PaperlayerView exact revision on arXiv

Source-linked AI summary

VisCAD: A Foundation Model Suite with Multimodal Industrial CAD Intelligence

Guanlin Li, Zhichao Huang, Huimu Yu, Yichen Long, Hongsen Liu, Linxin Cai, Qiuhe Hong, Ziqi Liu, Luya Wang, Wenxiang Wu, Ning Zhang, Yuchen Wang

arXiv:2609.03811v2cs.CVcs.CL

TL;DR

Specialized AI CAD models often generalize poorly beyond their training domains. VisCAD pairs multimodal part generation with a CAD-native assembly harness; VisCAD-M1 leads evaluated part-level systems on two benchmarks, with further gains from parallel test-time scaling.

  • Problem

    Specialized AI CAD models often generalize poorly beyond their training domains.

    Source: §1 · p. 2

  • Method

    VisCAD pairs 27B multimodal VisCAD-M1 for intent-to-program part generation with VisCAD-H1, a structured, CAD-native assembly harness.

    Source: §1 · p. 2, §3 · p. 3

  • Results

    Across PubCADBench and RealCADBench, VisCAD-M1's part-level profile average is 0.5540 versus 0.5496 for the strongest frontier model; parallel TTS raises it to 0.5797.

    Source: §1 · p. 2, §5.2 · p. 7, §7 · p. 9

  • Takeaways & Limitations

    The suite demonstrates that domain-specific training can support broad part generation, while its CAD-native harness can extend frontier CAD intelligence to complex assemblies and multiple CAD backends.

    Source: §1 · p. 2

Abstract

from arXiv · show

AI-assisted computer-aided design (CAD) for industrial products involves two challenging phases. Part-level generation maps diverse forms of user intent, including renders, text descriptions, 2D drawings, and real photographs, to executable programs in a CAD domain-specific language. Assembly-level generation must additionally handle interacting parts, plan mating relations, estimate poses, and place all parts correctly. Existing specialized CAD models are commonly trained on narrow input domains, such as renders or texts, and often generalize poorly, while general-purpose frontier models cover broader inputs but perform inconsistently across CAD domains. We present VisCAD, a foundation model suite designed to provide both broad generalization and strong CAD capability for realistic industrial products. At its core is VisCAD-M1, a 27B model trained through mid-training and post-training for part-level design generation. On PubCADBench and RealCADBench, VisCAD-M1 achieves the highest average part-level score among the evaluated models, reaching 0.5540 compared with 0.5496 for the strongest frontier model. Reusing VisCAD-M1 as a test-time verifier can further raise the score to 0.5797, an approximately 5 percent relative improvement over the previous state of the art. VisCAD also includes a domain-specific harness that leverages frontier models for complex assembly generation and demonstrates advantages over general-purpose harnesses in both quantitative and qualitative evaluations.

1 INTRODUCTION

Specialized AI CAD models often generalize poorly beyond their training domains, while industrial design requires diverse intent modalities and assembly-level reasoning. VisCAD combines multimodal part generation with a structured assembly harness.

1 INTRODUCTION

  • VisCAD-M1 maps text, engineering drawings, product photographs, and rendered views to executable FreeCAD Python programs, while VisCAD-H1 decomposes assembly intent into parts, mating relations, and placements.Its DSL-agnostic intermediate representation separates assembly reasoning from backend syntax.Source: §1 · p. 2, Figure 1 · p. 2
  • 0.5540 average profile score exceeds the strongest compared frontier model’s 0.5496; test-time scaling raises it to 0.5797, while VisCAD-H1 scores approximately 85.0 versus Codex’s 68.0 and Claude Code’s 52.0.The assembly comparison is from a 50-instance study.Source: §1 · p. 2
  • Parallel test-time scaling reranks several VisCAD-M1 rollouts for the same intent, improving its prediction by 2.7 points and by more than 5% relative to frontier SOTA.Source: §1 · p. 2
  • Curation uses image-editing and image-to-3D tools for raw-data processing, augmentation, and recursive program search, while multi-stage training uses CAD data of varying amounts and quality.Source: §1 · p. 2

2 RELATED WORK

Prior learning-based CAD work has advanced program-native generation and broadened input modalities, but remains largely part-centric. This section also distinguishes benchmark scopes and assembly-harness evaluation from sampled part-program evaluation.

