Repository navigation
tutorials/3d/trellis2 #1869
Replies: 7 comments 1 reply
|
With the described workflow, there is no way to upload an image! You can only use text prompts. |
|
There is no LoadImage node |
|
Hi @guntherramthun-bit, thanks for flagging this. I checked the exact template this page links to (
One detail you mention does not match this template, though: it has no text-prompt input at all (no To help me pin it down, could you share:
That will tell me whether it is a stale template on your end, or something the docs should call out more clearly. |
|
Can you do multiview image input based on that template? |
|
Hi @Ryptun, good question. Short answer: not with that template, but there is a dedicated multiview workflow for the shared Pixal3D / TRELLIS.2 pipeline.
One caveat: the multiview path is wired for Pixal3D, not TRELLIS.2. In ComfyUI the TRELLIS.2 conditioning node takes a single image and the multiview template has no TRELLIS.2 switch, so you cannot currently feed a turnaround sheet into the TRELLIS.2 model through this workflow. If you tell me what object you are trying to reconstruct, I can point you at the right template. |
|
Will this run on a Mac? I've spent a full 8 hours trying to get trellis running on ComfyUI. I finally managed (thanks to Chat GPT) to get Trellis-Mac and its related custom node working, but the output is very poor quality compared to the demo on Hugging Face. So much ComfyUI stuff demands CUDA, but it is never called out. I just want to know if I should give this one a try. Thanks. |
|
Hi @alsmithson, yes, worth trying, but through the native workflow on this page rather than the trellis-mac custom node path. Two separate things explain what you saw. The quality gap is that port's settings, not your install. trellis-mac defaults to its On Apple Silicon, prefer the bf16 checkpoints. The template's default checkpoints are int8 Expect it to be slow. The native pipeline is pure PyTorch/SciPy, so it does run without CUDA, but sparse sampling and mesh post-processing are much slower on MPS than on an NVIDIA GPU. Budget 24 GB+ unified memory and a long run per mesh. If you just want to see the documented output level first, run the same template on Comfy Cloud: https://cloud.comfy.org/?template=3d_pixal3d_trellis2_image_to_model If the native workflow errors or crawls on your Mac, paste your chip, unified memory and the console traceback and I will take a look. |
Uh oh!
There was an error while loading. Please reload this page.
tutorials/3d/trellis2
Generate high-fidelity 3D models with full PBR textures from a single image using TRELLIS.2, Microsoft's open-source 4B-parameter image-to-3D model.
https://docs.comfy.org/tutorials/3d/trellis2
All reactions