Running Hunyuan3D and TRELLIS Locally: What the Repos Require and How to Check

Documented requirements

TRELLIS.2 listing 24 GB as “at least” is the minimum, not a recommendation to buy more. All of these are vendor figures for their tested setups.

Where sources conflict

• VRAM, third-party claim. 3D AI Studio’s cost guide says “If you have an NVIDIA GPU with 8+ GB of VRAM, you can generate 3D models locally for free”. The sentence is a general local-generation figure and names no model; the article lists TripoSR, InstantMesh and TRELLIS as local options and gives no per-model VRAM figure or TRELLIS version. It is not a TRELLIS requirement, and the official READMEs say 16 GB for v1 and 24 GB for TRELLIS.2. Go by the READMEs.

• Speed. The TRELLIS.2 README gives about 3 seconds at 512³, 17 seconds at 1024³ and 60 seconds at 1536³, shape plus material, “Tested on NVIDIA H100 GPU”. trellis2.app (the authors say they operate a hosted TRELLIS.2 service) repeats those figures and gives roughly 10 to 30 seconds for Hunyuan 3D 2.x, labeled vendor-reported with hardware varying. It lists official repositories and release announcements as its sources, not runs of its own. 3D AI Studio’s cost guide quotes results “in under 2 minutes” on a cloud platform, a different measurement altogether. No consumer-GPU timing appears in any source we read.

• ComfyUI. The Hunyuan3D-2.1 README lists a community ComfyUI project under community contributions and makes no official ComfyUI statement; the TRELLIS v1 README does not mention ComfyUI.

• Licenses. TRELLIS.2 is MIT but depends on nvdiffrast and nvdiffrec under their own licenses. Hunyuan3D 2.1’s community license excludes the European Union, United Kingdom and South Korea from its territory and has a monthly-active-user clause. Check the open-weights licenses page before shipping output commercially. This is not legal advice.

• Community reports. The TRELLIS.2 issue tracker lists an issue titled “OOM on smaller GPUs” (#188, opened 2026-09-07). We read only the issue list, not the thread, so we cannot say what GPU it concerns or whether it was resolved. Read it before buying hardware. Its existence fits the README’s 24 GB minimum, but it is not a measurement. Two other titles are worth opening for the same reason: “Working environment setup for RTX 5090 (32GB VRAM) & WSL-Ubuntu” (#143) and “[Question] low_vram hardware requirement + OOM on A100 80GB when low_vram is disabled” (#55). Again, we read the titles only.

What a consumer card changes

Neither project publishes a consumer-GPU figure. TRELLIS.2’s README states verification on A100 and H100 data-center cards, TRELLIS v1 on A100 and A6000. A 16 GB or 24 GB gaming card is therefore outside the verified hardware, even when the VRAM number matches. For Hunyuan3D-2.1 the 10, 21 and 29 GB figures are separate for shape, texture and both, which means a 12 GB card might run shape only; whether you can get textures from a separate pass or another tool is a workflow decision, not a documented feature. If your card is below the documented minimum, the cheaper test is a hosted run before you build an install. Tencent Cloud sells a hosted Hunyuan 3D API with free starter credits; its billing page blocked our last automated check (November 2026), so read the current credit figures there. A hosted service is a separate product and may not run the same version as the open weights you would run locally.

How to check on your machine

1. Write down GPU model, VRAM, driver and CUDA version, OS (native Linux, WSL or Windows), and Python version. The TRELLIS READMEs say the code is tested only on Linux (the v1 README points to a community Windows thread it calls “not fully tested”), so a Windows run is outside the documented envelope; record it as such.
2. Time the install from a clean checkout to first successful sample, and write down every step that failed. Include dependency builds (the Hunyuan3D-2.1 repo has a custom rasterizer build).
3. While generating, watch peak memory with nvidia-smi sampled at one-second intervals (for example nvidia-smi --query-gpu=memory.used --format=csv -l 1), and keep the maximum.
4. Run a subset of the standard prompt set (image-input cases for these image-to-3D models; for text you must first make an image) and log generation seconds per output.
5. Run the mesh script on each output and compare with a hosted run if you have one, noting whether the hosted version matches yours.

Fields: gpu, vram_gb, os, install_minutes, install_failures, peak_vram_gb, generation_seconds, resolution_setting, completed (yes/out of memory/other error), triangles, loose_pieces, open_edges.

Ambiguous results

Rules to set before you start, for borderline readings; they are editorial, not outcomes we have seen.

• Runs, but only with --low_vram_mode or lower resolution: record “completed with flag”, the setting changed, and the extra seconds. Whether quality dropped is a mesh question; compare triangles and open edges against a hosted run of a matching version, not by eye.
• Peak memory sampled at one second may miss a spike: if you see an out-of-memory error with a sampled peak below your card’s size, trust the error and log the sample as a lower bound.
• Succeeds once, fails on the second image: count it as a failure at that resolution; memory use can depend on the input.
• Install works only on WSL or Windows: pass for you, but outside what the TRELLIS READMEs say was tested, so say so on any write-up.
• Output differs between runs: that is variance, not install failure; see the seed protocol.

Pass or fail

Pass for your hardware means: installs in a time you accept, completes the prompt subset without out-of-memory errors at the resolution you want, and produces meshes that meet your gate. Peak VRAM above what you have is a fail even if a flag like --low_vram_mode rescues it, but record what the flag costs in time and quality. Compare the hosted route too: best open-source AI 3D generators, Hunyuan3D vs TRELLIS.2, and reviews of Hunyuan3D and TRELLIS.2.