2 RELATED WORK

  • Program-native CAD methods recover parametric command sequences or code, but remain part-centric, usually generating one solid from CAD-native or synthetic supervision.Source: §2 · p. 3
  • A second line broadens inputs to drawings, sketches, and mixed visual-language evidence, yet engineering drawings and product photographs remain underrepresented relative to CAD-native renders.The output is still typically one part rather than a verified assembly.Source: §2 · p. 3
  • PubCADBench aggregates public program-generation datasets, while RealCADBench measures real-world factory-automation intents using executability, geometric, and visual-semantic scores.Source: §2 · p. 3
  • A harnessed assembly loop and a sampled part program are different evaluation objects; VisCAD-H1 is compared with coding agents as a delivered system.General coding harnesses are not built around structured bills of materials, per-part IR gating, or mating review.Source: §2 · p. 3

3 OVERVIEW OF VisCAD AND BASICS

VisCAD combines data and evaluation infrastructure with a part generator and an assembly harness. It frames CAD generation as mapping multimodal design intent to executable programs and verified shapes.

3 OVERVIEW OF VisCAD AND BASICS

  • VisCAD organizes infrastructure, part-level generation, and assembly-level generation into one suite.Its infrastructure includes data curation and evaluation; VisCAD-M1 generates parts, while VisCAD-H1 supports assembly planning, positioning, and verification.Source: §3 · p. 3
  • The shared abstraction represents CAD data as intent, program, and 3D-shape triples, allowing entries to be incomplete.The abstraction connects intent-to-program mapping with code verification and shape-based visual verification.Source: §3 · p. 4
  • VisCAD-M1 maps text, drawings, renders, or product photographs to executable FreeCAD Python for single-part design generation.The program is intended to run and export a 3D model.Source: §3 · p. 4
  • VisCAD-H1 uses an agentic workflow and a DSL-agnostic intermediate representation to turn multi-part intent into assemblies.The IR supports stable modeling across CAD environments.Source: §3 · p. 4

4 VisCAD: DATA CURATION, MODEL TRAINING AND HARNESS DESIGN

The suite combines heterogeneous CAD data curation, staged training for VisCAD-M1, and a domain-specific assembly harness. Its components organize diverse inputs and CAD tasks around executable programs and structured assembly plans.

4.1 DATA CURATION PIPELINE

  • Data curation unifies public CAD programs and renders, catalog photographs, and industrial assets and drawings as intent-program-shape data.A two-way pipeline completes missing elements: intent-led data can be normalized and reconstructed before program search, while shapes or programs can generate diverse intents and supervised programs.Source: §4.1 · pp. 4–5

4.2 MULTI-STAGE TRAINING OF VisCAD-M1

  • VisCAD-M1 training progresses from executable CAD priors to industrial visual coverage and then fine-grained geometric fidelity.Mid-training uses nearly 1M executed CAD programs; robust SFT uses roughly 150k product images; continual high-quality SFT uses roughly 10k curated examples in a generate–verify–feedback loop.Source: §4.2 · p. 5

4.3 THE DESIGN OF VisCAD-H1

  • VisCAD-H1 structures assembly generation through gated stages, from multimodal understanding and per-part IRs to global review and CAD realization.Per-part representations are checked before global review; the reviewed IR includes poses and is translated into backend modeling calls to build and assemble solids.Source: §4.3 · pp. 5–6, Figure 4 · p. 5
  • The DSL-agnostic IR separates assembly representation from CAD syntax, supporting translation of the same plan to different backends.The harness retains assembly reasoning in the IR while CAD realization uses backend-specific calls.Source: §4.3 · p. 6

5 EXPERIMENTS AND MAIN PERFORMANCES

Across PubCADBench and RealCADBench, VisCAD-M1 leads the reported part-level profile average, with test-time scaling raising its score further; VisCAD-H1 also scores above the compared harnesses in a 50-instance assembly study. Slice-level results show that geometry and judge scores do not always align.

5.1 SETUP

  • The assembly comparison uses a 50-instance subset, not the full assembly tracks, and its scores are not compute-matched to Table 1 or each other.The part-level sets total 1,100 PubCADBench tasks and 1,745 RealCADBench tasks.Source: §5.1 · p. 6

5.2 PART-LEVEL RESULTS

  • 0.5540 overall profile average for VisCAD-M1 exceeds Gemini-3.1-Pro at 0.5496 and GPT-5.5 at 0.5450; test-time scaling raises it to 0.5797.The base lead holds on PubCADBench (0.5596 vs. 0.5533) and RealCADBench (0.5485 vs. 0.5459) against Gemini-3.1-Pro.Source: §5.2 · p. 7

5.2.1 DETAILED PERFORMANCES ON EACH SLICE

  • On several slices, VisCAD-M1 leads on geometry overlap while another model scores higher on visual judgment.For RealCADBench rendered views, VisCAD-M1 has the highest Solid and Surface IoU (0.5858, 0.2581), while GPT-5.5 has a higher judge score (77.80 vs. 75.56); Gemini’s drawing judge score is 79.08 vs. 69.13 for VisCAD-M1.Source: §5.2.1 · p. 7

5.3 ASSEMBLY-LEVEL RESULTS

  • 85.0 assembly judge score for VisCAD-H1 exceeds Codex at 68.0 and Claude Code at 52.0 in the 50-instance study.Valid-sample counts were 49, 49, and 50, respectively.Source: §5.3 · p. 7

6 ANALYSES

The analyses report gains from mid-training and robust SFT, show that parallel reranking improves VisCAD-M1 while sequential refinement does not, and compare assembly quality and cross-platform stability.

