Spectral Labs releases SGS-1: the first generative model for structured CAD | Hacker News Reader

Spectral Labs releases SGS-1: the first generative model for structured CAD

spectrallabs.ai

319 pointsJumpCrisscross65 comments

> can generate fully [...] parametric 3D geometry.

This is obviously a lie without even testing it. STEP files do not have any sort of support for parametric features.

> SGS-1 outputs are accurate and can be edited easily in traditional CAD software.

I tested this on their own demo file to give it its best chance at working correctly and this claim they've made is just a complete lie. I've included the input and output for comparison below next to a correct part I modelled myself[1][2] and a list of errors. These aren't just incorrect dimensions but broken features that make these very hard to edit.

I don't know why they'd lie about this and them provide a demo the shows they're lying with their own input files, and even text on their own page which makes a claim that can not be true based solely on the information further down that page. Is it to get news headlines? Do they want to sell this to people that don't know any better? Is this simply just another case of CS people thinking they've solved a problem without having any domain knowledge to know their claims are nonsense?

- Every dimension is wrong aside from the one that I corrected to get the same scale (there doesn't appear to be any correct relative to each-other which is why I just picked one at random)

- One hole doesn't go all the way through

- The closest hole isn't round (it's two holes with slightly different diameters that overlap with a sum larger than either hole)

- The fillets are not fillets

- The top hole is offset

- The front chamfer goes down past the base

- The not-fillets don't have the same radius

- The two top holes are offset from each other in Z

- The front chamfer is joined to the circle in different ways on each side (to be fair the drawing is nonsensical here, I just went for a tangent with the circle on my part)

- Much more that I've probably missed

[1]: https://files.catbox.moe/mzb9bb.png

[2]: https://files.catbox.moe/5xkna1.png

> This is obviously a lie without even testing it. STEP files do not have any sort of support for parametric features.

Yes, this is also confusing me to no end. How can they make such a claim? They even explicitly state that they generate a B-rep (boundary representation) output only, then in their roller example they say "as the output is parametric, dimensions can easily be adjusted." Erm, no? I'd rather model it again with the proper feature history tree and constraints instead of fiddling with a step file.

I think the point they are making us that the intermediate representation the model works with is parametric and then converted to step for use with other tooling, I could be wrong, but I understand the argument both ways of their solution enables editing of that parametric version before conversion out.

You seem to know what you're talking about. I am completely ignorant of this field, but I have been hearing about https://zoo.dev/design-studio (nee. "kitty cad") for a while because I follow one of the lead devs and it seems like it offers similar features.

I would be very interested to get a comparison from someone who understands the terrain.

Pure text to CAD doesn't seem particularly useful because of how hard it would be to express all constraints in text. Perhaps it's is useful for some kind of mock-up work but that's not really my area.

I agree, text-to-mesh makes much more sense for 3D artist than for CAD. If I'd wanted to have all my parts in text, I'd just use OpenSCAD, writing this down in prose sounds horrible.

And lo, OpenSCAD is great, but is basically unused in the professional world... mainly because textual representation of models amounting to anything more complex than basic 3d shapes slapped together is difficult to achieve. You just can't do so many things in OpenSCAD that you need once you start doing professional designn

I really appreciate OpenSCAD. It’s my go-to whenever I need to quickly define a basic shape to 3D print. Unfortunately I predictably hit a wall every time the complexity increases beyond the basics.

The kernel is not robust to the stupid things I naively ask it to do. Often my code makes sense to me, but OpenSCAD refuses to create an object.

Performance falls off a cliff. You can work around it by pausing previews and adjusting resolution, but that’s a big UX compromise.

Still, I’ve tried a few other options and keep cycling back to OpenSCAD. The barrier to entry is very low, coding AI does a pretty good job, and there’s a decent ecosystem of community modules.

I also religiously use OpenSCAD for hobbyist stuff. For making some simple component/replacement stand or enclosure in my house, it's fantastic, don't need the full on parametric modeller for that, plus it's way more modular and extensible with code

Great until you want to fillet or chamfer a complex edge, which professionals do all day long

Though it's possible that generating the STEP first is easier to do, and that the plan could be backporting the feature tree using another method / model would then enable editing.

