Meta 3D Gen

Original Article ↗ Hacker News Discussion ↗

Mesh and topology quality

  • Many commenters focus on topology as the main weakness.
  • Generated meshes are described as “blobby” with poor, unusable wireframes, similar to photogrammetry scans.
  • Automatic retopology/remeshing tools (built into DCC apps or tools like Instant Meshes, TopoGun) are seen as helpful but far from solving production needs, especially for animation/rigging and thin features.
  • Some argue that “fixing topology” is nearly solved in geometry processing; professionals in games/VFX strongly disagree, saying manual retopo is still standard.
  • Meta’s paper is criticized for not addressing topology much; Rodin is cited as a contrasting model that claims “clean topology.”

Textures, detail, and VR constraints

  • Textures are viewed as improved versus older work but still low-quality: blown-out highlights, odd colors, and “uncanny smoothness.”
  • Many note the heavy reliance on fake surface detail via textures/normal maps rather than actual geometry.
  • VR is said to be especially unforgiving: stereoscopic depth reveals goopy low-res geometry quickly, though others counter that current VR games already rely heavily on baked/normal-mapped detail and look fine at typical distances.
  • Displacement mapping is discussed: useful mostly offline; real-time engines mostly rely on normal maps.

Image-to-3D and current tools

  • Meta 3D Gen is primarily text → multi-view images → 3D reconstruction → mesh, not single-image-to-3D.
  • Other systems mentioned: Rodin, AssetGen, Meshy, Luma Genie, Meshroom, MeshAnything.
  • MeshAnything is praised for good low-poly topology but limited (~800 polys), so not yet suitable for high-detail assets.

Practical usefulness for artists and pipelines

  • Several people who tried commercial text/image-to-3D services report results as “unusable garbage” for real pipelines: baked, tangled meshes, awkward UVs, and textures that don’t match normal workflows.
  • For 3D printing/CNC, some think bad topology “doesn’t matter”; others say it absolutely does (holes, flipped normals, non-manifold geometry).
  • Consensus: promising for background props, prototypes, or as rough starting points, but far from “game-ready” or “rig-ready” assets.

Research, hype, and trajectory

  • Multiple comments stress the gap between flashy demos/papers and production reality, but also point out how quickly 2D generation improved, suggesting 3D may follow.
  • There’s enthusiasm about lowering barriers for VR, games, and 3D printing, but skepticism that 3D gen AI will spread as fast or widely as 2D, given 3D’s complexity, format fragmentation, and higher integration costs.