Blueprint2CAD: Benchmarking Foundation Models on Engineering Drawing-to-CAD Render Correspondence

Alexander Valverde; Danny Xie-Li; Zhichao Wang; Marcos V. Conde

cs.CVNov 2, 2026

Blueprint2CAD: Benchmarking Foundation Models on Engineering Drawing-to-CAD Render Correspondence

Authors: Alexander Valverde, Danny Xie-Li, Zhichao Wang, Marcos V. Conde

Organizations: Novineer · Instituto Tecnológico de Costa Rica · University of Würzburg

Abstract

Engineering drawings specify mechanical parts through orthographic projections. Recognizing which part a drawing depicts underlies part search, design reuse, and digitization, yet no public benchmark measures whether vision models can do it. We introduce Blueprint2CAD, a benchmark for drawing-to-render correspondence: given an orthographic line drawing and a shaded render, usually from another viewpoint, decide whether they show the same part. Its data layer, BlueCAD-30K, applies an open drafting pipeline to 29,925 Zero-to-CAD models, adding view-aligned renders, clean and dimensioned drawings from six viewpoints, and surface-type metadata. Frozen encoders retrieve the matching part from a 10,000-model gallery with up to 65% Recall@1 when viewpoints are known but 43% when they are not, and the four encoders we test lie within six points of each other: viewpoint, not backbone, is the bottleneck. Six zero-shot vision-language models reach only 0.52-0.67 verification accuracy, with model-specific Yes/No biases that chain-of-thought shifts but does not remove; a quick LoRA fine-tuning on in-domain pairs improves verification accuracy by a few points but sharpens these biases and leaves the viewpoint failure intact.

Figures & tables

Panel of Figure 2 — Figure 2: Examples of CAD models from the two VLM evaluation subsets: shaded render, clean 2D drawing, and dimensioned 2D drawing.

Figure 3: A CAD example with its six camera directions and paired render, drawing, and dimensioned drawing views.

	R@1 , view-agnostic $\uparrow$		R@1 , view-aligned $\uparrow$		Al-mean
Backbone	Max	Mean	Al-max	Al-mean	R@1 (R $\rightarrow$ D) $\uparrow$	MRR $\uparrow$	MedR $\downarrow$
Silhouette	7.85	20.16	10.96	48.54	48.08	0.555	2
DINOv3	29.10	43.11	34.67	65.14	57.57	0.728	1
TIPSv2	34.39	34.65	38.58	64.65	58.67	0.724	1
SigLIP2	29.52	38.07	33.54	64.15	57.15	0.714	1
RADIO	27.46	36.54	30.98	59.51	49.58	0.675	1

Table 1: Experiment 1: frozen-backbone cross-modal retrieval. R@1 (%, D

\rightarrow

R) for the four scoring strategies defined in the protocol: Max / Mean are view-agnostic, Al-max / Al-mean assume known viewpoints. MRR, median rank, and reverse direction for Al-mean. Silhouette : non-learned outline baseline, six filled

16{\times}16

view masks plus aspect ratio (the hard-negative descriptor of Sec. D ). Bold: best per column. Full metrics in the supplementary.

Views	R@1 $\uparrow$	R@5 $\uparrow$	R@10 $\uparrow$	MRR $\uparrow$	MedR $\downarrow$
front	14.88	25.98	31.38	0.206	62
+ back	15.76	27.08	33.11	0.217	48
+ bottom	42.19	61.12	68.20	0.512	2
+ left	52.15	70.74	76.93	0.607	1
+ top	65.59	81.90	87.15	0.731	1
+ right (all 6)	65.14	81.79	86.87	0.728	1

Table 2: Experiment 2: cumulative view fusion (frozen DINOv3, Al-mean, D

\rightarrow

R). Gains concentrate where a new viewing axis is introduced and on the top view; the opposing back and right views add almost nothing. Single- and two-view results for all encoders in the supplementary.