Yes, it would seem post hoc feature tree requires the constraints that come from context in your head, but I could imagine that for most cases a "drafter's intuition" in AI may be sufficient, and you could build an interface to allow that to be mostly given up front and then through iterate post generation.

I could imagine the stepwise approach may allow AI training to be more constrained / efficient that trying to do the whole thing in one go.

Right. If you can express the constraints in text, why wouldn't you just draw it in a CAD program?

This is also why SQL is cursed for LLMs. For queries that are actually valuable to the business we tend to have more constraints than these models can tolerate.

By the time you get done explaining the meaning of your schema, you might have run out of context. Not that it would matter either way. I've never seen the attention mechanism lock onto more than ~10 hard constraints at a time.

Maybe you need to approach SQL the way code generation must be approached: don’t develop the whole statement or script at once, instead put together a plan and execute it step-by-step.

Not everything must be done via LLMs themselves. You could use one or more tools to help generate parts of the query.

You might be interested in this:

https://www.pedronasc.com/articles/lessons-building-ai-data-...

Yes, it is a bunch of lies. I wonder what who they pretend to deceive? May be investors?

I’m not deeply familiar with these formats, but isn’t a defining feature of STEP files that they store “B-rep” (boundary representation) data, unlike eg the polygons in an STL file, which are the geometric relationships between surfaces and can be easily used by a parametric editor? That seems to be the point they are making, vs other 3D generation models that output polygons.

STEP files do store higher level geometry data than STLs, making it possible to accurately modify the imported data in parametric modeling software like Fusion, so in that sense they’re “parametric compatible”? They don’t store the design history or intermediate setup, though, so you can’t “go back in time” and change a parameter that was set during the design of the part, and have everything recalculate to accommodate this.

I agree that STEP files are not parametric. But for cases where thats all you've got, Altair Inspire is pretty good at letting you use them in a CAD system:

https://altair.com/inspire

It identifies features as such even from step.

Yes, but is it even saving you any time, then? For complicated parts, maybe, but the example above is not complicated and the AI is already doing a pretty terrible job. Someone familiar with CAD is able to quickly do this in 15min or so, and then you have all the history and constraints to adapt the design simply by just changing parameters. I'm pretty sure I'd spent much more time fiddling with a rough approximation in the form of a step file.

Even simpler tools like Solid Edge in synchronous mode will do a decent job of editing STEP files, as long as they're not just full of polygons.

AI news lying about their real capabilities?

I bet somebody in the comments will also say it works for them and then provide vague or not details at all.

For parametric what the mean is that it generates fusion360 solids that fit into a parametric workflow (unlike say a mesh)

Remember Adept?

I think there is a market for AI that would clean up the goo imports of 3D scanners. Imagine how beneficial that would be.

Also to generate clean 3D meshes from points cloud, while identifying the various objects via the colours/lighting. That would be also really interesting. It could as well describe the world, and the object's meta data.

As for something like this, it removes the fun of CAD design more than solve a problem, I think we best focus AI for repetitive boring tasks, rather than design. This may wow investors, and may save professional minutes, but it does not really solve the bigger problems.

Nobody in this space cares about anything except wowing the investors. They would willingly sap every ounce of joy out of every single knowledge worker’s professional lives and leave us being either AI process overseers or scrubbing toilets if some MBA could envision using that technology to lay people off. From where I’m standing, it looks like it’s just going to kill the demand for labor in a bunch of fields, tanking previously reliable salaries, and transfer the savings directly to shareholders.

The "subconscious" part of the brain processes huge amounts of information, and a lot of thought is actually subconscious manipulation. Conscious attentive deliberate thought is much slower and typically estimated at ~10 bits per second (if that is even the right unit!!!).

When I was nude drawing, the first few poses were always very short (half a minute or a minute) and towards the end of a session the poses took longer. Forcing the brain to make quick decisions (and yes, err along the way) is a fantastic method to force the brain to learn, the sense of urgency and the sense of importance are very much related.

