TRELLIS Image-to-3D · 6 GPUs measured first-party · image-to-3D · Updated October 2026
What GPU Do You Need for TRELLIS Image-to-3D?
TRELLIS is Microsoft's first-generation image-to-3D model: drop in a picture, get a textured 3D asset. We measured it on 6 GPUs all the way to a finished GLB file: 167 assets an hour on an H100 (21.5 seconds each), and the 3D generation is only 3-4 seconds of that. The rest is the GLB export, which simplifies the mesh and bakes its textures.
Benchmarked weights: microsoft/TRELLIS-image-large
NVIDIA H100 80GB HBM3
167.2 assets/hour on TRELLIS Image-to-3D, the ceiling. Measured on our bench. 80GB of VRAM, $30,000 at launch.
- 167.2 assets/hour on TRELLIS Image-to-3D
- 80GB, clears the TRELLIS Image-to-3D floor
- Rentable by the hour rather than bought
- 700W board rating
- Datacenter or workstation hardware, not a retail purchase
What GPU Do You Need for TRELLIS Image-to-3D?, assets/hour by GPU
Measured on our own bench. A card absent from this chart has not been run on this model yet, or cannot fit it.
Value: assets/hour per $1,000 of MSRP
Launch price, not street price, so it ages. A speed leaderboard always crowns the most expensive card; this is the counterweight.
TRELLIS Image-to-3D. Measured 3D generation speed by GPU
| GPU | Assets/hour | s per asset | Peak VRAM | Avg power |
|---|---|---|---|---|
| NVIDIA H100 80GB HBM3 | 167.2 | 21.54 | 11.7GB | — |
| NVIDIA H200 | 134.6 | 26.74 | 11.7GB | — |
| NVIDIA L40S | 134.5 | 26.76 | 11.7GB | — |
| NVIDIA A100 80GB SXM4 | 116.4 | 30.93 | 11.7GB | — |
| NVIDIA A10G | 94 | 38.31 | 11.7GB | — |
| NVIDIA L4 | 68.8 | 52.35 | 11.7GB | — |
Where v1 stands now: the volume tool. TRELLIS.2 has superseded this model on quality, and for hero assets that's the page to read. Measured honestly, to a finished GLB on both, v1 is about twice as fast on the same card and needs half the memory. For bulk props, background clutter and candidate shapes to pick from, that's the trade worth making.
Where the time goes. The 3D generation is the quick part: 3.3-4.2 seconds per asset on an H100, 5.3-6.9 on an L40S, 12.8-24.9 on an L4. The GLB export (mesh simplification plus a texture bake that runs on the GPU) takes 70-85% of every asset's time, and it grows with detail: 9 seconds for a pistol on an H100, 45 for a leafy tree. So the card matters less than the chart suggests for simple props, and a lot for detailed ones. If you only need a quick look, TRELLIS's splat preview skips the export entirely.
About TRELLIS Image-to-3D. TRELLIS Image-to-3D: from microsoft, on Hugging Face since December 2024, MIT licence. 2,209,080 downloads in the last 30 days and 19 community quantizations.
How it compares. H100 80GB HBM3: TRELLIS Image-to-3D 167.2 assets/hour, TRELLIS.2 Image-to-3D 87.9 (4B), TripoSG 559.4 (1B), TripoSR 2107.4. 2 of 3 beat TRELLIS Image-to-3D here. These aren't like-for-like: TripoSR returns a rough vertex-coloured mesh and TripoSG bare geometry, while both TRELLIS numbers are timed to a finished, textured GLB.
Cost on a rented GPU. 1,000 assets of TRELLIS Image-to-3D: $5.87 on a L40S ($0.79/hr, 7.4 hours), $12.78 on a H100 80GB HBM3 ($2.14/hr, 6.0 hours, 2.2x the cost).
TRELLIS Image-to-3D: cost per 1,000 assets on rented GPUs
| GPU | Cheapest rate | Speed (assets/hour) | Cost per 1,000 assets |
|---|---|---|---|
| NVIDIA L40S | $0.79/hr | 134.5 | $5.87 |
| NVIDIA L4 | $0.44/hr | 68.8 | $6.40 |
| NVIDIA A100 80GB SXM4 | $0.95/hr | 116.4 | $8.14 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 167.2 | $12.78 |
| NVIDIA H200 | $3.59/hr | 134.6 | $26.67 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for TRELLIS Image-to-3D. 60-360 assets/hour: 6 (H100 80GB HBM3, H200, L40S). 360 assets/hour is ten seconds per mesh.
VRAM for TRELLIS Image-to-3D. Measured peak 11.7GB, so 16GB is the smallest common card size; smallest card it ran on: A10G (24GB).
Our verdict
TRELLIS v1 is still the fastest way to a textured GLB we measured: 167 an hour on an H100, 135 on an L40S, about twice TRELLIS.2's pace on the same card. It used about 12GB at its peak. Use it for background props and for drafts you'll pick from; use TRELLIS.2 for the assets people look at closely.
FAQ
- Is TRELLIS v1 obsolete now that TRELLIS.2 exists?
- Not for volume work. Measured to a finished GLB, v1 makes about 1.9 assets for every one TRELLIS.2 makes on the same card (167 against 88 an hour on an H100), and it needs about 12GB against TRELLIS.2's 24GB. TRELLIS.2's shapes and textures are clearly better, so use it for the assets that matter.
- What GPU does TRELLIS need?
- Its measured peak was about 12GB (11.7GB), so a 16GB card runs it with room to spare and 12GB is tight; Microsoft lists 16GB. Renting, the L40S is the value pick: 135 assets an hour at about $0.79/hr, well under a cent per asset.
- How fast is it really?
- The 3D generation takes 3-4 seconds on an H100 and 5-7 on an L40S. The GLB export takes longer: 9-14 seconds for simple props on an H100, and 45 seconds for our tree, whose detail takes the longest to bake. A full asset averages 21.5 seconds on an H100 and 52 on an L4.
- Why are these numbers lower than before?
- Our first run, in July, timed only TRELLIS's Gaussian splat output, which is fast (864 an hour on an H100) but isn't a mesh a game engine can use. Since October 2026 every TRELLIS number on this site is timed to a finished, textured GLB, the file you actually import.
- What's the best workflow combining v1 and v2?
- Draft with v1, or with TRELLIS.2's 512 setting (about 6 seconds to generate on an H100), pick the shapes you like, then make the finals with TRELLIS.2 at its default setting. Keep one worker running for the whole batch: a fresh TRELLIS.2 worker spends its first few assets tuning kernels.
How we test
Warm worker: one pass over five game-asset pictures (rifle, pistol, magazine, crate, tree, as transparent cut-outs) to warm up, then the same five timed one after another. Each asset is timed from picture to a finished GLB: TRELLIS-image-large's default sampler settings, then the official to_glb export at 95% simplification with 1024px textures. Peak VRAM is the most PyTorch allocated during the timed pass. Datacenter cards on Modal, October 2026.