Why is 3D modeling still so complicated?

01From an idea to a model you can use0:00

Narrator: I'm Khalil, and this is Maya. Suppose you want to design a chair. First, you want to play with the shape. Later, you need dimensions, joints, materials. Then someone asks for a wider version.

Maya: A reasonable request. Four words. Potentially your afternoon.

Narrator: Because the shape, the rules, and the next change don't always fit together neatly. Some tools make starting easy. Others give you powerful control, after you've learned how they think.

Maya: And found the command. And selected the face behind the face you keep selecting.

Narrator: That whole journey is what interests me: getting from an intention to a detailed model you can revise and share. AI could help with all of it. First, let's look at the work we're actually asking it to help with.

02The bargains each CAD tool makes0:45

Maya: Take SketchUp. Draw a face, pull it into a volume. That feels wonderfully immediate. But raw geometry can stick together, so you learn groups, components, and which editing context you're in.

Narrator: It has parametric components too. Someone still has to decide what happens to the shelves and doors when a cabinet gets wider. That behavior doesn't follow automatically from drawing the cabinet.

Maya: Plasticity gives you precise direct modeling without a history tree. Great for exploring a shape. A face edit still needs extra logic if a whole family of parts should follow.

Narrator: Onshape records a sequence of features. Change an early dimension and later features can update. Remove a face they depend on, and they may fail. There are repair tools. There is also something to repair.

Maya: Shapr3D combines direct editing and parametric history. A friendly gesture helps, but you still need to understand the dependencies.

Narrator: Rhino and Grasshopper let you build elaborate geometric procedures. Then you meet data trees: lists with branches that control how things match up. You came to design a facade. Your points have joined the wrong branch.

Maya: An administrative problem for the facade.

Narrator: Revit organizes buildings through families, hosts, and parameters. A change can affect one instance or a whole type. A family can look perfect and expose completely unsuitable controls for your project.

Maya: Each approach earns its place. The hard part is predicting the structure you'll need before the design has settled. I'd like to start simply, then add rules as they become useful.

03What the chat box actually has to solve2:23

Narrator: That's a promising job for AI. Help me express a relationship, explore an alternative, or build a procedure without finding every command myself.

Maya: But we're putting very different things under 'AI for CAD.' Making an image, generating a mesh, writing modeling code, explaining a failed feature. A good result in one doesn't prove the others work.

Narrator: Imagine asking for a housing. The code runs. The solid is valid. The wall is still twice as thick as you wanted.

Maya: Three separate tests. Does it run? Is the geometry usable? Is it the design I asked for? The render passes a fourth test: would I click on this?

Narrator: Autodesk Research tested CAD editing systems and found a substantial gap from professional expectations. Its automatic visual and geometric scores also missed things experts cared about. That describes those tests, but it makes a useful point: looking convincing is an incomplete measure.

Maya: There is progress with solver feedback for generated constraints. You can check whether relationships behave correctly, instead of trusting a plausible picture. Even a fully constrained sketch still needs engineering.

Narrator: And an assistant inside an existing application can be valuable. The real test comes after the first answer: can I point at something, change it, inspect the consequences, and carry on?

Maya: Let's open the generated file, then. This is usually where the demonstration gets interesting.

04What comes out of the generator3:51

Narrator: Mesh generators can make recognizable objects quickly. Meshy's own guide describes guessed hidden geometry, inaccurate details, and separate objects merging. A convincing front view leaves quite a lot happening around the back.

Maya: Even when the shape is close, you may need topology cleanup and dimensional corrections. A background prop and a part that must fit another part have different standards.

Narrator: Generating a boundary representation, or B-rep, can give you native CAD geometry. But the output format cannot tell you whether the design is right.

Maya: The housing looks right. Its wall thickness is wrong. A groove is pretending to be a joint. The holes nearly line up. Nearly is doing a lot of work.

Narrator: The B-rep isn't inherently approximate. The generator has misunderstood the object. Benchmarks also show difficulties with complex features; some still evaluate single bodies without assembly constraints.

Maya: Gaussian splats solve a different problem: reproducing views with lots of soft, overlapping primitives. They can give a captured environment a strong sense of presence.

Narrator: You can select, move, and color-grade splats in tools like SuperSplat. That does not give you a dimensioned wall with editable openings. Captured lighting and surface appearance can also be mixed together.

Maya: Relighting is progressing, including in V-Ray. Still, changing a tint and changing how a material responds to new light are different jobs.

