MeshAxiom-Base

MeshAxiom: An Agent-Native 3D Foundation Stack for Programmable Worlds

Project · Code

Public Base weights for Exact image-to-native-mesh inference. This repository contains vertex EMA 180K, topology EMA 10K and topology VAE EMA 30K. Install the MeshAxiom source package for the Python API, CLI and agent tools. No optimizer, training data or alternative inference preset is included.

from meshaxiom import MeshAxiomPipeline

with MeshAxiomPipeline.from_pretrained(
    "AlayaLab/MeshAxiom-Base",
    revision="859a522c77e25770619f16e89a57447a0bb9bc16",
) as model:
    result = model.generate("image.png", "outputs/object", face_budget=4000)

The pinned revision is the weight-identity baseline; current validation is recorded in this repository's latest VALIDATION.json. Install MeshAxiom code version 0.2.0.dev1 or later for the Hub-cache and worker-startup fixes. A local model directory can also be used. This repository supplies model data; the source package supplies the inference implementation. model_index.json is descriptive metadata, not a Diffusers pipeline.

Input is a 512×512 RGB image on a black background. The recipe uses MBR8, 20 DPM++2M steps and topology CFG 3. The requested face budget guides generation; actual face counts can differ. OBJ/JSON coordinates are Z-up, and exported GLB files include a transform into glTF's Y-up frame.

Verification and limitations

All 2,128 EMA tensors were preserved and the three models loaded strictly on CPU. Version 0.2.0.dev1 passes 95 CPU tests. Uploaded weight hashes and authenticated sample byte ranges match the verified local files. See WEIGHTS.json, VALIDATION.json and SHA256SUMS for the recorded model identities and validation scope.

Sho Kuno contributed the Exact acceleration. His tested platform is H200 with NVIDIA PyTorch 25.03. This integrated release also passed a bounded H200 check: one image at 4K and 8K requests, with all four geometry files identical to the reference and resident workers reused between requests. This covers only the recorded inputs, seeds, weights and stack. A cold Docker build remains unverified; run meshaxiom verify-exact on your target setup. Geometry can have open boundaries or degenerate faces.

License

The authors' standalone release license is pending; see LICENSE.md. These checkpoints contain frozen DINOv3 tensors subject to external/dinov3/LICENSE.md. The included architecture and processor configurations avoid a separate backbone weight download. Third-party terms and notices are retained.

Paper

The arXiv link and citation will be added when the preprint is available.