CADFusion turns text prompts into parametric 3D CAD models

Anyone who has spent hours in CAD software knows how painful parametric 3D modeling can be. Sketching constraints, defining offsets, extruding faces, and fixing broken dependencies require serious engineering expertise. If modern LLMs are already so good at coding, why can't we just describe a mechanical part in plain English and let the model write the CAD instructions?

That is the premise behind Text-to-CAD. Until recently, though, attempts to do this were pretty clumsy. Models were trained strictly on raw command sequences, which meant they often spit out code that compiled fine on paper but rendered into absolute monstrosities.

To bridge this gap, researchers introduced CADFusion, a framework that pairs large language models with a visual feedback loop to generate cleaner, more accurate parametric 3D designs.

Why standard sequence training broke down

A parametric CAD model is essentially a structured recipe of sequential commands. Previous systems took a straightforward approach: grab human-authored parametric sequences and train LLMs to predict the next token, exactly like code generation.

There is a fundamental catch here. CAD is intrinsically multimodal. It is a piece of code, but its real-world identity is the rendered 3D visual object.

Furthermore, the mapping from parametric instructions to a 3D shape is many-to-one. Multiple command sequences can produce the same physical geometry, while a tiny syntax hiccup might keep the code valid but completely warp the physical geometry. When models are trained exclusively on text tokens, they have no clue what their generated models actually look like.

Alternating between code and eyes

CADFusion fixes this by essentially giving the model a visual reality check during training.

The framework uses an LLM as its foundation and alternates continuously between two distinct training phases.

The first is the Sequential Learning (SL) stage. Here, the LLM is trained on ground-truth parametric sequences. This teaches the model the grammar, hierarchy, and logical constraints needed to write valid CAD programs.

The second is the Visual Feedback (VF) stage. The model's generated parameters are rendered into actual 3D visual representations. Sequences that render into clean, visually correct objects receive positive rewards, while those that produce distorted shapes get penalized.

By cycling between these two stages throughout training, CADFusion keeps the model balanced. It doesn't forget its syntactic logic, but it learns how visual renderers actually perceive the geometry. The researchers report substantial gains in both quantitative benchmarks and visual fidelity.

Why it matters

Most generative 3D tools today output chaotic polygon meshes or point clouds. They look decent in a demo video, but real engineers cannot manufacture them, let alone modify a single dimension in SolidWorks or Fusion 360.

CADFusion matters because it outputs parametric CAD sequences. That means the resulting files retain full design trees, constraints, and editable dimensions. If you are a hardware hacker, a 3D printing enthusiast, or a product designer, you get something you can actually open and tweak.

We think the team's alternating training architecture is a clever move. Instead of overcomplicating the model with a giant end-to-end multimodal pipeline, they let the LLM focus on sequence generation while using reinforcement-style visual feedback to steer the outcome.

That said, full-scale industrial CAD is notoriously unforgiving. Real-world machinery requires micron-level tolerances and complex assembly hierarchies that a text prompt cannot fully capture yet. But as a bridge between conversational prompts and real parametric CAD, CADFusion looks like a solid step forward.