6.2 THE EFFECTIVENESS OF TEST-TIME SCALING

  • With k = 16, parallel TTS reaches 0.5833, more than 2.4 points above naive VisCAD-M1, while sequential TTS does not improve results.The listwise reranker selects among the model’s own rollouts; k = 8 yields 1.3 points without thinking and 1.7 points with thinking.Source: Table 3 · p. 8, §6.2 · p. 8

6.1 THE EFFECTIVENESS OF MID-TRAINING AND ROBUST SFT

  • Mid-training raises many IoU-related entries by about 1 point and judge-score entries by about 5 points across PubCADBench slices.Table 2 covers five evaluation slices and reports success rate, Solid IoU, Surface IoU, and Judge Score.Source: Table 2 · p. 7, §6.1 · p. 8
  • Three robust-SFT target programs per input outperform one toward the end of training.Figure 5 compares Solid and Surface IoU curves for k = 3 and k = 1 across the last five checkpoints, with independent subplot y-axes.Source: §6.1 · p. 8, Figure 5 · p. 8

6.3 ASSEMBLY DESIGN CASE STUDY AND IR STABILITY

  • In representative assembly cases, VisCAD-H1 more often preserves local parts after establishing an envelope, while Codex often stops at a coarse envelope and Claude Code more often distorts topology or fails to export.Figure 6 presents two cases; results for all 50 instances are listed in the appendix.Source: §6.3 · p. 8, Figure 6 · p. 9
  • VisCAD-H1’s platform-agnostic IR yields more stable final assembly designs across FreeCAD, Autodesk Fusion, and SolidWorks than Codex.Figure 7 compares the same assembly realized in the three CAD environments.Source: §6.3 · p. 8, Figure 7 · p. 9

7 CONCLUSION

VisCAD combines multimodal part generation with structured assembly modeling, reporting strong benchmark results for both. Its conclusion also presents reusable data and training infrastructure and identifies unified agent-native CAD intelligence as a future direction.

7 CONCLUSION

  • VisCAD-M1 achieves a 0.5540 part-level profile average versus 0.5496 for the strongest frontier models, while VisCAD-H1 scores 85.0 versus 68.0 for Codex and 52.0 for Claude Code under the assembly judge.Parallel test-time scaling further improves the part-level result.Source: §7 · p. 9
  • VisCAD-M1 maps text, engineering drawings, product photographs, and renders to executable CAD programs, while VisCAD-H1 supports structured assembly modeling.The conclusion frames the suite as spanning part generation and assembly modeling.Source: §7 · p. 9
  • The data-curation pipeline converts heterogeneous, incomplete evidence into intent–program–shape triples and synthetically fills missing intents, programs, or geometry.The multi-stage recipe combines narrow-domain mid-training, general-domain robust supervised fine-tuning, and continual high-quality fine-tuning across data sources of varied quantity and quality.Source: §7 · p. 9
  • Assembly generation follows an intent–IR–program workflow, using a backend-agnostic representation of parts, geometry, mates, and placements across FreeCAD, Fusion, and SolidWorks.The conclusion describes the IR as a link between otherwise isolated CAD dialects.Source: §7 · p. 9
  • The authors identify integrating VisCAD-M1 and VisCAD-H1 into one agent-native system for design, assembly reasoning, execution, verification, and feedback as future work.The stated goal is to unify these functions in one model–agent loop.Source: §7 · p. 9

A DATA CURATION AND TRAINING DETAILS

The appendix describes the raw-data sources and processing steps used to build VisCAD-M1’s dataset, including a category filter for factory-automation products.

A DATA CURATION AND TRAINING DETAILS

  • Three source families provide raw data: factory-automation catalog imagery, purchased industrial CAD assets and drawings, and public CAD code and drawings.Public code is retained when it compiles to a non-empty solid.Source: A DATA CURATION AND TRAINING DETAILS · p. 12
  • Catalog construction expands product terms with an LLM, retrieves related SKUs through a search API, and keeps the main product image; purchased CAD is rendered to single- and multi-view images.The source report also describes raw-data processing across these sources.Source: A DATA CURATION AND TRAINING DETAILS · p. 12
  • A four-step category filter removes software, electronics, consumables, and weak-CAD items, retaining 21 major categories and 128 subcategories.Figure 8 depicts the category-construction flow; Table 4 shows a sample of the filter.Source: A DATA CURATION AND TRAINING DETAILS · p. 12, Figure 8 · p. 12, Table 4 · p. 12

B EVALUATION PROTOCOL

The evaluation protocol scores executed CAD artifacts and uses separate geometry, executability, and judge measures. It specifies alignment and aggregation rules, and notes that geometry scores do not test millimetre-accurate dimensions.

