How design semantics enter a generative model

01Define what the model must understand

Understanding industrial design goes beyond recognizing a vehicle or a machine. Designers work with users, assembly relationships, materials, manufacturing processes, service access and brand language. Compressing these into a style caption risks teaching appearance without the conditions that make a design useful.

For HuanYu domain research, the design task is the organizing unit: requirements, references, proposals and review feedback form linked records. Views and revisions of one product belong to a sample family. This preserves the design process and creates a defensible boundary between training and evaluation sets.

• Object: product category, intended users and context.
• Constraints: structure, materials, processes and operating boundaries.
• Expression: viewpoint, composition, visual language and design stage.

02Preserve the design process in the dataset

Data preparation begins with provenance, rights and revisions before visual quality. Near-duplicate detection must consider crops, altered backgrounds and multiple views, not only file hashes. Splitting by design task helps prevent a proposal from appearing in training and returning in a test set under a different view.

Annotations should separate observable properties from expert judgments. Tools can assist with colors and component positions; maintainability or process suitability needs supporting evidence. Keeping an attribute unknown is more useful than filling a complete-looking record with an unsupported claim.

03Separate the backbone from domain conditioning

DiT explores a generative backbone that processes latent image patches with a Transformer. Flow Matching learns vector fields connecting noise and data distributions. Network representation and the training objective address different parts of the problem; neither label alone establishes industrial design capability.

Our research design organizes domain conditions explicitly: text expresses intent, structured fields carry constraints, and references describe an existing form. Encoded conditions guide generation while retaining whether each condition was provided or inferred. An iteration can then change a chosen condition without silently rewriting the whole brief.

Domain continuation, conditioning adapters and parameter-efficient fine-tuning are engineering options to compare. Their effects on constraint adherence and distribution shift matter more than dismissing a method category. Ablations must determine the appropriate training combination.

04Evaluate the image and the conditions

Evaluation should retain separate outcomes instead of a single aesthetic score. Check required objects, components and relationships; then identity across views or revisions; then ask domain reviewers whether the concept merits further development. Model-based ratings are supporting evidence, not the only judge.

A shared task set should cover ordinary briefs, conflicting constraints and unfamiliar products with fixed references and rubrics. Comparisons between general models, domain adaptation and domain training should record inference cost, failure categories and human revision effort, rather than only selected images.

05Hand the result to the next design stage

A rendering can express structural intent without being a verified engineering drawing. Production still requires dimensions, materials, assembly, tolerances and process validation. The research goal is a result with enough explicit intent and constraints to become a useful starting point for 3D modeling and engineering judgment.

A generation artifact can therefore include the image, task conditions, reference provenance and revision links. Retaining this context alongside an API artifact makes changes traceable and allows designers to continue the work. This note describes a research and artifact design; integration documents define the actual interface.