by Stability AI
Research method adding typed 3D constraint meshes (hull, avoidance, touch) as control for text-to-3D generation on TRELLIS.2.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Arbor is a research method for controllable text-to-3D asset generation. It introduces constraint meshes as a native 3D control interface: hull regions mark where geometry should exist, avoidance regions mark space that should stay empty, and touch regions mark surfaces the generated object should contact. These meshes are treated as typed local requirements rather than target shapes.
Technically, Arbor keeps the text-conditioned generator (TRELLIS.2) and its OVoxel geometry encoders frozen. Constraint meshes are fused into a surface signal, encoded into tokens, and routed to local regions of the sparse structure grid. A learned residual geometry branch inside each denoising block attends to this memory; only the geometry projection, positional embedding, summary modules and grounding adapters are trained. YouTube coverage describes a small (~600 MB) model with released code and a Blender add-on, though these details are not confirmed on the project page text provided.
The project page reports evaluation on automatic and artist-curated Toys4K benchmarks against TRELLIS, Gradient, SpaceControl and Spice-E, plus Point-E, SPAR3D and Hunyuan3D-Omni in a variation track. The Arbor family won 59.2 percent of pairwise choices in a 27-participant, 404-trial preference study. It was developed by Jan-Niklas Dihlmann (University of Tübingen) during an internship at Stability AI. It is a research release aimed at 3D authoring workflows.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
3D generation control system that takes constraint meshes specifying where geometry should or shouldn't exist and feeds them to existing generators like Trellis without retraining. Small 600 MB model, code released, usable as a Blender add-on.
Autonomous research system that builds a hypothesis tree with a coordinator and executors, saving evidence across experiments. Outperforms Claude Code and Codex harnesses on several benchmarks; Apache 2.
Closest neighbours by meaning, not just by category.
Model
Research text-to-3D method using block-wise diffusion over shape tokens, about 5x faster than autoregressive baselines.
Model · Tencent
Tencent's unified multimodal model for 3D understanding, text-to-3D generation, instruction-based editing and part generation.
Model
Roblox research model that generates text-to-3D objects as separate part meshes from a user-defined part list.
Model · Hi3D
Open-source image-to-3D and text-to-3D generation model from Hi3D, formerly Sparc3D, trained on ~2M curated 3D assets.
Agent
Open-source multi-agent framework that generates editable CAD parts and assemblies from text prompts using build123d.
Research · Tencent
Tencent Hunyuan research framework that turns one text prompt into an explorable, editable open-world 3D scene.