Narrator: Then come the retries. Fix the hole. Check the neighboring features. Explain what must stay fixed. Eventually, editing by hand starts to look relaxing.

Maya: Measure the time to a result you can trust and keep editing. That brings us to what the geometry actually preserves.

05Choose the representation for the next operation5:42

Maya: Say I need to change the radius of a curved part. What information would I want the model to keep?

Narrator: A mesh gives you connected surface faces, often triangles. Useful for rendering, sculpting, scans, and many fabrication jobs. A very detailed mesh still may not contain a named radius you can change.

Maya: More triangles can improve the surface. They cannot tell me why it has that shape.

Narrator: A boundary representation describes the boundary and how its pieces connect. In typical CAD systems, faces and edges refer to mathematical surfaces and curves. That helps preserve a cylindrical hole or a precise mating face.

Maya: With tolerances, and operations that can still fail. Precision does not guarantee a peaceful afternoon.

Narrator: A function representation uses a function over space to distinguish inside, outside, and the boundary. A signed distance field is one kind. Fields are useful for smooth blends and structures like lattices. Displaying or exporting them still takes work.

Maya: So choose for the job. Precise surfaces where fit matters, meshes where they help, fields where the design benefits. Then be explicit about what a conversion loses.

Narrator: And the kernel is another question. Open Cascade, Parasolid, and ACIS are geometry engines. They perform geometric operations; the application still has to decide how you work.

Maya: The engine can calculate the hole. Something has to remember what it lines up with.

Narrator: Which is where a shape starts becoming a design you can change.

06Keep the relationships through the edit7:19

Narrator: Let's use a window frame. You want it wider, while the frame members stay sixty millimeters thick.

Maya: If I stretch the whole thing horizontally, the vertical members get thicker too. The tool did exactly what I told it.

Narrator: If width and thickness are separate parameters, I can change one and preserve the other. Same starting picture. Different behavior.

Maya: Now make it a cabinet. Longer shelves, redistributed doors, hardware a fixed distance from the edges. Without those relationships, I become the update system.

Narrator: Excellent flexibility. Slightly dependent on coffee.

Maya: And I need to know the scope. Am I changing this one cabinet, every instance of it, or the rule that generates them? If I make a local exception, it should survive the next update.

Narrator: Some changes really do break a relationship. A host face disappears. Constraints become impossible. Show me what broke and help me repair it. Quietly attaching a feature to a different face can hide a design error.

Maya: AI could help explain and author those relationships. But its explanation, the controls, and the actual geometry have to agree.

Narrator: And the conversation has to move at a useful pace. Otherwise every small decision becomes a scheduled event.

07The whole cost of trying another idea8:35

Maya: Find the command. Select the part. Change a value. Wait. Inspect. Repair. That's the cost of one idea, even if the generation itself was quick.

Narrator: Complex geometry and relationships can mean real computational work. Revit documents that tradeoff. Sending every adjustment through a fresh AI generation can add another delay.

Maya: Sometimes the computer is fast and the interface is the problem. Which value controls this? Is that preview current? Can I compare two options without losing one?

Narrator: A better loop reuses valid work, updates what changed, gives responsive previews, and makes undo dependable. If an operation fails, the explanation should give you a next step.

Maya: 'Cannot do that because these constraints conflict' is useful. I can fix it. A cheerful wrong answer gives me another inspection job.

Narrator: The article's timing diagram is a schematic, not a benchmark. The point is to count inspection and revision too. Cheaper iteration lets us try more directions. Then we need to decide whether the result works as an actual object.

08Detail has to mean something9:45

Maya: Our chair has beautiful rounded edges now. How are the legs attached?

Narrator: An excellent question for the chair.

Maya: Stock thickness, joints, clearance for moving parts, fasteners that avoid each other. Those decisions make detail useful. Adding more geometry cannot resolve them by itself.

Narrator: Keep the relationships in the model and they can survive a change. Keep them entirely in someone's head, and that person's holiday becomes a dependency.

Maya: There is another distinction here. Parallel faces do not prove a bracket can carry a load. A closed solid does not prove the part can be manufactured by the process you chose.

Narrator: Those checks need appropriate assumptions and evidence. The model should tell us which requirement drives a detail, what has been checked, and what remains unresolved. AI can help manage that information. Its explanation cannot stand in for the check.

Maya: Now we have something another person could use. Provided that information actually reaches them.

