Vibe coding and 3D design: the next evolution of modelling

An interior designer is dressing the show flat for a new residential development. They need forty props: chairs, lamps, vases, books, textiles. Three style options. Until recently, every piece meant searching libraries, adapting someone else's models or building it vertex by vertex.

Today they can describe an armchair and receive four versions at different mesh densities. They can ask an agent to inspect the scene and tell them which objects are too heavy. It is the same shift vibe coding brought to software, applied to geometry.

3D modelling is moving from pushing vertices to directing outcomes. What does not change is who is responsible for the measurements, the scale and the intent behind every piece.

Four ways of modelling that now coexist

Modelling has not evolved by replacement. It has built up in layers:

• Manual. Extruding, cutting and adjusting vertices. Still the most precise way to control a piece.
• Procedural. Modifiers and node systems that generate geometry from rules. Blender 5.2 LTS, released on 14 July 2026, added a Mesh Bevel node with detailed procedural control over edges and vertices.
• Generative. Text or images turned into a 3D mesh.
• Conversational. An AI agent that operates the modelling software from written instructions.

A studio with sound judgement does not pick just one. It decides which layer suits each part of the project.

What is new: generated meshes you can actually edit

On 21 September 2026, Tripo AI launched Tripo P2.0, a model that generates meshes with native quad topology. The company describes it as a first for AI 3D generation.

It sounds like a technical footnote, but it was the bottleneck between a curiosity and a production tool. Generated meshes tended to be dense, irregular triangle soup. Subdividing, texturing or animating them properly meant rebuilding their topology by hand, and that work swallowed much of the time saved.

The press release also lists:

• Triangle meshes of up to 50,000 faces and quad meshes of up to 25,000.
• Up to four versions of the same asset, each with a different face count, from a single prompt.
• Mesh Edit, which regenerates one region of a model without altering the rest.

These are the manufacturer's claims. The test that matters is different: open the asset in your own pipeline and check whether it holds up in a close-up, a material change and a scale review.

When the agent drives Blender

The other half of the shift does not generate objects; it operates the software. On 28 April 2026, Anthropic announced Claude connectors for creative tools, including Blender, Autodesk Fusion and SketchUp. In Fusion, subscribers can create and modify 3D models through conversation; in SketchUp, a conversation becomes the starting point for a model.

Blender's approach is telling. Its official MCP server, developed at Blender Lab, is presented as a natural-language interface to the software's Python API. Its examples are not "build me a building". They are analysing scene performance, finding which objects use a material, fixing data-block names and documenting a node setup. It requires Blender 5.1 or later.

This is where vibe coding and modelling meet. The agent writes code that changes the scene, and you judge the result in the viewport rather than the editor. It is the same logic we described when explaining how a game was built by conducting an orchestra of agents, with one difference: in architecture, a scale error does not show up in a playthrough. It shows up in a home.

What to delegate, and what not, on an architecture project

The useful question is not whether AI can model. It is which pieces tolerate an approximation and which demand accuracy.

Worth delegating:

• Secondary dressing: decorative objects, books, tableware or background planting that nobody will sell or build.
• Style variations to compare options before investing in the final one.
• Scene housekeeping: renaming, organising collections, assigning materials in bulk or finding objects with too many polygons.
• Early massing to discuss with the client, always flagged as provisional.

Best kept out of AI's hands:

• The project geometry. Walls, openings, slabs and joinery come from drawings or BIM models with verified dimensions.
• Real products. If the buyer will receive a specific sofa or tap, the model must be the manufacturer's own or a faithful reproduction.
• Materials in the specification. A similar floor is not the floor in the contract.
• Scale and ergonomics. A generated chair can look right and still be ten centimetres too tall.

A hybrid workflow we would propose

Picture the show flat from the opening, intended for stills and for an interactive model in the browser. A sensible division of labour could be:

1. Build the architecture from the drawings, with checked dimensions and no automatic generation.
2. Sort the pieces into three groups: real products, visible dressing and background dressing.
3. Generate only the dressing, requesting several mesh densities: more detail for stills, less for the real-time experience, where every polygon counts.
4. Review every piece at real scale beside a human reference, checking topology, UVs and materials.
5. Hand scene housekeeping to the agent on a working copy: naming, collections and a polygon report.
6. Keep art direction with people: light, composition and consistency across every piece.

Time saved on secondary objects should go where the client notices it: lighting, materials and the visual story. If it only produces more of the same images, the shift adds little.

Risks to manage from day one

• Security. Blender warns that its MCP server runs model-generated code without safeguards against data being deleted or sent elsewhere, and recommends a virtual machine or a system without sensitive information. For a studio handling confidential projects, that warning defines the working environment.
• Licensing. Every generation service has its own terms for commercial use. Check them before delivering an asset to a client.
• Consistency. Forty objects generated separately may not share a style, level of detail or material treatment. Pulling them together is still design work.
• Traceability. Keep a record of which pieces were generated with AI, with which tool and how they were reviewed. It makes for a clear answer if a client or buyer asks.

The evolution of modelling does not remove the craft. It shifts it towards direction, verification and judgement.

If you are preparing a development, a showroom or a 3D experience and want to know which parts of the modelling can be accelerated without losing precision, tell us about your project. At Viseni, we can help you draw the line between what is generated and what is modelled exactly.