Trellis 2 review: image-to-3D tests and GLB export

This review tests Microsoft's Trellis 2, an open source image-to-3D model, with several pictures rather than relying only on supplied samples. The presenter leaves resolution and advanced settings unchanged for the first run, inspects the preview with different materials, then exports a GLB file. The description identifies an NVIDIA A100 as the test GPU.

Results vary by image. The presenter finds the gear reconstruction unusable and rejects another result after a reported five-second generation. A propeller captures a plausible helix pitch, but the speaker notes inaccurate areas and inferred geometry. The unicorn earns the strongest praise. These examples support the presenter's advice to choose clear pictures with suitable lighting and angles, rather than expecting precise reconstruction from every input.

The tested interface accepts one photo at a time and has no text prompt option. The presenter suggests online conversion of GLB files to STL or OBJ. A Meshmixer inspection finds that the exported unicorn needs editing before 3D printing; the video does not establish print readiness for other outputs.

For running this local AI model, the description lists 8 GB NVIDIA VRAM with optimized community workflows, or 16 to 24 GB for native configurations at higher resolutions of 1024³ to 1536³. It lists 16 GB system RAM minimum and 32 GB recommended. Reported generation times describe this test, not guaranteed performance on those configurations.