Nvidia's LLaMA-Mesh weights have been released on HF. Inference code is available.

[-]

Xanian123@reddit

What will this help us do? I don't see the logic behind this

[-]

Comms@reddit

Mesh, create a series of cabinets that fit in this workspace: x=50.8 y=102 z=96.

Maximum dimensions per cabinet: height=34.5, depth=24, width=32. Kick is inset: depth=3, height=4

Material thickness: Frame=0.75, shelving=0.5, stretchers=0.75, drawer_support=0.75

Create as many cabinets to that specification as will fit in the space provided.

[-]

I can see all kinds of uses for this if you use CAD. You can have a materials library, examples library, fixtures library, appliances library, etc. Just dropping the space dims and getting a kitchen layout generated into a STEP file with all parts grouped into components would be a really nice time-saver.

Same for any craft that uses CAD. Furniture making, machining/CNCing machine parts for small scale production—I can't imagine this would be that useful for a large manufacturer. Training it on how to build and design shop machines would be really useful too. Different sized belt-grinders based on custom specs.

All of this can just be done in CAD by hand, but getting something generated that gets you 80% of the way there would be great. I have libraries and templates for everything I CAD but the early stages of design are frequently the same thing. The real work is in the details and making everything work.

[-]

Xanian123@reddit

By God, AI generated and 3d printed one of a kind slop is coming, isn't it?

[-]

Comms@reddit

Probably. Thingiverse is going to be slammed with dogshit.

That said, the tool itself could be useful.

[-]

Any_Pressure4251@reddit

Eventually produce assets for games, films. Web pages, 3d printers, applications etc.

[-]

Beautiful_Help_3853@reddit

create a 3D model of a cube :

That definitively worth the cash spent to train that model.

[-]

KrazyKirby99999@reddit

They need Niantic's dataset https://lunduke.locals.com/post/5756204/the-cia-nsa-and-pok-mon-go

[-]

hugganao@reddit

Lol

[-]

JustinPooDough@reddit

I told it to generate a 3D model of a "Phallus of the Species Homo Sapien". It said no. Boring.

[-]

tucnak@reddit

Honestly, their dataset would be more interesting than the model itself.

[-]

nahojjjen@reddit

Interesting concept/research, even if this first iteration seems quite limited. :) Guess what I asked for to get this 3d model:

[-]

DIBSSB@reddit

Can we use this in ollama or llmstudio ?

[-]

Illustrious-Lake2603@reddit

I made one at q6 on hugging face https://huggingface.co/CronoBJS/LLaMA-Mesh-Q6_K-GGUF

It works well in LM Studio. Wayyyyyyyyyy faster than their app. The only thing is the visualization of the mesh. I updated their app.py to remove the interference portion and left the visualization so I can test what it came out of it. I'm at work right now so I can upload the gradio app.py yet

[-]

Ylsid@reddit

Thanks for the q6! Can we get a q4? :)?

[-]

DIBSSB@reddit

Thanks

[-]

MrTubby1@reddit

One step closer to the holodeck.

[-]

i_m_old_rabbit@reddit

One step closer to холодец.

[-]

sluuuurp@reddit

Is there any example of it generating something that doesn’t already exist? An object that you can describe in text, but that you can’t find on google?

If not, then I guess it’s an interesting research step, but pretty useless for now.

[-]

ortegaalfredo@reddit

Al last I can say "Computer kick up the uh, 4d3d3d3"

[-]

Pro-editor-1105@reddit

how tf do i use this?

[-]

FullOf_Bad_Ideas@reddit (OP)

HF space is here, free use is limited though.

Inference code is here.

Demo code runs on Windows quite fine, a bit slow (could be just my as I often have this issue on Windows), it should be painless on Linux though.

[-]

Pro-editor-1105@reddit

it is garbage, I tell it to make a 3d model of a garden and it just tells me what a garden is.

[-]

Admirable-Star7088@reddit

Have you tried to tell it to make a 3D model of garbage?

[-]

FullOf_Bad_Ideas@reddit (OP)

I did try, it went quite terrible.

https://pixeldrain.com/u/ZsBaWj9k

[-]

Pro-editor-1105@reddit

i ran out of requests before I could.

[-]

FullOf_Bad_Ideas@reddit (OP)

It can do simple objects only.

[-]