TRELLIS.2 4B

TRELLIS.2-4B is Microsoft's 4-billion-parameter image-to-3D model (Tech report, 2025). It uses a flow-matching transformer on top of a sparse 3D VAE to convert a single image into a textured GLB mesh with PBR materials — geometry, base color, roughness, metallic, normal, and opacity all aligned. MIT-licensed.

Yes — TRELLIS.2-4B is our default image-to-3D backend at /image/3d/ and is included in the free daily token allowance. Each generation costs about 5,000 tokens (well within the free 30K/day for signed-in users).

TRELLIS v1 (2024) reconstructed meshes via Gaussian Splats + iso-surface fields. TRELLIS.2 (2025) introduces O-Voxels — an omni-voxel structure that natively handles arbitrary topology (open surfaces, non-manifold geometry, full PBR including translucency) and runs ~5-10× faster end-to-end. The output is a single textured GLB instead of a multi-asset bundle.

A single still image. Photo, render, sketch, or AI-generated image all work. Best results: object centered, lit head-on, on a plain background. The model is not currently tuned for multi-view input.

GLB with WebP-compressed PBR textures (geometry + base color + normals + roughness + metallic + opacity). Decimated to ~1M faces by default with a 2048×2048 texture atlas. GLB opens natively in any modern browser, Google AR Quick Look, Blender, Unity, and Unreal.

On Free.ai's H200 GPU: ~3 seconds at 512³ voxel resolution, ~8-17 seconds at 1024³ (our default), and ~60 seconds at 1536³ (hero-asset quality). Cold-loading the model adds ~20 seconds.

TRELLIS.2-4B fits in ~24 GB resident VRAM and transiently peaks higher during the VAE decode + texture-bake stages. Free.ai runs it on a single H200 (141 GB) shared with other heavy tools via a cross-tenant semaphore that caps concurrent heavy gens.

Yes — fully MIT-licensed by Microsoft. No commercial restrictions, no territorial carve-outs, no monthly-active-user caps. The meshes you generate are yours to ship, sell, modify, and 3D-print.

Yes, with the caveat that the raw mesh may have small holes or topological discontinuities (Microsoft documents this in the model card). Free.ai's wrapper runs a best-effort hole-fill pass on every export, but for critical prints, run the GLB through Blender's 3D-print toolbox or Meshmixer's "Make Solid" first.

TRELLIS.2 produces dramatically better meshes — textured PBR materials vs untextured geometry, watertight-aware topology vs hole-prone, arbitrary surface types vs closed-only — at higher token cost (~5K vs ~2K) and similar wall-clock time on modern GPUs. We keep TripoSR as a fast-fallback option; new traffic defaults to TRELLIS.2.

Default settings target a 1M-face decimation with 2048-px textures. Power users can pass `decimation_target` and `texture_size` to the API for higher-fidelity hero assets (4096-px textures, fewer-decimated meshes) — bumps generation time and token cost slightly.

Yes — POST multipart to /v1/3d/generate/ with `image` (file) and `model=trellis2-4b`. Returns JSON with `file_url` pointing at the GLB. Curl + Python + Node snippets at /api/. X-User-Id required for authenticated quotas.