Benchmarking Trellis 2: GLB Output Quality Across Six Input Types

Everyone shows off AI image-to-3D with the same glossy product render on a white background. That tells you almost nothing about how the tool behaves on your input. So I ran a proper bake-off: one model — Trellis 2, the SLAT-based image-to-3D generator — fed six deliberately different input types, all judged on the output that actually matters to devs: the GLB you import into your engine.

The test protocol

Six input types, one subject each, all run through Trellis 2 in the browser (it's cloud-based, so no local GPU variables in the mix). Scoring: silhouette accuracy, topology sanity (can the mesh be decimated without exploding), texture coherence, and time-to-import into Unity.

1. Studio product shot (white background) — the control group
2. Amateur phone photo (cluttered desk background)
3. Game concept art (stylized painting)
4. Screenshot of an existing low-poly model (testing the "remaster" use case)
5. Hand-drawn sketch (pencil on paper)
6. Multi-view collage (front+side pasted into one image)

Results

Studio shot: Flawless, as advertised. Clean topology, 2/2 import. This is the marketing photo use case and it delivers.

Phone photo: The surprise — the model ignored the desk clutter impressively well. Silhouette true, some texture hallucination on the back side (which no single-photo method can know). Usable for background props with 10 minutes of cleanup. 8/10.

Concept art: This is where SLAT's learned-prior approach shines. The stylized painting produced a stylized mesh that matched the art direction — something photogrammetry literally cannot do (it would need the object to exist). For stylized games, this is the killer input type. 9/10.

Low-poly screenshot: The "remaster" workflow works: feed an old asset's screenshot, get a modernized mesh. Ironically the output sometimes inherited the input's angularity — garbage in, slightly better garbage out. Still a legitimate retro-asset pipeline. 7/10.

Sketch: The wildest card. A pencil sketch produced a coherent, charming "concept-model" — wrong in the ways your drawing was wrong. As a tool for pre-visualizing your own sketch in 3D, genuinely useful. As production asset, no. 6/10 for production, 9/10 for ideation.

Collage: Mixed views in one image confused the reconstruction more than helped. Front-only input beat my hacky collage. Don't do this — if you want multi-view, that's what dedicated modes are for. 4/10.

Practical takeaways

• Best input: clean silhouette, one dominant view, honest lighting. Backgrounds matter less than I feared.
• Topology: GLB exports imported cleanly every time (6/6); mesh density skews dense, so plan a decimation pass for mobile targets.
• Speed: every generation landed in the 1–3 minute window, cloud-side, browser-only.
• The honest ceiling: single-image methods can't know your object's back side. For hero assets, plan manual back-side cleanup or multi-view inputs.

Verdict

For game devs, the sweet spot is inputs 2 and 3: real-world reference photos and concept art. That covers about 80% of "I need a prop by Friday" situations. Benchmark score: Trellis 2 is the first image-to-3D tool where I'd trust a phone photo enough to skip the studio-shot ritual — and that's the whole ballgame.