Reflexive / reaction speed computer games force a player to learn.

I believe its possible to upload neural network weights to the human brain by reaction speed games.

I agree with your assessment of the driving force behind machine learning (laying off workers), but I believe it will usher in a new Enlightenment era, where the tremendous energy intensive computations to summarize human knowledge into a compressed form of neural weights results in the democratization of all this knowledge (and if those MBA's had this foresight, they wouldn't share those weights at all! unless they secretly "fight babylon from the inside out").

I will soon try this on a smaller model (that has basic ~100-language knowledge).

The main issues are transforming the model weights so that all weights are embedding weights (moving the attention and feed-forward weights to token weights), but this requires knowledge distillation, and I know what form I want, but not sure if I have the requisite compute to do it.

The second issue is figuring out how many weights per day one can learn.

Not everyone thinks CAD is fun. When prototyping things for 3D printing, I just want a “mount for this controller board with Xmm spacing and N holes”; I would be completely happy using AI for most of it.

“Connect these two parts”, “add screw holes here”, “make a snap fit joint”, “make it 8cm wide”, “move these holes to the other side” is what I dream of!

I played around with a trial of the generative design stuff in Fusion, and my biggest takeaway was that it’s at least as much work to set up all your loads and attachment points and keep-outs and so forth as it is to just model the part myself. Similar to vibe coding in a language I know, really.

I think the demo is cool, but the single most important question is how good it is at converting written constraints into the design.

Designing something that has the right features is easy, the hard part is creating a design that is manufacturable, fits inside the allocated space, has the desired mechanical properties (e.g. rigidity) at the desired cost (uses less material, is easy to manufacture).

As an example, parts for 3D printing have various design constraints, as you want to avoid support material and want to print the part in a certain orientation. Being able to verbally tell an AI model to incorporate these constraints would be very beneficial.

I wonder where they'd get the training data for something like that.

Neat to see the parent article. In the early months of ChatGPT4, I had it produce an open markup standard to support LLM-mediated mechatronic design. Basically I felt that human-readable markup language was needed for an LLN to train on many examples (or for an imagined highly capable LLM to absorb the logic of the markup language de novo), so that it could be used to rapidly iterate subsequent mechatronic designs posed by users. Mostly as a demonstration-of-concept so others could build on it or be inspired to roll their own.

https://github.com/ricksher/ASimpleMechatronicMarkupLanguage

In any case, hopefully it anticipated some future company’s would-be IP thereby ensuring Freedom-to-Operate for future open source hardware tinkerers.

> Given an image or a 3D mesh, SGS-1 can generate CAD B-Rep parts in STEP format. Unlike all other existing generative models, SGS-1 outputs are accurate and can be edited easily in traditional CAD software

This is a game changer, all the models before that output meshes were a toy at best. Super excited to see where they can take this.

I wander if the next step is for a step -> proprietary format (SolidWork, NX etc) model that can infer constraints.

I agree, even if it just does a decent job of creating sane STEP geometry out of 3D scan meshes it would be a huge win.

I’d pay a subscription for that.

There are so many hobbyist 3D printing things I’d like to do around my abode by taking some existing piece and tweaking it. Creating a model for a one off part is pretty tedious though.

I think the bar is even lower than that: generating sane STEP geometry from STL files generated by other CAD software is already a huge win. Autodesk Fusion pretends to be able to do that, but it only works for easy demos.

Thanks for this! We are actively considering this in our next model. What did you have in mind specifically?

As an engineer I just wanted to pipe in and say that this AI removes the part of design that isn’t actually the difficult part.

The difficult part is the manufacturing, also coming up with the plausible load paths and deciding on the geometry of the parts according to the actual loads.

As an engineer, I can confirm that this AI (if it works well) would compress the prototyping stage significantly, resulting in a better product at reduced cost.

> also coming up with the plausible load paths and deciding on the geometry of the parts according to the actual loads

It would help with that as well.