Backbone	R@1 (%) $\uparrow$	MRR $\uparrow$	MedR	Worse (%)	Better (%)
DINOv3	65.14 $\rightarrow$ 58.10	0.728 $\rightarrow$ 0.663	1 $\rightarrow$ 1	29.1	13.4
TIPSv2	64.65 $\rightarrow$ 61.61	0.724 $\rightarrow$ 0.697	1 $\rightarrow$ 1	23.2	16.0
SigLIP2	64.15 $\rightarrow$ 61.51	0.714 $\rightarrow$ 0.692	1 $\rightarrow$ 1	23.7	20.3
RADIO	59.51 $\rightarrow$ 58.84	0.675 $\rightarrow$ 0.673	1 $\rightarrow$ 1	22.5	22.9

Table 3: Experiment 3: annotation ablation (normal

\rightarrow

dim, Al-mean, D

\rightarrow

R). Worse / Better: share of queries whose rank degrades / improves with dimension annotations; the rest keep their rank. Bold: best value before and after annotation. Full metrics in the supplementary.

Figure 4: Examples of 2D drawing retrieval tasks with four candidate CAD renders. Green border: ground-truth render; blue border: model selection. A correct selection makes the two borders coincide.

	Task 1		Task 1 + DIM		Task 2		Task 3
Model	Acc. $\uparrow$	F1 $\uparrow$	Acc. $\uparrow$	F1 $\uparrow$	Acc. $\uparrow$	F1 $\uparrow$	Acc. $\uparrow$	F1 $\uparrow$
Closed-source models
Claude Sonnet 5	0.635	0.631	0.590	0.560	0.374	0.372	0.520	0.409
Gemini 3.6 Flash	0.654	0.620	0.670	0.646	0.690	0.689	0.562	0.491
GPT-5.6 Luna	0.670	0.670	0.654	0.649	0.637	0.636	0.676	0.676
Open-source models

Table 4: Main VLM results (accuracy, macro-F1) . Task 1: verification without and with dimensions (DIM),

n{=}1000

for zero-shot models (one item lacks its render; supplementary Sec. C.1 ) and 1,000 for FT; per-model 95% Wilson intervals are at most

\pm 0.031

, and paired model comparisons are in supplementary Tab. 14 . Task 2: four-way retrieval (chance 0.25). Task 3: unguided complexity probe scored against surface-type class (chance 0.50). Bold: best model per column, not including the fine-tuned (FT) row.

Model	Acc. $\uparrow$	Macro F1 $\uparrow$	Complex Rec. $\uparrow$	Simple Rec. $\uparrow$
Closed-source models
Claude Sonnet 5	0.520	0.409	0.086	0.954
Gemini 3.6 Flash	0.562	0.491	0.188	0.940
GPT-5.6 Luna	0.676	0.676	0.712	0.640
Open-source models
Mistral-S 3.2 24B	0.521	0.409	0.087	0.954

Table 5: Complexity probe (Task 3) : per-class recall, overall accuracy and macro-F1. Labels are B-rep-derived (prismatic = Simple, freeform = Complex); the task probes alignment between an unguided notion of complexity and surface type.

	Undimensioned			Dimensioned
Model	Pos. $\uparrow$	Neg. $\uparrow$	$\Delta$	Pos. $\uparrow$	Neg. $\uparrow$	$\Delta$	$\Delta_{\text{DIM}}$
Closed-source models
Claude Sonnet 5	0.728	0.541	$+0.187$	0.848	0.331	$+0.517$	$+0.120$
Gemini 3.6 Flash	0.356	0.952	$-0.596$	0.412	0.928	$-0.516$	$+0.056$
GPT-5.6 Luna	0.703	0.637	$\mathbf{+0.066}$	0.772	0.535	$\mathbf{+0.237}$	$+0.069$
Open-source models

Table 6: Task 1 response bias: accuracy on positive (

n{=}500

) and negative (

n{=}499

; 500 for FT) pairs, without and with dimensions.

\Delta

= Pos.

