#422 TRELLIS.2
About TRELLIS.2, an open-source image to 3D generation model. TLDR - very disappointing results from my initial tests.
Notes
TRELLIS.2 is an open-source 4B-parameter image-to-3D model producing up to 1536 PBR textured assets, built on native 3D VAEs with 16× spatial compression, delivering efficient, scalable, high-fidelity asset generation.
A ran a few tests using the demo hosted at huggingface, but am not getting great results.
Line Drawing Trial
Image generated with ChatGPT:
create a black and white outline drawing of a taiko drummer, in dramatic pose holding bachi. Only the figure, on a clear background without taiko or other elements
Not a bad image, but note the legs are not complete. This could be a common challenge for 3D model generators, given incomplete source material.
Let’s see how TRELLIS.2 performs, using the demo hosted at huggingface. I used default settings all the way:
I exported the 14.1MB GLB file and loaded into blender:
Evaluating the results:
- It wasn’t smart enough to complete the missing leg detail
- Hidden hand details are poorly completed anatomically incorrect
- Generated some random “floating” elements
In all, a decent start, but would require a lot of cleanup to be a usable 3D model.
Photo Trial
I started with this image from https://southwestfolklife.org/ken-koshio-taiko-player/
I generated a model using default settings using the demo hosted at huggingface:
Evaluating the results:
- useless; very disappointing
- clearly failed to properly comprehend the geometry of the source