> It would help with that as well.

Beyond having a cad file where there wasn’t one… how?

But can you really confirm it?

Have you even used it and know if it does what it claims?

Sorry, you can't confirm shit.

I’ll bet they’re not a mechanical/manufacturing/etc. engineer.

Remove the boilerplate and focus on the value adding parts is the philosophy that actually happens in software development ai usage, so it might be useful elsewhere as well.

This is being AI’d as well: https://limitlesscnc.ai/

as a founder working on a direct competitor to this product, I am happy to see there's so many opinions and interest. this means we are on to something. I am totally behind the idea, so I am super excited for Pranav to get this in the wild.

however, the output of the current SGS-1 model leaves a lot to be desired in terms of usability and manufacturability.

as sir liam powell so explained, the STEP generated is indeed a STEP file in the fact that it will load on your CAD of choice (CATIA, NX) but it's not even close to being useful for the designer/engineer/fabricator much less a factory or even a 3d printer.

Most importantly: -Surfaces are not G1, G2, or even G0 (un-CAE-able and un-manufacturable) -simple shapes are hyper overcomplicated (a fillet turns into 10++ surfaces mashed together) -and most unfortunately in one part I generated, the output was a collection of separate STEP child parts (meaning, when importing, your single "part" is actually a product containing 10000+ child parts)

looking forward to V2, keep it coming

shameless plug: my entire career has been in CAD and that is why I started Transfigure. to eliminate my own job as a CAD monkey: and get on with the engineering. already having anticipated the limitations which SGS-1 is dealing with, Transfigure's AI architecture is approaching the problem from the perspective of a mechanical engineer whose customers rely on clean data, ready for simulation and manufacturing. because if I submitted data generated by SGS-1 to our fab shop... I would be fired.

yo@xfgr.ai

sounds interesting when are you planning on making more info publicly available?

not a great answer but the true answer would be only God knows. Personally my goal is 1 month.

posting random "artistic" snapshots of various aspects of Transfigure (instagram: itstransfigure) if you care to follow.

without saying too much, the competitive advantage is in training data and I think that will be the deciding factor as to whether or not a company is successful in (Transfiguring? rezzing? summoning?) a real part from a sketch, that can be made into physical stuff, without looking like shapenet cartoons.

I tried one of the samples and it wasn't watertight, with spurious little holes in the corners. Still mind-boggling though, and that can be healed easily enough.

Depending on pricing, this could be a boon for 3D printing. The really hard part for hobbyists creating novel models has been the CAD. If they can price this appropriately I can see it selling really well.

This would be incredible if it shipped with bambustudio or orca slicer. This is the first AI release in months that I've been excited for.

Related: All LLMs have struggled helping me generate OpenSCAD models for 3D printing. They can’t seem to reason through XYZ space and 3D math. It’s my current litmus test for LLM performance.

I have had moderate success describing things as geometric primitives. I.e. making a simple phone stand is hard to one shot. Had to do it in steps as make a plane with XYZ dimensions, rotate it X degrees up, calculate the height using trig, create another plane with that height, translate to X position. Etc.

This has been my experience also. Curious if this will improve with just more training data, or if another method will need to be introduced.

Building models in code like this is really cool; it’s great to be able to import, for example, a “gears” library and automatically generate the geometry for complex bevel gear setups. I’ve seen this approach used for more than CAD too; there’s a Python library (GDSFactory) for building photonic ICs in code as well, and I’m sure plenty more.

This workflow is a big motivator for my data notebook project (https://mnty.sh/#serenity), since I would like to build everything for a project in code and have visualizers for each component in one notebook.

I found that an MCP server which can render the model and send the image to the LLM helps a lot.

* https://github.com/jabberjabberjabber/openscad-mcp

this doesn't make sense from a theoretical pov at least to me. like is it also defining tolerances etc?

No? Do you mean fits? If you have to go in later and define dimensions and tolerances by hand, that's still way less than modelling from scratch.

Oh man! The sketch into CAD is going to be huge for the restoration market! Holy crap!

