Introduction

AlayaLab just open-weighted MeshAxiom-Base: you give it one image and a requested face count, and it builds a native triangle mesh with that budget baked into generation — not bolted on afterward with a remesher. Code is a public release candidate (0.2.0.dev1), weights live on Hugging Face, and a 123-asset GLB kit is already browsable in the browser. The authors' standalone license is still pending, and commercial clearance is not there yet.

That face-budget-first angle is the whole pitch. Dense image-to-3D pipelines (TRELLIS.2 and friends) give you lush geometry you then carve down. MeshAxiom aims for the low-poly topology game artists usually spend evenings fixing by hand.

What shipped

  • Artifact: MeshAxiom-Base on Hugging Face (vertex EMA 180K, topology EMA 10K, topology VAE EMA 30K) plus Exact inference code in AlayaLab/MeshAxiom
  • Project page: alaya-lab.github.io/MeshAxiom with a Space Saga trailer and 123-asset library (input, gray mesh, topology, materials, GLB download)
  • Modality: image-to-3D. Input is a 512×512 RGB image on a black background plus a face budget; outputs are mesh.obj, mesh.json, mesh.glb (Z-up → glTF Y-up transform), vertices.npz, topology.npz
  • Open vs API: public weights and code, no hosted API. License pending — see LICENSE.md. Checkpoints embed frozen DINOv3 tensors under DINOv3 terms. Treat this as research weights until the authors assign a release license
  • Where to run today: self-host on Linux + CUDA. Python API, meshaxiom generate, or meshaxiom serve for a local upload page at 127.0.0.1:7860. Author-tested stack: NVIDIA PyTorch 25.03, CUDA 12.8, Python 3.12, H200. Weights total about 21 GB (11.24 + 9.42 + 0.45 GB) — plan for a big-VRAM GPU; no consumer VRAM figure is published. No hosted Space yet

How the mesh is made

Two diffusion transformers share the job. One grows vertices coarse-to-fine over a sparse octree. The second predicts connectivity in the latent space of a topology autoencoder. The face budget is a generation condition, not a post-hoc triangle cull — README examples are 2K / 4K / 8K / 16K requests, and realized counts can differ.

Default recipe on the card: MBR8, 20 DPM++2M steps, topology CFG 3. Pin the verified Hub revision from the GitHub README (b80240ce…) with code ≥ 0.2.0.dev1, not the older pin on the model card.

from meshaxiom import MeshAxiomPipeline

with MeshAxiomPipeline.from_pretrained(
    "AlayaLab/MeshAxiom-Base",
    revision="b80240ce4f77a28bc4debdcca895e96c78ddb294",
) as pipeline:
    result = pipeline.generate("image.png", "outputs/object", face_budget=4000)

python -m pip install -e '.[geometry,hub,background,postprocess]'
meshaxiom generate --config models/MeshAxiom-Base/meshaxiom.json \
  --image image.png --face-budget 4000 --output outputs/object
meshaxiom serve --config models/MeshAxiom-Base/meshaxiom.json --workspace outputs/server

Run meshaxiom verify-exact on your own box before you trust bit-identical geometry. The authors report byte-identical meshes between reference and Exact paths on one H200 canary (4K and 8K), plus 95 CPU tests and roughly 1.6× median Exact speedup on 400 author-side 4K requests — evidence, not a new public benchmark.

Agent-native, geometry-first

The stack is built so agents call file tools — generate_mesh, plus separate texture_mesh / enhance_texture material helpers — over JSON or a local HTTP API and keep compact summaries plus rendered views for later edits. Base only ships geometry and connectivity. UVs, materials, rigs, scene assembly and choreography sit in a workflow the authors say will follow; training code and the paper are not out yet.

Space Saga shows twelve authored scenes (stations, fleets, shipyards, moon bases, encounters) with Materials / Gray mesh / Topology views. Showcase scenes use selected generated meshes plus bounded repairs and external materials — honest about what the raw model vs the authored demo includes.

Caveats worth reading twice

  • Geometry can have open boundaries or degenerate faces
  • Cold Docker build is unverified
  • License pending — not cleared for commercial shipping
  • Validated on H200 only; other GPUs need your own verify-exact
  • Traction is fresh: public weights landed overnight with almost no stars yet — substance is the ~21 GB EMA payload and the runnable RC, not social proof

If you already use Espalier-2 or Effigy for dense TRELLIS.2 pipelines, MeshAxiom is a different bet: ask for the polycount up front and get explicit triangles you can inspect in an agent loop.

Original Source

https://huggingface.co/AlayaLab/MeshAxiom-Base

https://github.com/AlayaLab/MeshAxiom

https://alaya-lab.github.io/MeshAxiom/

Conclusion

MeshAxiom-Base is a rare day-0 drop that treats face budget as a first-class generation input for game-shaped meshes, with real weights, Exact inference code, and a browsable asset kit — while staying honest that license, paper and full workflow tools are still coming. Clone the RC, pin the verified revision, verify on your GPU, and keep commercial use on hold until LICENSE.md says otherwise.

—Titus