-

Neg.;

\Delta_{\text{DIM}}

= change in Pos. from adding dimensions. Bold: best zero-shot model per column and smallest

|\Delta|

Model	Acc. $\uparrow$	F1 $\uparrow$	Pos. $\uparrow$	Neg. $\uparrow$	$\Delta$ (forced-choice)	$n$
Closed-source models
Claude Sonnet 5	0.673	0.664	0.515	0.830	$-0.315$ ( $+0.187$ )	996
Gemini 3.6 Flash	0.667	0.639	0.386	0.949	$-0.563$ ( $-0.596$ )	989
GPT-5.6 Luna	0.681	0.679	0.604	0.759	$\mathbf{-0.155}$ ( $+0.066$ )	959
Open-source models
Mistral-S 3.2 24B	0.545	0.523	0.758	0.331	$+0.427$ ( $-0.896$ )	1000

Table 7: Task 1 under chain-of-thought prompting (undimensioned). Accuracy, macro-F1, positive- and negative-pair accuracy, and

\Delta

= Pos.

-

Neg., with the forced-choice

\Delta

from Table 6 in parentheses.

n

: items with a parseable final Yes / No . Bold: best zero-shot model per column and smallest

|\Delta|

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Signal	Weight	Normalization
Freeform face ratio	25	$\min(1,\,2r_{\text{freeform}})$
Face count (log)	15	$\min(1,\,\log_{10}(\text{faces})/3)$
Non-planar face ratio	15	$r_{\text{curved face}}$
Surface-type entropy	15	$\min(1,\,1.5\,H)$
Inner boundary loops	12	$\min(1,\,\text{loops}/12)$
Curved edge ratio	10	$r_{\text{curved edge}}$

Appendix

Table 8: Selection-score signals, weights (sum to 100), and normalization to

[0,1]

r

: raw ratio for the named signal;

H

: Shannon entropy of the surface-type histogram, normalized by

\log(11)

;

c

: Wadell’s sphericity,

c=\sqrt[3]{36\pi V^{2}}/A

, clipped to

[0,1]

. Weights are fixed manually: the freeform-face ratio as the clearest primitive-vs-organic separator, surface-type entropy after the information-theoretic measures of SHREC’21 [ 3 ] with the per-face classification of [ 32 , 44 ] , and compactness as Wadell’s sphericity [ 77 ] entered as

1-c

	Simple	Complex	Full
Faces	23	42	37
Edges	120	209	190
Inner loops	0	4	4
Freeform face ratio	0.00	0.50	0.00
Non-planar face ratio	0.07	0.81	0.30
Curved edge ratio	0.08	0.86	0.28

Appendix

Table 9: Per-model medians for the VLM subsets (Simple = prismatic, Complex = freeform) and the full sample (

n{=}30{,}237

); interquartile ranges are visible in Fig. 5 . Bottom block: fraction of all faces per surface class.

Figure 5: Distributions of B-rep signals for the VLM evaluation subsets (simple in orange, complex in blue) against the full sample (grey). Face count, edge count, and aspect ratio are on a log axis.

Figure 6: Fraction of faces per surface type for the simple subset, the complex subset, and the full sample. Simple parts are almost entirely planar and cylindrical; complex parts are dominated by B-spline faces.

View	Proj. direction	Image $x$ -axis	Image up
Front	$(0,-1,0)$	$(1,0,0)$	$(0,0,1)$
Right	$(1,0,0)$	$(0,1,0)$	$(0,0,1)$
Back	$(0,1,0)$	$(-1,0,0)$	$(0,0,1)$
Left	$(-1,0,0)$	$(0,-1,0)$	$(0,0,1)$
Top	$(0,0,1)$	$(1,0,0)$	$(0,1,0)$
Bottom	$(0,0,-1)$	$(1,0,0)$	$(0,-1,0)$

Appendix

Table 10: Canonical camera definitions. Projection directions point from the part toward the viewer (FreeCAD convention); up

=

direction

\times

x

-axis. All vectors are in native STEP coordinates, with no per-model reorientation.

