Feature matrix
| Criterion | Model A | Model B |
|---|---|---|
| Primary input | Single image | Single image |
| Pipeline design | O-Voxel shape and material generation | Separate shape and PBR paint models |
| PBR output | Base color, roughness, metallic, opacity | Dedicated PBR texture synthesis |
| Official local memory note | 24 GB+ NVIDIA GPU baseline | 10 GB shape, 21 GB texture, 29 GB combined |
| Upstream requirements use different implementations and are not a speed or quality ranking | ||
| Code availability | Inference and training code | Weights and training code for shape and paint |
| Key inspection | Open surfaces, density, alpha setup | Shape-to-paint consistency and stage memory |
Choose by workflow
High-resolution O-Voxel assets and native PBR channel inspection
Workflows that benefit from separate shape and material iteration
Practical verdict
Start with TRELLIS.2 when O-Voxel topology handling and one integrated high-resolution PBR pipeline fit the task. Test Hunyuan3D 2.1 when separate geometry and paint stages are useful. Inspect both on your own asset class before committing.
Compare geometry without texture
Use silhouette, hidden-side, boundary, thin-part, and normal checks in a neutral clay render before appearance influences the judgment.
Compare material response
Inspect maps individually and under the same HDR environment. Note baked lighting, metal boundaries, roughness range, seams, and alpha configuration.
Method and limits
Use the same source image, preserve each model's recommended settings, render neutral clay and PBR views, and report geometry, channel completeness, file size, peak memory, hardware, and failures. Do not compare vendor timing claims from different machines.