There’s tons of old drawings but lacking measurements or something. The engineer knows how tall it is or what the overall dimensions are so they can easily use that to create a box and tell AI, make this part that fits this thing. This is going to change everything.

3D printing, restorative, imagineering, part manufacturing, everywhere where there is CAD…

There's this cool site that already done this for like 2 years https://text-to-cad.zoo.dev

Hey! I'm a cofounder and the CTO of this company. Want to lead with saying its an absolute privilege to be on the front page of this site and have all of you try this model and give us your feedback. Your time and attention is valuable and is absolutely not taken for granted. Our model is SGS-1, and it will continue to improve in future versions. Want to address a few points made by others here.

> can generate fully [...] parametric 3D geometry. >> This is obviously a lie without even testing it. STEP files do not have any sort of support for parametric features.

This is a simple misunderstanding/semantic issue. We aren't trying to misrepresent anything, we (and some others in the research community) interpret parametric as "being composed of primitives with parameters". We have an internal representation we don't expose to users, and we convert that to a B-Rep STEP file which we do expose.

>> Every dimension is wrong aside from the one that I corrected to get the same scale (there doesn't appear to be any correct relative to each-other which is why I just picked one at random)

Visually this output looks close. There are some gaps, but looking at your screenshot, features like the main hole going all the way through are underdetermined. It's very hard for a model to know because even if the hole did not go through the input would look the same. All of this can be fixed by having richer input conditions, and we are actively working on this for SGS-2

>Yes, this is also confusing me to no end. How can they make such a claim? They even explicitly state that they generate a B-rep (boundary representation) output only, then in their roller example they say "as the output is parametric, dimensions can easily be adjusted." Erm, no? I'd rather model it again with the proper feature history tree and constraints instead of fiddling with a step file.

SGS-2 will include a feature tree representation, so this should be more easily editable than this (which should already be easily editable in many cases as well with direct modeling). Several engineers already work with STEP files in their work for to the best of my knowledge.

Thanks once again for all the feedback! We are incorporating all this into our next model.

> This is a simple misunderstanding/semantic issue. We aren't trying to misrepresent anything, we (and some others in the research community) interpret parametric as "being composed of primitives with parameters". We have an internal representation we don't expose to users, and we convert that to a B-Rep STEP file which we do expose.

No officer I wasn't speeding, I have an internal definition of velocity that you don't have access too and it doesn't say I was speeding

"I was enjoying a beverage while travelling"

Look - this is a really bad answer in a lot of ways, but rather than excoriate you for it, i'll try to be helpful by pointing out a simple fact:

You are saying that it does some things. Literally everyone who could ever want to pay you money to use this or future models are going to try it, see that it does not appear to do what they expect for those things, and then:

1. Not want to pay you money

2. Be much harder to attract again in the future. Once bitten, twice shy.

This seems like a bad plan?

> we (and some others in the research community) interpret parametric as "being composed of primitives with parameters".

Well guess what? You're wrong.

You seem to be generating output files compliant with ISO 10303-214, this standard has been withdrawn and replaced by ISO 10303-242, edition 4 of 10303-242 has just been published.

I don't get why people are downvoting this due to the parametric claim.

Yes, STEP files don't have a feature tree and cannot be parameterized. I read it as OP saying that under the hood there __is__ parameterization before the final export to STEP.

This means that they could expose this in the future, say if they chose to output to FreeCAD FCStd or even some proprietary format like Solidworks (via $$$ CAD translator packages).

This is just my read - I'm not affiliated and have no internal knowledge to Spectral or SGS-1, but I have worked deep in building CAD plug-ins and custom software for manufacturing automation.

All that said, the demo model quality has issues and many are unmanufacturable (with subtractive means at least) - so there's still along way to go. But I don't think the presentation of their capabilities is disingenuous.

Congrats on the launch of v1, keep going!

If this really works, in conjunction with a 3D printer and a robotic arm, the implications will be huge…

Glad to see this is utter BS. Moving along.

Reply on news.ycombinator.com