Is TRELLIS v1 obsolete now that TRELLIS.2 exists?

For hero-quality assets, yes, v2 is the one to use. But v1 generates 13× faster in our measurements (864 vs 66 assets/hour) and needs ~7GB instead of ~30GB, so for bulk drafts, candidate generation and background props it remains the practical tool.

What GPU does TRELLIS need?

Any 8GB card. Measured peak was ~7GB. On rented hardware the L40S is the value pick: 776 assets/hour at 90% of H100 speed for far less per hour.

How fast is it really?

4.2 seconds per asset on an H100 in 10-batch pace; even an L4 takes only 12.6 seconds. For iterating on a concept, that's fast enough to stay in flow.

Textured 3D assets from a single input image, meshes usable in game engines and DCC tools, though production use typically wants retopology. Quality is respectable but visibly a generation behind TRELLIS.2.

What's the best workflow combining v1 and v2?

Volume-then-quality: batch v1 at hundreds of drafts per hour to explore, shortlist the winners, then re-run those through TRELLIS.2 at max quality. You spend the expensive 55-second v2 generations only on assets that earned it.

TRELLIS Image-to-3D · 6 GPUs measured first-party · image-to-3D · Updated July 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: 864 assets/hour on an H100 (4.2 seconds per asset in batch), with a ~7GB peak VRAM floor that puts it on ordinary 8GB cards.

Benchmarked weights: microsoft/TRELLIS-image-large

864.0assets/hr

Fastest: NVIDIA H100 80GB HBM3

measured, 3-run llama-bench

~7GB

VRAM needed (measured peak)

GPU-independent, applies to every card

GPUs measured

same pinned harness

4.17s

Per asset (10-batch pace)

warmup excluded

What GPU Do You Need for TRELLIS Image-to-3D?, assets/hour, fastest 6

NVIDIA H100 80GB HBM3

864 Assets/hour

NVIDIA H200

818.8 Assets/hour

NVIDIA L40S

776.1 Assets/hour

NVIDIA A100 80GB SXM4

488.7 Assets/hour

NVIDIA A10G

293.6 Assets/hour

NVIDIA L4

284.8 Assets/hour

Measured on our own bench. A card absent from this chart has not been run on this model yet, or cannot fit it.

TRELLIS Image-to-3D. Measured 3D generation speed by GPU

NVIDIA H100 80GB HBM3864

NVIDIA H200818.8

NVIDIA L40S776.1

NVIDIA A100 80GB SXM4488.7

NVIDIA A10G293.6

NVIDIA L4284.8

GPU	Assets/hour	s per asset	Single run (s)	Avg power
NVIDIA H100 80GB HBM3	864	4.17	4.09	323.8 W
NVIDIA H200	818.8	4.4	4.44	316.8 W
NVIDIA L40S	776.1	4.64	4.98	267.8 W
NVIDIA A100 80GB SXM4	488.7	7.37	7.63	225.2 W
NVIDIA A10G	293.6	12.26	12.34	196.6 W
NVIDIA L4	284.8	12.64	12.27	68.3 W

Where v1 stands now: the throughput tool. Let's be straight, TRELLIS.2 has superseded this model on quality, and if you're producing hero assets, that's the page to read. But calling v1 obsolete misses what the numbers say: 864 assets an hour against TRELLIS.2's 66, thirteen times the throughput, and a ~7GB floor against ~30GB. That's a different tool, not a worse one. Bulk-drafting game props, generating candidate meshes to cherry-pick from, populating background geometry: v1's speed-per-dollar at those jobs is untouchable, and it runs on hardware you already own.

The measurements. The chart is unusually flat for a generation model, H100 at 864, H200 at 819, L40S at 776 assets/hour, and even the humble L4 (68W measured) produces an asset every 12.6 seconds. Diffusion-based 3D at this resolution is compute-light by modern standards, so mid-tier cards deliver most of the experience. If you're renting, the L40S is the value row: 90% of H100 throughput at a fraction of the hourly rate.

Our verdict

TRELLIS v1: 864 assets/hour peak, ~7GB floor, outclassed on quality by TRELLIS.2, unbeatable on volume. Use v2 for the assets people will look at closely; use this to generate the hundred candidates you pick them from. On an 8GB card, this is the only image-to-3D game in town.

FAQ

Is TRELLIS v1 obsolete now that TRELLIS.2 exists?: For hero-quality assets, yes, v2 is the one to use. But v1 generates 13× faster in our measurements (864 vs 66 assets/hour) and needs ~7GB instead of ~30GB, so for bulk drafts, candidate generation and background props it remains the practical tool.
What GPU does TRELLIS need?: Any 8GB card. Measured peak was ~7GB. On rented hardware the L40S is the value pick: 776 assets/hour at 90% of H100 speed for far less per hour.
How fast is it really?: 4.2 seconds per asset on an H100 in 10-batch pace; even an L4 takes only 12.6 seconds. For iterating on a concept, that's fast enough to stay in flow.
What does it output?: Textured 3D assets from a single input image, meshes usable in game engines and DCC tools, though production use typically wants retopology. Quality is respectable but visibly a generation behind TRELLIS.2.
What's the best workflow combining v1 and v2?: Volume-then-quality: batch v1 at hundreds of drafts per hour to explore, shortlist the winners, then re-run those through TRELLIS.2 at max quality. You spend the expensive 55-second v2 generations only on assets that earned it.