This is the kind of porting work I always hope for when I see a CUDA-only release. Have you thought about publishing the gather-scatter sparse 3D convolution and SDPA attention swaps as a standalone toolkit or writeup? A lot of folks running models locally on Apple Silicon hit the same wall with flash_attn, nvdiffrast, and custom sparse kernels and end up redoing the same work.
Most 'runs on Mac' ports are a wrapper around a cloud call or a quantized shell of the original model. Going after the CUDA-specific kernels with pure-PyTorch alternatives is the kind of work that ages well, because the next CUDA-locked research release is three weeks away. One question: how much of the gather-scatter sparse conv is reusable for other TRELLIS-like architectures, or is it bespoke to this one?
Nice work. Although this model is not very good, I tried a lot of different image-to-3d models, the one from meshy.ai is the best, trellis is in the useless tier, really hope there could be some good open source models in this domain.
Hey, thanks for sharing this. I'm sure TRELLIS.2 definitely has room to improve, especially on texturing.
From what I've seen personally, and community benchmarks, it does fair on geometry and visual fidelity among open-source options, but I agree it's not perfect for every use case.
Meshy is solid, I used it to print my girlfriend a mini 3d model of her on her birthday last year!
Though worth noting it's a paid service, and free tier has usage limitations while TRELLIS.2 is MIT licensed with unlimited local generation. Different tradeoffs for different workflows. Hopefully the open-source side keeps improving.
Meshy is indeed great but I am terminally put off by their alltogether terrible, sleazy, gamified, opaque web UI. It's like aliexpress and a lootbox game had a baby that's into mesh generation. Ugh.
How much RAM does this use? Only sitting on 8 GB right now, I'm trying to figure out if I should buy 24 GB when it's time for a replacement or spring for 32.
The model needed about 15GB at peak during generation - the 4B model loads multiple sub-models (1.3B each for shape and texture flow). 8GB won't be enough, but both 24GB and 32GB both should be fine.
Thanks! Could it conceivably load the sub-models in series rather than parallel? 8 still won't be enough but I wonder if those with 16 could eke something out.
That’s always been possible with MPS backend, the reason people choose to omit it in HF spaces/demos is that HF doesn’t offer an MPS backend. People would rather have the thing work at best speeds than 10x worse speeds just for compatibility.
IMO TRELLIS.2 is slightly different case from the HF models scenario. It depends on five compiled CUDA-only extensions -- flex_gemm for sparse convolution, flash_attn, o_voxel for CUDA hashmap ops, cumesh for mesh processing, and nvdiffrast for differentiable rasterization. These aren't PyTorch ops that fall back to MPS -- they're custom C++/CUDA kernels. The upstream setup.sh literally exits with "No supported GPU found" if nvidia-smi isn't present. The only reason I picked this up because I thought it was cool and no one was working on this open issue for Silicon back then (github.com/microsoft/TRELLIS.2/issues/74) requesting non-CUDA support.
It’s always been possible, but it’s not possible because there’s no backend, and no one wants to it to be possible because everyone needs it 10x the speed of running on a Mac? I’m missing something, I think.
I thought it was cool and then I found the open issue mentioned above, that convinced me its def something more people want.
It IS significantly slower, about 3.5 minutes on my MacBook vs seconds on an H100. That's partly the pure-PyTorch backend overhead and partly just the hardware difference.
For my use case the tradeoff works -- iterate locally without paying for cloud GPUs or waiting in queues.
This is the kind of porting work I always hope for when I see a CUDA-only release. Have you thought about publishing the gather-scatter sparse 3D convolution and SDPA attention swaps as a standalone toolkit or writeup? A lot of folks running models locally on Apple Silicon hit the same wall with flash_attn, nvdiffrast, and custom sparse kernels and end up redoing the same work.
that makes so much sense...I am exploring if I can find someone who has done this well...If not I'll try to do it myself.
