Start with design practice
How form serves use, how parts make a whole and how materials respond to context: IDENIFE brings these industrial design judgements into HuanYu’s development.
HuanYuby IDENIFEIndustrial Design Generation API
From design intent.
To product form.
Design intelligence. A new expression of imagination.
Grounded in industrial design data. Turn conversations, product semantics and visual conditions into concepts—through an API inside your platform.
Explore the technologyAutomatic animation: large mosaic tiles progressively subdivide to reveal the silhouette, surfaces, materials and lighting of a sharp car concept. This visualises generation using a concept asset; it is not a live API response.
“Design an electric GT.”
Low stance. Liquid silver. A slender light signature.From pixels to design.
Illustration of the generation processBorn from industrial design.
First, understand design.
Then, teach it to a model.
Behind a rendering are decisions about use, form, materials and manufacturing. IDENIFE’s path from industrial design to model development makes those decisions the starting point for HuanYu.
Turn experience into data
Connect briefs, renderings, part descriptions and revision feedback. The learning task includes the design problem an image responds to, alongside the image itself.
Put generation to work
Connect through an API to design tools, enterprise platforms and industry software, keeping briefs, concepts, reviews and revisions in a shared business context.
Industrial data & semantics
More than an image.
A design with a reason.
HuanYu is trained on industrial design data. Relationships among briefs, form, parts and materials retain the professional context of each visual sample, beyond a style label.
Design brief
A low-slung GT. Long wheelbase, a compact cabin, silver metallic bodywork and slim lighting.
Form & proportionParts & layoutColour & material
Form
Stance · proportion
Parts
Lighting · layout
CMF
Colour · material · finish
Data dimensionConnected informationWhat it teaches
Briefs & concepts
Briefs, use contexts, concept images
Connect the reason for a design decision with its visual expression.
Parts & form
Silhouettes, proportions, zones, part relationships
Relate the overall form and its local features to a coherent product.
Materials & finish
CMF descriptions, seams, surfaces, connections
Distinguish material appearance from structural intent, rather than learning style alone.
Revisions & review
Source concept, edit instruction, preserved regions, review
Extend single-image generation into revisions that retain the design direction.
Evaluate across product families and projects.
Verify rights and provenance, deduplicate and retain version relationships. Similar views of one product are not independent evidence of generalisation.The generation architecture
Design conditions.
Present at every step.
From intent and multimodal conditions to joint latent modelling and image decoding: professional generation connects data, the model backbone, control mechanisms and evaluation.
01
Encode conditions
Design brief
Text embeddings
Reference / sketch
Visual & spatial context
Language defines intent. Visual conditions guide its arrangement.
02
Multimodal modelling
Multimodal TransformerText & visual conditioning
Noisy latents + timestep
Image and text representations interact, with spatial conditions guiding each step.
03
Iterative integration
Velocity integration
NoiseFormDetail
Repeatedly predict velocity with the backbone; a solver advances the latents towards the concept.
04
Image decoding
Visual decoder
Concept imageDecode latents into pixels for design review and the next revision.
Design conditions stay in the loop.
Illustrative technical pathway01Multimodal Transformer
Model language and vision together
Attention exchanges information between text representations and image latents. Category, form, material and spatial relationships jointly condition a concept, rather than acting as isolated words.
02Flow Matching
Form emerges in latent space
Flow Matching trains a velocity field. During inference, the transformer predicts velocity from the current latents, timestep and design conditions. A numerical solver advances the latents, before a decoder produces the image.
03Spatial Conditioning
Make spatial intent a condition
References, contours, depth and region masks provide different constraints. Condition strength and creative freedom are balanced around established proportions, layout and identity.
04Representation Alignment
Guide learning with semantic features
During training, intermediate generative features can be aligned with pretrained visual representations to improve semantic learning. This training mechanism is distinct from inference-time conditioning.
These diagrams explain the development approach. Public research is linked below; model configuration and supported controls are defined for each integration.
Generation is the beginning
Open up the possibilities.
Keep the design direction.
Industrial design needs deliberate variation. Organise conditions around form, CMF and local revisions so every generation addresses a more specific design question.
Establish the design language.
Anchor exploration in product category, silhouette and key part locations. Retain the low stance and long proportions while exploring surface treatment and detail.
Preserve: proportions, cabin placement, silhouette
Make material choices visible.
Describe colour, material and finish separately. Explore how they affect highlights, surfaces and character, giving CMF reviews a common visual language.
Explore: liquid silver, dark glass, fine metallic finish
Change a detail. Keep the direction.
Condition an edit on the source, the current instruction and the target region. Separate what changes from what stays, then assess both the edit and identity preservation.
Review: edit accuracy, untouched regions, product identity
Images illustrate control dimensions; they are not a comparison of separate generation or editing results.
One revision. Three conditions.
“Keep the proportions and cabin. Simplify the front lighting.”
- Preserve
- Silhouette, cabin and established materials
- Change
- Lighting region, signature and local detail
- Validate
- Edit accuracy and preservation of untouched regions
Domain learning & design alignment
Specialisation.
Built as a complete system.
The difference begins with how the learning task is defined.
LoRA is an efficient parameter adaptation method. Domain capability also depends on data relationships, generation and editing tasks, conditioning and evaluation. HuanYu’s development is organised around that complete design process.
01
Build visual-language foundations
Learn relationships between language and visual structure, establishing object, spatial, form and material representations.
02
Learn industrial design tasks
Organise domain images, text, part and CMF semantics by task, strengthening the relationship between brief and concept.
03
Connect generation and editing
Connect initial generation with source-conditioned revisions, focusing on context and identity across turns.
04
Learn from design judgement
Compare candidates for the same brief. Separate aesthetic preference from constraint compliance and use difficult cases to guide refinement.
Connect data versions, task definitions and evaluation results. Give the next iteration evidence beyond visual preference.
Quality for design work
A good image
moves the work forward.
Assess image quality, condition adherence and design usefulness separately. The goal is a concept that supports informed review and further development.
Are the conditions respected?
Check category, critical parts, layout and CMF; record omissions and conflicts.
Do the parts make sense?
Review part completeness, connections and local contradictions, separately from engineering validation.
Does an edit preserve identity?
Evaluate changed and preserved regions together for unintended changes or drift.
Does it advance the next design step?
Designers assess usefulness for selection, communication and subsequent modelling.
Move ideas towards manufacturing.
Concept3D developmentEngineering validationManufacturing
Generated images inform design decisions. Dimensions, tolerances, structure, materials and processes are validated during engineering.
An API inside your product
Your design platform.
HuanYu intelligence.
Keep your accounts, projects and workflows. Integrate specialist generation into existing software and build the experience around your product.
Define the design input
Brief · references · preserve / change conditions
Your system owns project context, permissions and source provenance, with a clear input version for each generation.
Manage generation tasks
Task identity · state · errors · version linkage
Fit generation into an asynchronous workflow with agreed polling or callbacks, error handling and duplicate-request behaviour.
Use the design artifacts
Image assets · specifications · conditions · review history
Route artifacts into your preview, selection, review and 3D development workflow so each result can be referenced, compared and revised.
Technical references
Original work on model architectures, conditioning and training methods.
DiTTransformer backboneFlow MatchingLearning generative pathsMMDiTJoint text-image modellingControlNetSpatial conditioningREPARepresentation alignment
These papers provide technical background; they do not establish HuanYu’s implementation or performance.