Introduction

Espalier-2 is an open MIT toolkit from over|yonder that runs Microsoft's TRELLIS.2 image-to-3D model at its highest setting — the 1536 cascade — on a 16 GB consumer GPU. Microsoft lists 24 GB as the minimum, and the stock 1536 path runs out of memory in the shape decoder and again in the GLB export. Espalier-2 tiles those stages and splits generation from export so the same weights and settings finish on an RTX 5070 Ti. The repo landed on GitHub on October 14 (create 21:25 UTC), with a measured write-up on over-yonder.tech dated October 15. Modality: single image → textured mesh (GLB). Open vs API: local open pipeline (MIT patches + Microsoft MIT weights). Where to run it today: Linux + NVIDIA ≥16 GB, CUDA 12.8+, then ./install.sh and ./espalier input.png out/.

What it unlocks

TRELLIS.2 already ships open weights and four pipelines (512, 1024, 1024_cascade, 1536_cascade). The top cascade builds on a 1536-voxel grid and is the quality setting most indie 3D / game-asset makers want — and the one that OOMs on 16 GB cards. Espalier-2 does not retrain or swap weights. It patches three places in TRELLIS.2 75fbf01 and CuMesh 12289e1, then runs generation and export as separate processes:

  • Shape and texture decoders: the two highest-resolution stages run in overlapping spatial tiles (default tile edge 256 at 1536).
  • Narrow-band remesh: connectivity and triangle steps chunk over voxels/quads; output bit-identical to stock.
  • GLB export: material volume stays in system RAM until bake; old mesh released before the remeshed one is built.

On Microsoft's published catapult input, with the desktop already using 2.1 GiB of the card, driver peaks are 11.9 GiB for generation (5 minutes) and 13.3 GiB for export (51–60 seconds). Microsoft quotes about 60 seconds end-to-end on an H100. Texture bake defaults to 4096 and face target 1,000,000.

Install and run

Needs Linux, an NVIDIA GPU with 16 GB, CUDA toolkit 12.8 or later (GCC that toolkit supports), git, and uv. Tested on RTX 5070 Ti (Blackwell, compute capability 12.0). install.sh clones TRELLIS.2, CuMesh, FlexGEMM, nvdiffrast and nvdiffrec at pinned commits, applies the patches, and builds CUDA extensions into .venv for your card (PyTorch 2.8 + CUDA 12.8, flash-attn 2.8.3).

git clone https://github.com/overyonder/espalier
cd espalier
./install.sh
./espalier input.png out/

That writes out/model.glb plus out/raw.pt. Useful flags: --pipeline 1536_cascade (default), --seed 42, --tile 256, --texture 4096, --faces 1000000, --no-rembg if the input already has alpha.

Two HF models TRELLIS.2 loads are gated: DINOv3 ViT-L/16 and RMBG-2.0. Request access, then hf auth login. RMBG-2.0 is non-commercial and only needed for opaque inputs.

Pro Tip

Start with Microsoft's own catapult image (ships with TRELLIS.2) and seed 42 before you trust a hero prop. If something OOMs, check the desktop is not sitting on 2+ GiB of the card, and try --tile larger so FlexGEMM's kernels behave more like the full-object path.

Why game makers should care

A lot of local image-to-3D stacks quietly drop to 512 or 1024 on consumer cards. Espalier-2 keeps the 1536 cascade — the setting on Microsoft's own gallery quality bar — available on a single 16 GB GPU, and it still exports a textured GLB you can drop into Blender, Godot, Unity or Unreal. For prop and kitbash pipelines that already assume TRELLIS.2, this is the missing low-VRAM path rather than a new model family. HF card: overyonder/espalier-2 (points at the GitHub patches; no re-hosted weights).

Caveats

  • 0★ day-zero drop: no vendor X thread yet; verify the install on your driver before you schedule overnight batch jobs.
  • Linux + NVIDIA only in the README; Apple Silicon / Windows are not claimed.
  • DINOv3 and RMBG licences sit beside MIT Espalier / TRELLIS.2 — especially RMBG's non-commercial term if you strip backgrounds inside the tool.
  • Tiling at 1024 shows tiny floating-point drift vs untiled (largest feature diff ~0.016 on a 0–1 scale); 1536 uses large tiles to stay closer to stock behaviour.
  • Generation is slower on 16 GB (5 min) than on H100 (60 s) — the win is fitting, not beating datacenter latency.

Conclusion

Espalier-2 is a measured, MIT-licensed way to run TRELLIS.2's top 1536 cascade on a 16 GB card instead of the 24 GB floor Microsoft lists. If you already generate game-ready props with TRELLIS.2 and your workstation tops out at 16 GB, clone the repo, run install.sh, and bake one GLB before you change any defaults.

—Titus