Most 'runs on Mac' ports are a wrapper around a cloud call or a quantized shell of the original model. Going after the CUDA-specific kernels with pure-PyTorch alternatives is the kind of work that ages well, because the next CUDA-locked research release is three weeks away. One question: how much of the gather-scatter sparse conv is reusable for other TRELLIS-like architectures, or is it bespoke to this one?
Does it support multi-view input?
Nice work. Although this model is not very good, I tried a lot of different image-to-3d models, the one from meshy.ai is the best, trellis is in the useless tier, really hope there could be some good open source models in this domain.
Hey, thanks for sharing this. I'm sure TRELLIS.2 definitely has room to improve, especially on texturing.
From what I've seen personally, and community benchmarks, it does fair on geometry and visual fidelity among open-source options, but I agree it's not perfect for every use case.
Meshy is solid, I used it to print my girlfriend a mini 3d model of her on her birthday last year!
Though worth noting it's a paid service, and free tier has usage limitations while TRELLIS.2 is MIT licensed with unlimited local generation. Different tradeoffs for different workflows. Hopefully the open-source side keeps improving.
Meshy is indeed great but I am terminally put off by their alltogether terrible, sleazy, gamified, opaque web UI. It's like aliexpress and a lootbox game had a baby that's into mesh generation. Ugh.
Great. Potentially can go much faster rewriting it in terms of Metal shaders.
This is fantastic, great work. I will attempt to run it on my 16GB M1 but I doubt it'll run.
Out of curiosity, how did you go about replacing the CUDA specific ops? Any resources you relied on or just experience? Would love to learn more.
So much effort, but no examples in the landing page.
You're right, thanks for flagging this, let me run something and push images
added! will add more, maybe even a GIF
How much RAM does this use? Only sitting on 8 GB right now, I'm trying to figure out if I should buy 24 GB when it's time for a replacement or spring for 32.
The model needed about 15GB at peak during generation - the 4B model loads multiple sub-models (1.3B each for shape and texture flow). 8GB won't be enough, but both 24GB and 32GB both should be fine.
Thanks! Could it conceivably load the sub-models in series rather than parallel? 8 still won't be enough but I wonder if those with 16 could eke something out.
That’s always been possible with MPS backend, the reason people choose to omit it in HF spaces/demos is that HF doesn’t offer an MPS backend. People would rather have the thing work at best speeds than 10x worse speeds just for compatibility.
IMO TRELLIS.2 is slightly different case from the HF models scenario. It depends on five compiled CUDA-only extensions -- flex_gemm for sparse convolution, flash_attn, o_voxel for CUDA hashmap ops, cumesh for mesh processing, and nvdiffrast for differentiable rasterization. These aren't PyTorch ops that fall back to MPS -- they're custom C++/CUDA kernels. The upstream setup.sh literally exits with "No supported GPU found" if nvidia-smi isn't present. The only reason I picked this up because I thought it was cool and no one was working on this open issue for Silicon back then (github.com/microsoft/TRELLIS.2/issues/74) requesting non-CUDA support.
Are you saying the original one worked with MPS? Or are you just saying it was always theoretically possible to build what OP posted?
It’s always been possible, but it’s not possible because there’s no backend, and no one wants to it to be possible because everyone needs it 10x the speed of running on a Mac? I’m missing something, I think.
I thought it was cool and then I found the open issue mentioned above, that convinced me its def something more people want.
It IS significantly slower, about 3.5 minutes on my MacBook vs seconds on an H100. That's partly the pure-PyTorch backend overhead and partly just the hardware difference.
For my use case the tradeoff works -- iterate locally without paying for cloud GPUs or waiting in queues.
Well done
rad. how long does output take? trellis is a fun model.
i was able to get it in 3.5 mins from a single image on my 24gb m4 pro macbook
I'm still working on this to try to replicate nvdiffrast better. Found an open source port, might look it tonight
thanks!