B EVALUATION PROTOCOL

  • VisCAD uses the RealCADBench metric vector and judge prompts, with prompts ratified by CAD practitioners before scoring.The protocol reports executability, geometry, and judge outcomes, which can disagree.Source: B EVALUATION PROTOCOL · pp. 12–13
  • Systems emit FreeCAD Python, which a shared runtime executes and exports as an STL; syntax errors, runtime failures, missing files, or empty geometry fail executability.The protocol scores the executed solid rather than a particular code path.Source: B EVALUATION PROTOCOL · p. 12
  • Solid IoU and Surface IoU use the same deterministic PCA-signed alignment, which removes pose and uniform scale, and a voxel grid of resolution R = 96 with 2% padding.Filled occupancy measures volume, while boundary occupancy is more sensitive to thin structures; scale removal means drawing results do not measure millimetre-accurate dimension recovery.Source: B EVALUATION PROTOCOL · pp. 12–13
  • Executability uses every assigned task, while each quality aggregate uses evaluator-available artifacts and does not zero-fill missing values.The part-level AVG in Table 1 is the profile average.Source: B EVALUATION PROTOCOL · p. 13
  • Judge columns remain on a 0–100 scale, and Equation 4 divides them by 100; the assembly table reports the assembly judge directly.The judge scores part identity and salient features, or assembly component geometry, accuracy, and system design, without access to the reference STL.Source: B EVALUATION PROTOCOL · p. 13

C PART-LEVEL DETAILED RESULTS

This section maps the PubCADBench and RealCADBench slices and identifies detailed per-slice performance tables, alongside a representative assembly-harness comparison.

C PART-LEVEL DETAILED RESULTS

  • PubCADBench covers 1,100 tasks, while the RealCADBench part-level slices cover 1,745 tasks.Table 6 attributes RealCADBench to Team, 2026.Source: Table 5 · p. 13, Table 6 · p. 13
  • Figure 9 shows PubCADBench slices spanning text, renders, textured views, multi-view, and orthographic settings; Figure 10 shows RealCADBench text, drawings, photographs, and rendered CAD views.Tables 5 and 6 define the slices behind Table 1; Figures 9 and 10 reproduce the source-report galleries.Source: C PART-LEVEL DETAILED RESULTS · p. 13, Figure 9 · p. 14, Figure 10 · p. 14
  • Table 7 and Table 8 report detailed per-slice performance for PubCADBench and RealCADBench, respectively.Both tables report AVG as the profile average in Equation 4.Source: Table 7 · p. 14, Table 8 · p. 14
  • Table 9 compares harnesses on 50 assembly instances, split evenly between PubCADBench and RealCADBench.The table covers a representative subset of assembly slices.Source: Table 9 · p. 15

D ASSEMBLY-LEVEL ADDITIONAL RESULTS

The section presents a 50-instance assembly comparison and describes separate judging protocols for part quality and assembly quality. The protocols distinguish geometry, assembly accuracy, and system design, with explicit boundaries on what each dimension scores.

D ASSEMBLY-LEVEL ADDITIONAL RESULTS

E.1 PART JUDGE

  • The Part Judge evaluates the target part itself, excluding pose, contact, collision, mating, and global scale; mesh integrity is a separate programmatic audit.It uses the original global reference image to locate the target part, while the BoM is only for disambiguation and identity context.Source: §E.1 · p. 19
  • Part judging separates geometry realization (P1) from functional design (P2), with fixed subdimensions and weights across observed, mixed, and inferred evidence.Observed evidence uses the reference image; mixed evidence treats visible and invisible portions differently, while inferred evidence does not claim ground-truth fidelity.Source: §E.1 · pp. 19–20

E.2 ASSEMBLY JUDGE

  • The Assembly Judge independently scores Q geometry, F assembly accuracy, and D system design, without using Part-Judge outputs.Asset availability, mating distance, and collision proxies are audit facts whose consequences are assigned to Q, F, or D, not extra dimensions or gates.Source: §E.2 · p. 21
  • Assembly Q scores geometry for all BoM parts, including missing or fallback parts, while excluding pose, gaps, collisions, connectivity, and system function.Its fixed weights are 35/30/25/10 for primary form, defining features, detail fidelity and completeness, and surface refinement.Source: §E.2 · p. 21
  • Assembly F scores pose and silhouette, module layout, mating and connectivity, and non-mating collision or clearance; D scores system function and engineering logic.D weights topology, module completeness, load/support logic, and interface organization at 35/25/25/15; the same fact may affect Q, F, and D differently, without repeating a rationale.Source: §E.2 · pp. 21–22

Related papers

Loading 2609.03811v2…