MICROSOFT DROPPED A 4B PARAMETER MODEL THAT TURNS ONE IMAGE INTO A 3D ASSET IN 3 SECONDS and it's open source TRELLIS.2 fully textured, physically accurate 3D models with PBR textures out of the box not a rough mesh..not a placeholder roughness, metallic, opacity the kind of detail that... makes things look real under any lighting and it handles the weird stuff too..open surfaces, hollow interiors, geometry that breaks every other tool the model doesn't know the word "limitation" apparently demo is live on hugging face right nowshow more

The entire team is Chinese

It is amazing how well it did, and just as amazing at how wrong some simple parts were.

Yeah it works ok

Hmmmmm look at the prerequisites... 😶

this is one of those few times when i don’t have to check the actual model to call bullshit on your shilling ass, because it’s far too obvious. shame on you.

Women are going to use this to 3D print all the dick pics they get.

the gap most people miss: 'fully textured 3D' from one image is viewer-grade, not engine-grade. great for product previews and AR mockups. drop one into Unity and you'll spend the saved time on retopology + UV unwrap + LODs anyway. tested 3 of these last month for a video pipeline — only 1 mesh was clean enough to skip cleanup.

just tried it...horrible model. Maybe take a bit longer than 3 seconds?

4 months ago, these reposts are cancer.

huggingface demo: Error. After nothing, got the message: You reach you limit. Please upgrade to 9$ tier. As a good µ$oft tool it's a joke 😜

Trellis2 has been out for a while. That is an impressive demo that you showed. Yes, it does what yuou described; but the geo that is generated is, currently, good in a limited fashion. As a professional 3D artist the resulting geo has a limited scope. Will this change? I am sure so, but right now - it is nice, but limited in scope and application.

Like 5 months ago lol

bullshit you cant extrapolate interior shapes w exacting precision from a single photo

Open source 3D from a single image in 3 seconds. The demo is impressive. Maintaining texture quality across different lighting conditions is where it gets real.

Optimized topology or a million billion vertices?

I've done some benchmarking and analysis on this model running locally on consumer hardware. It's also worth noting that this model has been out for some time.

@TencentHunyuan THIS is how you do OPEN SOURCE ... open and everyone can actually use it ...

Damn, microsoft finally taking a teny tiny step.

ok but what happens when artists start feeding it concept art directly

Things are becoming cheaper

nice does it work with Bambu A1 printer

Yeah but it’s been out since dec 2025, right ? Or am I missing something ?

Yes

Here are some tips, use an image where the background is transparent and the subject is in the foreground. Have the subject completely in frame. Pick something that is easy or naturally 3D to start with.

@grok can this do sdf and frep ?

I wish @MaxonVFX could do this. Anytime I import a CAD, texture it, and then attempt to split it with Boole it more than likely will fail (and very possibly crash the software). Voronoi fracture is INCREDIBLY slow.

Sweet, now, 3D print

That 4B parameter model's open-source - that's a big deal. Could make 3D asset creation way more accessible.

3d printing about to become much easier

@llmgram

Microsoft was silent all these days and now cooking with launching free open source model for the users that are really efficient; Previously it was a 7B voice model and now with a 4B 3D asset model

Problems with every single model as of April 27, 2026. Zero Clean code. Zero TDD compliance. Zero Observability. Zero Verification workflows. Nothing is built in. AI models cannot become productive until these guard rails are built in. Vibing with zero verifiable results...

i'm having trouble understamding what this thing actually does or how a 4n parameter model works, can someone break dowb the math behind it for me?