TRELLIS.2 is Microsoft Research's 4-billion-parameter image-to-3D generative model, released under **MIT** and accompanied by the paper *[Native and Compact Structured Latents for 3D Generation](https://arxiv.org/abs/2512.14692)*. The repository has accumulated roughly **9,950 stars and 1,201 forks**, and the accompanying checkpoint [`microsoft/TRELLIS.2-4B`](https://huggingface.co/microsoft/TRELLIS.2-4B) has been downloaded about **1.44 million times** on Hugging Face with 1,063 likes — figures that put it among the most widely pulled open 3D generation models available. ## The Core Idea: O-Voxel Most 3D generative models represent geometry as an implicit field — an SDF or occupancy grid — and then run marching cubes to extract a mesh. That pipeline carries a structural cost: iso-surface extraction can only produce closed, manifold, watertight surfaces. A shirt, a leaf, a wire frame, or a bottle with liquid inside all have to be approximated as sealed solids, and the approximation is where fidelity leaks away. TRELLIS.2 replaces the field with **O-Voxel**, described in the paper as a *field-free* sparse voxel structure. Because it never passes through an iso-surface, it handles categories that field-based methods structurally cannot: - **Open surfaces** — clothing, foliage, thin shells - **Non-manifold geometry** — edges shared by more than two faces - **Internal enclosed structures** — geometry sealed inside an outer shell A Sparse 3D VAE with **16× spatial downsampling** compresses assets into the latent space, which is what keeps a 4B model tractable at high output resolutions. ## Speed and Resolution The README publishes a resolution/latency table measured on an NVIDIA H100, split between shape and material generation: | Resolution | Total time | Shape + Material | |---|---|---| | 512³ | ~3s | 2s + 1s | | 1024³ | ~17s | 10s + 7s | | 1536³ | ~60s | 35s + 25s | Three seconds for a textured 512³ asset is interactive-adjacent, and 1536³ in a minute puts high-detail generation inside a normal iteration loop rather than an overnight batch. The conversion utilities are equally lean and, importantly, **rendering-free and optimization-free**: under 10 seconds on a single CPU to turn a textured mesh into O-Voxel, and under 100 ms on CUDA to go back the other way. ## PBR Materials, Not Just Vertex Color The model predicts arbitrary surface attributes rather than baked color — **base color, roughness, metallic, and opacity**. That is the difference between an asset that looks acceptable in a turntable preview and one that survives being dropped into a real renderer with an HDRI environment map. The example pipeline loads an EXR environment map, runs the image through `Trellis2ImageTo3DPipeline`, and exports a PBR-ready GLB with a 4096px texture and a 1M-triangle decimation target: ```python from trellis2.pipelines import Trellis2ImageTo3DPipeline pipeline = Trellis2ImageTo3DPipeline.from_pretrained("microsoft/TRELLIS.2-4B") pipeline.cuda() mesh = pipeline.run(Image.open("input.png"))[0] ``` A second entry point handles **shape-conditioned texture generation** — feeding in existing geometry and generating PBR textures for it — which is the workflow most studios actually need, since they rarely want the model inventing the topology. ## Complete, Including Training Every item on the project roadmap is checked off: paper, image-to-3D inference code, the 4B checkpoint, a Hugging Face Spaces demo, shape-conditioned texturing, and — the item most releases quietly omit — **the full training codebase**. `train.py` supports multi-node distributed runs with `--num_nodes`, `--node_rank`, and `--master_addr`, plus checkpoint resumption and a `--tryrun` dry-run mode, and a separate `data_toolkit` documents converting raw 3D assets into the O-Voxel representation. Fine-tuning on a proprietary asset library is a documented path here, not a reverse-engineering exercise. ## What It Costs to Run The hardware and platform requirements are strict and worth reading before cloning. The code is **tested only on Linux**. An NVIDIA GPU with **at least 24 GB of VRAM** is necessary, and verification was done on A100 and H100 class hardware — so the published latency numbers describe datacenter GPUs, not a workstation card. CUDA Toolkit 12.4 is the recommended build target. Installation is the heaviest part of the experience. `setup.sh` compiles a stack of custom and third-party CUDA extensions — `flash-attn`, `nvdiffrast`, `nvdiffrec`, `cumesh`, `o-voxel`, and `flexgemm` — and the README itself warns the process "may take a while" and suggests installing flags one at a time if it fails. GPUs without flash-attn support need a manual `xformers` install and an `ATTN_BACKEND` override. Smaller friction points: exported GLB files default to `OPAQUE` mode, so the alpha channel is present in the texture but must be manually wired to the material's opacity input in your DCC tool before transparency works. The repository carries **143 open issues**, and the most recent commits date to June 2026 — the current star surge appears to be driven by adoption rather than fresh releases. ## Verdict TRELLIS.2's contribution is specific and well-argued: dropping the iso-surface field removes a class of failure that image-to-3D systems have been working around rather than solving, and doing it while shipping PBR materials, sub-minute generation, MIT licensing, and complete training code is an unusually generous package from a corporate research lab. The barrier is hardware — a 24 GB NVIDIA card on Linux and a demanding compile — so the practical audience is technical artists and pipeline engineers with datacenter or high-end workstation GPUs. Anyone who just wants to try the model should start at the Hugging Face Spaces demo; anyone building an asset pipeline around open 3D generation should treat this as the current reference implementation.