MechBench-36

36 hard-surface, single-object evaluation cases for image- and text-conditioned 3D generation.

Each case is one mechanically complex object — vehicles, robots, instruments, aerospace mechanisms, sci-fi hardware — rendered as a clean studio product shot, together with the text prompts used to condition and to evaluate generation. The suite was built to stress exactly what hard-surface objects punish: open frames, thin supports, coaxial nesting, articulated linkages, repeated and mirrored parts, deliberate asymmetry.

MechBench-36 is the evaluation suite of the paper Procedura: A Procedural Modeling Agent for 3D Shape-as-Code Generation (paper page).

Fields

field	type	description
id	string	case identifier (e.g. assault_buggy)
image	image	the reference render, 1254×1254 RGB, white/light studio background, three-quarter product view
condition_prompt	string	short conditioning prompt describing the object — what an image- or text-conditioned method receives
spec	string	the full evaluation spec: the condition prompt plus what is (and is not) gradeable from the single reference view
reference_generation_prompt	string	the prompt used to generate the reference image itself (provenance / reproducibility)
track	string	benchmark track (object)
difficulty	string	hard, very_hard, or legacy_unrated
domain	string	object domain (e.g. road_vehicle, aerial_robotics, astronomy_instrument)
challenge_tags	list[string]	geometric challenges the case was authored to contain

One split: test (36 rows). There is deliberately no train split — this is an evaluation benchmark.

Usage

from datasets import load_dataset

ds = load_dataset("LoYoT/mechbench-36", split="test")
case = ds[0]
case["image"]              # PIL image, 1254x1254
case["condition_prompt"]   # text conditioning

Typical protocol: condition a method on image (image-to-3D) or on condition_prompt (text-to-3D), and evaluate the produced mesh against the reference view. The spec field states which claims the single reference view can and cannot support, so judges do not grade occluded or prompt-only detail.

Provenance and license

The reference images are AI-generated product renders authored for this benchmark (generation prompts included in reference_generation_prompt); the prompts and specs were written by the authors. No scanned, photographed, or third-party assets are included. Released under CC-BY-4.0.

Citation

@article{procedura2026,
  title  = {Procedura: A Procedural Modeling Agent for 3D Shape-as-Code Generation},
  author = {Lin, Youtian and others},
  journal = {arXiv preprint arXiv:2608.26238},
  year   = {2026}
}