Vi3W
Skip to content

Guides

VRAM for Hunyuan 3D, TRELLIS and other AI 3D models: a hardware guide

VRAM for Hunyuan 3D, TRELLIS, TRELLIS 2, TripoSR, SF3D, TripoSG and UniRig, plus Mac, low-VRAM tricks and when to go hosted.

2 min read

Three graphics cards side by side, each with a small figurine standing on top
Illustration. More VRAM unlocks bigger models and texture stages.

Quick answer

How much VRAM do I need for Hunyuan 3D?

Per Tencent's READMEs, Hunyuan3D 2.0 needs about 6 GB of VRAM for shape and 16 GB for shape plus texture; Hunyuan3D 2.1 needs 10 GB for shape, 21 GB for texture and 29 GB for both, with a low-VRAM mode. For comparison, TRELLIS needs 16 GB, TRELLIS 2 24 GB, and TripoSR and Stable Fast 3D about 6 GB.

VRAM by model (from each README)

About 6 GB

  • TripoSR (default settings, one image)
  • Stable Fast 3D (default settings, one image)
  • Hunyuan3D 2.0, shape only

About 8 to 10 GB

  • TripoSG (at least 8 GB)
  • UniRig generation (at least 8 GB)
  • Hunyuan3D 2.1, shape only (10 GB)

16 GB

  • TRELLIS (at least 16 GB; verified on A100 and A6000)
  • Hunyuan3D 2.0, shape plus texture

24 GB and up

  • TRELLIS 2 (at least 24 GB; verified on A100 and H100)
  • Hunyuan3D 2.1 texture (21 GB) or shape plus texture (29 GB), or use --low_vram_mode

NVIDIA is the default

Most open 3D models depend on CUDA extensions (custom rasterisers, sparse convolutions, attention kernels). An NVIDIA card with enough VRAM is by far the easiest path. TRELLIS and TRELLIS 2 are tested on Linux only.

Mac and AMD

  • Mac: Hunyuan3D 2.0/2.1 list macOS support; Stable Fast 3D has experimental MPS support and recommends the CPU path under 32 GB unified memory. See Hunyuan 3D on Mac.
  • AMD: possible for some models with ROCm builds of PyTorch, but custom CUDA extensions are the usual blocker.
Three graphics cards with increasingly detailed statues above them
Illustration. Match the model to the card's VRAM, not the other way round.

Squeezing into less VRAM

  • Run shape only, texture later or elsewhere.
  • Use smaller variants: Hunyuan3D 2mini, TRELLIS instead of TRELLIS 2.
  • Use low-VRAM flags and quantised (GGUF) community workflows in ComfyUI, accepting slower or lower-quality output.
  • Close Blender, browsers and games while generating.

System RAM, disk and CPU

  • RAM: 32 GB is comfortable for these pipelines.
  • Disk: model weights take many gigabytes each; keep an SSD with room.
  • CPU: matters little once the GPU does the work.

Blender on the same machine

Blender's Cycles renders on NVIDIA (CUDA/OptiX), AMD (HIP), Intel Arc (oneAPI) and Apple (Metal). See Blender system requirements.

When a hosted tool is cheaper

A 24 GB card is a big purchase for occasional use. Vi3W runs TRELLIS, TRELLIS 2, Hunyuan 3D and commercial models on its own servers, so you need only a browser. Free plan included. Try Vi3W.

Frequently asked questions

hunyuan 3d local requirements

A CUDA-capable NVIDIA GPU is the smoothest path: 6 GB for Hunyuan3D 2.0 shape, 16 GB with texture, up to 29 GB for 2.1 with texture. macOS is listed as supported but is much slower.

Last reviewed