Image-to-3D generation has a credibility problem: demos look great, but the exported mesh often isn't something you can actually open in Blender. Trellis 2 takes a different angle — it's a browser-based generator built on Microsoft Research's SLAT (Structured LATent) architecture, and the entire pipeline ends in a real GLB export you can drop straight into Blender, Unity, or Unreal.

The research under the hood

Trellis 2 is built on Microsoft Research's SLAT representation, scaled up to 2 billion parameters trained on 500K+ diverse 3D objects. According to Microsoft Research, SLAT enables 3D generation with exceptional geometric detail and material fidelity, supporting both 3D Gaussians (efficient rendering) and meshes (direct import into standard 3D software). Trellis 2 packages that research as a usable product: upload an image, get a previewable model, export.

The workflow

  1. Upload a single image (JPG/PNG/WebP) — single-image or multi-view.
  2. Configure: output resolution (1024p balanced default), texture quality (2K standard), mesh quality, optional mesh optimization.
  3. Generate — a usable, previewable 3D model in 1–3 minutes.
  4. Export GLB instantly, with cloud storage included.

The headline constraint that matters: no GPU required. The pipeline runs in the cloud, so a laptop without a discrete GPU produces the same result as a workstation. For solo devs and hobbyists, that removes the usual "I'd need a 4090 to run this locally" wall.

Styles and use cases

Trellis 2 offers 25+ 3D styles — photorealistic, cartoon, anime, clay, low-poly — which changes who it's for:

  • Game developers: environment dressing and props in low-poly or stylized finishes, exported as GLB.
  • 3D printing: photo-to-mesh for miniatures and decor.
  • E-commerce: turn product photos into previewable 3D for listings.
  • Architecture: quick massing models from reference images.

There's currently a launch sale (20% off credit packages), and generation starts at 2 credits per model, so evaluating the pipeline is cheap.

Honest expectations

Image-to-3D is still image-to-3D: flat-on photos of objects work best; occluded geometry gets invented. Start with a clean, well-lit single-object photo, check the preview before exporting, and treat the GLB as a strong first draft rather than a final asset. With that framing, the 1–3 minute loop is genuinely useful — especially as a sketching tool for game jams and concept validation.

→ Try it with one of your own photos: https://trellis2.app/

Disclosure: independent write-up based on the product's public documentation; not sponsored.