09What survives the handoff10:47

Maya: A client wants to review the proposal. A colleague wants to revise the assembly. A fabricator needs dimensions, tolerances, and the correct revision. One attractive viewport is a rather optimistic handoff.

Narrator: There are serious collaboration tools already. Onshape has versions and branches for controlled revisions and alternatives. But crossing between applications raises another question: which design information travels?

Maya: Onshape's import documentation is clear: imported CAD does not bring its original feature tree. You can still work with the solid. You haven't received its original modeling recipe.

Narrator: Useful reuse needs that recipe, or controls that expose the behavior you need. In the article's example, several windows share a width. One gets a local override. It should keep that exception while still benefiting from relevant changes to the shared definition.

Maya: And the drawings and quantities should tell me which revision they describe. Otherwise the model and the documents quietly develop separate opinions.

Narrator: Preserving those decisions is valuable beyond CAD too. It changes what we could ask creative AI to do.

10The scene behind creative AI11:57

Narrator: This is my bigger bet: reliable, editable 3D generation is the most important next step for creative AI. Images, video, games, film, architecture, design. They all benefit when we can decide what exists and keep control of it.

Maya: A scene can keep the objects, dimensions, positions, materials, lights, and cameras together. Make another shot from the same decisions. Change a decision deliberately.

Narrator: Adobe already has Scene to Image, using a constructed 3D scene to guide composition and viewpoint. A rough blockout can establish the layout. An accurate product model matters when the product itself must stay correct.

Maya: Our chair should not have to reapply for its fourth leg in every campaign image.

Narrator: For video, spatial information can help maintain a set as the camera moves. NVIDIA has research using an explicit 3D cache for that. It is evidence of a useful direction, with limits.

Maya: In games, a door also needs motion and collision geometry. In architecture, the plan, material study, and walkthrough should describe the same design. These are different outputs with shared decisions underneath.

Narrator: Film and game production already work around shared scenes. Open U S D helps compose and exchange scene information. AI could make creating and revising those scenes accessible to many more people.

Narrator: But for repeatable spatial control, I think the editable scene becomes the most valuable thing we keep. It can support another shot, another material, another medium. That's why making it easier to model matters so much.

Maya: Not every image needs a 3D model. And using one as a reference doesn't guarantee the generator obeys it. You still check the result.

Narrator: Making the scene approachable gives more people a way to direct that result.

11Let the workspace grow with the design13:46

Maya: Then the entry point has to be approachable. Let me explore a form before asking me to plan its entire future.

Narrator: Add dimensions when they matter. Add relationships when repetition appears. Develop the parts as the idea becomes clearer. The model should get more capable as you need more from it.

Maya: Sometimes I want to drag a part. Sometimes I want a hundred parts to follow a rule. Sometimes I'd like help figuring out the rule.

Narrator: Words, sketches, selections, and dimensions can all contribute. AI can propose an edit, help develop a detail, or explain a dependency. I should be able to inspect the proposal and continue working on the same model.

Maya: So the demonstration needs a second act. Make the model, then change your mind.

Narrator: Then hand it to someone else and see what they can do with it. That is a demanding test for any new tool.

Maya: I assume you haven't brought all of this up purely to improve our imaginary chair.

Narrator: The chair has been very helpful.

12Why we are building Eskiz14:47

Narrator: We're building Eskiz. A new approach to 3D modeling, enhanced by AI collaboration.

Maya: From an intention to a highly detailed model, with room to explore and change it along the way.

Narrator: That's the ambition. Direct modeling, reusable logic, and AI assistance working together. Easier iteration. Changes that propagate through the relationships they should affect.

Maya: And you'll show how that works through actual design work?

Narrator: Yes. There's a lot underneath it, including the work we've been doing on procedural design and editable modeling systems. We'll share the machinery as it becomes useful to explain the work.

Maya: What should people watch for next?

Narrator: We'll share details and progress weekly as we work toward the reveal. You can join the Eskiz waitlist through the link in the article.

Maya: I'd like to spend more of the afternoon trying designs.

Narrator: So would I. More of the work between an idea and a finished design should be design work. That's what we're building toward.

13Until next time15:42

Narrator: Thanks for listening. Follow along for more articles, experiments, and things that worked on my machine.

Maya: Some of them might even work on yours.

Narrator: No promises. Have a lovely day. Go make something.

Maya: Preferably a cup of tea first. Bye!