LL3M: Large Language 3D Modelers

Perceived Usefulness and Current Capabilities

• Many see LL3M as a “cute” but impressive early-stage tool: fun toy, already usable for rough props, prototypes, Roblox‑style games, or as a starting point to edit in Blender.
• It fits into broader workflows where LLMs script tools (Blender, FreeCAD, OpenSCAD, Aseprite, etc.) or where image→3D tools (e.g. meshy.ai) provide a base mesh that artists refine.
• High-poly, messy topology makes these assets unsuitable for production games or animation, but potentially fine for quick visualization or communicating ideas to a 3D artist.

Skepticism from Experienced 3D Artists

• Experienced Blender users argue the showcased models are trivial; with a day or two of tutorials most technically inclined people could make better results directly, while gaining real skills.
• Critiques: bland output, bad topology, excessive polygon counts, no attention to constraints like 3D printability or performance; risk of people using AI instead of learning fundamentals.
• Some stress that LLMs are text models; the real work for high-quality 3D will need specialized geometry/vision models, not “Blender via Python” alone.

Accessibility vs. Craft and “Gatekeeping”

• Non‑artists and those who have repeatedly failed to learn 3D (or lack strong visualization ability) find this kind of tool “insanely useful” just to get a passable dog model or simple game assets.
• Others push back that wanting results without learning the craft should not be conflated with genuine creative expertise; AI may lower entry barriers but won’t replace deep skill.
• This leads to accusations in both directions: “shitty gatekeeping” vs. “shitty optimism” and hand‑wavy “it’ll get better” arguments.

Future Directions and Architectures

• Strong interest in using AI as assistive tooling for tedious steps: retopology, UVs, rigging, auto‑constraints, shader wiring, asset search, and geometry‑nodes boilerplate.
• The paper’s multi‑agent approach (planner, coder, critic, visual checker, BlenderRAG, etc.) is seen as a promising pattern: orchestrated specialists rather than a single monolithic LLM.
• Some speculate this style of modular, multimodal system is closer to eventual AGI, and that everything (including geometry) will become just another token space.
• Others warn against over‑extrapolating from current “low‑hanging fruit,” pointing to previous tech hype cycles and uncertain progress beyond current plateaus.