+corpus images ingested
30crafts catalogued by functional logic
8+LoRA training runs across four craft instruments
24versioned grammar sketch sheets
Team-level evidence.

Each image carries source, licence and provenance records. Each new craft follows the same route: corpus → curated set → LoRA + grammar maps → registered runs.

02 · Generation architecture

Four controlled channels assign different responsibilities inside one generation run.

01 · Material · Live instrumentMiao batik material channel reference
Trained LoRA carries material signature, relief and grounding.
02 · Structure · First runOpen weaving grammar sketch for structure channel
Grammar sketch + ControlNet carries construction logic.
03 · Form · In buildTarget silhouette for form channel
Target silhouette or depth constrains object form.
04 · Prompt · Live input{ material, structure, form, object, context, intent }

A structured concept brief binds the roles without pretending that text alone is the pipeline.

03 · Controlled evidence

Same seed, same brief—the instruments reveal what actually changes.

Vehicle generation with LoRA switched off
LoRA off · a lattice sits on the panel.
Vehicle generation with bamboo LoRA switched on
LoRA on · lashed joints and load paths shape the object.

The comparison asks whether the model merely applies a surface pattern or begins to reorganise the object through a craft-derived logic.

Seed-locked ablations, strength ladders and material swaps keep the change attributable and inspectable.

04 · First end-to-end run

A first test of material, structure and prompt together—not yet a proof.

Dense vertical weaving structure map
Flat Herringbone weave grammar.
Open vertical weaving structure map
Pre-projected Herringbone weave grammar.
Radial weaving structure map
Pre-projected Radial grammar.

05 · Traceability infrastructure

Every source, dataset and generation run remains inspectable.

Latent Heritage generation system and visual generation register
The research system connects four input channels, generation, provenance-aware storage, internal review and a controlled external layer.
Internal Latent Heritage source database and dataset export interface
Source database. Records provenance and usage permission, generates training annotations and exports datasets. Internal access while the research is ongoing.
Internal experiment gallery showing parameters, comparison matrices and generation records
Experiment register. Connects parameters, inputs, outputs and findings for every formal run. Internal access while the research is ongoing.
Selection of Latent Heritage experiments across garments, lighting, furniture and spatial design
A curated cross-domain field of current experiments; selected here as evidence of transfer rather than as a complete gallery.

Research boundary and next phase

Novelty remains human, governed and culturally situated.

The project is non-commercial research. Publication-grade or commercial reuse requires renewed licence review, and minority-craft work requires cultural-advisory participation rather than technical validation alone.

Next: complete structure and form channels, conduct Chinese–Western reception research, develop pairwise human curation and formalise cultural-advisory review.