Why do you recommend the newest architecture?

The data: an A100 80GB, same memory class as the H100, takes 110 seconds per asset where the H100 takes 57. This is a compute-dense diffusion workload that scales hard with architecture; Ampere-era cards, including the RTX 3090, will be painfully slow.

How long does one asset take?

55-57 seconds on H200/H100 at max quality (66 and 63 assets/hour in sustained batch), at 630-643W measured draw. Budget roughly a minute of top-tier GPU time per hero asset.

Is TRELLIS.2 worth it over v1?

For quality, unambiguously. It's the best open-source image-to-3D output there is. For volume, no: v1 is 13× faster. The strong workflow is both: draft with v1, finalize the shortlist with v2.

Should I rent or buy for it?

Occasional use: rent, an H200 hour yields ~66 max-quality assets. Pipeline use: this model is the best case in our data for owning 32GB flagship consumer silicon, since the max-quality floor clears 24GB cards out of the running entirely.

TRELLIS.2 Image-to-3D (1536³ max quality) · 3 GPUs measured first-party · image-to-3D · Updated July 2026

What GPU Do You Need for TRELLIS.2 Image-to-3D (1536³ max quality)?

TRELLIS.2 is the current king of open-source image-to-3D, and the most demanding model in our entire database. At its maximum-quality 1536³ setting we measured ~30GB peak VRAM and 55-57 seconds per asset on H200/H100 silicon, drawing over 630W sustained. This is a heavyweight workload, and the numbers deserve respect before you plan hardware around it.

Benchmarked weights: microsoft/TRELLIS-image-large

65.8assets/hr

Fastest: NVIDIA H200

measured, 3-run llama-bench

~30GB

VRAM needed (measured peak)

GPU-independent, applies to every card

GPUs measured

same pinned harness

54.74s

Per asset (10-batch pace)

warmup excluded

What GPU Do You Need for TRELLIS.2 Image-to-3D (1536³ max quality)?, assets/hour, fastest 3

NVIDIA H200

65.8 Assets/hour

NVIDIA H100 80GB HBM3

62.7 Assets/hour

NVIDIA A100 80GB SXM4

32.7 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.2 Image-to-3D (1536³ max quality). Measured 3D generation speed by GPU

NVIDIA H20065.8

NVIDIA H100 80GB HBM362.7

NVIDIA A100 80GB SXM432.7

GPU	Assets/hour	s per asset	Single run (s)	Avg power
NVIDIA H200	65.8	54.74	55.64	643.2 W
NVIDIA H100 80GB HBM3	62.7	57.41	57.74	633.2 W
NVIDIA A100 80GB SXM4	32.7	109.93	109.28	291.9 W

The one to use, with eyes open about hardware. On output quality there's no debate: TRELLIS.2 is the best open image-to-3D you can run, and it's the model we'd build an asset pipeline around. The catch is cost per asset. At the 1536³ max-quality preset we benchmarked, peak VRAM hit ~30GB, beyond every 24GB consumer card, and even an H200 needs 55 seconds per asset at 643W. Lower-resolution presets reportedly fit 24GB cards, but we haven't measured those yet and this page won't pretend we have. What we can say from the data: the A100's 110 seconds per asset shows older architectures falling off a cliff: this workload wants the newest compute you can get, and anything Ampere-era will crawl.

Planning numbers. 66 assets/hour on the H200, 63 on the H100, 33 on the A100 80GB, that architecture gap (2×, same generation of memory capacity) is the loudest signal in the chart. If you're renting: an hour of H200 produces about 66 max-quality assets; price that against your pipeline. If you're buying for local use, this model is the single best argument in our database for 32GB flagship consumer silicon over 24GB, the measured max-quality floor simply doesn't fit 24GB.

Our verdict

TRELLIS.2: the best open image-to-3D, measured at ~30GB peak and 55+ seconds per asset at max quality, datacenter cards or 32GB-class flagship silicon, newest architecture strongly advised (the A100 takes 2× the H100's time). Draft in volume on v1, spend v2's expensive minutes only on assets that matter.

FAQ

What GPU do you need for TRELLIS.2?: At the 1536³ max-quality setting we measured: ~30GB peak VRAM: an H100/H200 class card, or 32GB flagship consumer silicon. No 24GB card fits that preset; reduced-resolution presets reportedly fit 24GB, but we haven't measured them yet.
Why do you recommend the newest architecture?: The data: an A100 80GB, same memory class as the H100, takes 110 seconds per asset where the H100 takes 57. This is a compute-dense diffusion workload that scales hard with architecture; Ampere-era cards, including the RTX 3090, will be painfully slow.
How long does one asset take?: 55-57 seconds on H200/H100 at max quality (66 and 63 assets/hour in sustained batch), at 630-643W measured draw. Budget roughly a minute of top-tier GPU time per hero asset.
Is TRELLIS.2 worth it over v1?: For quality, unambiguously. It's the best open-source image-to-3D output there is. For volume, no: v1 is 13× faster. The strong workflow is both: draft with v1, finalize the shortlist with v2.
Should I rent or buy for it?: Occasional use: rent, an H200 hour yields ~66 max-quality assets. Pipeline use: this model is the best case in our data for owning 32GB flagship consumer silicon, since the max-quality floor clears 24GB cards out of the running entirely.