Backbone	Direction	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	R@5 $\uparrow$	R@10 $\uparrow$	MRR $\uparrow$	MedR $\downarrow$
Silhouette	D $\rightarrow$ R	48.54	55.41	59.04	62.79	68.34	0.555	2
Silhouette	R $\rightarrow$ D	48.08	55.40	58.75	63.14	68.98	0.554	2
DINOv3	D $\rightarrow$ R	65.14	73.56	77.67	81.79	86.87	0.728	1
DINOv3	R $\rightarrow$ D	57.57	66.34	70.32	75.28	81.44	0.659	1
TIPSv2	D $\rightarrow$ R	64.65	73.34	77.08	81.63	86.42	0.724	1
TIPSv2	R $\rightarrow$ D	58.67	66.60	71.11	75.55	81.18	0.665	1

Appendix

Table 11: Complete frozen-backbone retrieval results (extends Tab. 1 ). View-aligned setting (Al-mean, viewpoints known), both directions, all metrics; Silhouette is the non-learned outline baseline. Gallery:

N{=}10{,}000

-model sample; chance R@1

=0.01\%

, chance median rank

\approx 5{,}000

Model	Identifier	Temp.	Reasoning	Max tok.
Claude Sonnet 5	claude-sonnet-5	default	off	2000
Gemini 3.6 Flash	gemini-3.6-flash	default	default	2000
GPT-5.6 Luna	gpt-5.6-luna	default	default	2000
Mistral-S 3.2 24B	mistralai/Mistral-Small-3.2-24B-Instruct-2506	default	n/a	2000 (20 in T3)
InternVL3.5 14B	OpenGVLab/InternVL3_5-14B	default	off	2000 (100 in T3)
Qwen3-VL 8B	Qwen/Qwen3-VL-8B-Instruct	default	n/a	2000

Appendix

Table 13: VLM inference settings. Identifier: API model alias (closed, queried 15–25 August 2026) or Hugging Face checkpoint (open). Temp.: no run set a temperature, so “default” is the provider’s or the serving engine’s default. Reasoning: “off” if thinking is disabled, “default” if left unset, n/a for models without a reasoning mode. Max tok.: output-token limit passed in the call, the same for forced-choice and CoT prompts; for GPT-5.6 Luna it also counts reasoning tokens.

Comparison	$\Delta$ Acc	$b{+}c$	$p_{\text{Holm}}$
Task 1, leading models (undimensioned)
GPT-5.6 vs. Qwen3-VL	$+0.000$	342	1.000
GPT-5.6 vs. Gemini	$+0.017$	431	1.000
GPT-5.6 vs. Claude	$+0.036$	344	0.295
Qwen3-VL vs. Gemini	$+0.016$	418	1.000
Qwen3-VL vs. Claude	$+0.035$	295	0.285

Appendix

Table 14: Paired exact McNemar tests on the items both systems answered.

\Delta

Acc: accuracy of the first system minus the second on those items;

b{+}c

: discordant items;

p_{\text{Holm}}

: Holm-corrected within each block.

		Without DIM				With DIM
	Model	Acc.	F1	Rec.	Prec.	Acc.	F1	Rec.	Prec.
Complex	Claude Sonnet 5	0.665	0.665	0.666	0.668	0.611	0.597	0.615	0.637
	Gemini 3.6 Flash	0.717	0.699	0.713	0.777	0.735	0.724	0.732	0.773
	GPT-5.6 Luna	0.711	0.710	0.710	0.711	0.705	0.704	0.707	0.711
	Mistral-S 3.2 24B	0.545	0.428	0.537	0.675	0.563	0.488	0.556	0.628
	InternVL3.5 14B	0.659	0.638	0.655	0.698	0.663	0.639	0.659	0.711

Appendix

Table 15: Task 1 VLM results by surface-type class and pair tier, without (left) and with (right) dimension annotations. Bold: best zero-shot model per column; the fine-tuned (FT) row is excluded.

Model	front	right	back	left	top	bottom
Positive pairs ( $n$ = 86, 81, 92, 79, 78, 84)
Claude Sonnet 5	0.92	1.00	0.89	1.00	0.53	0.74
Gemini 3.6 Flash	0.35	0.56	0.27	0.54	0.42	0.36
GPT-5.6 Luna	0.76	0.95	0.74	0.89	0.63	0.68
Mistral-S 3.2 24B	0.16	0.21	0.20	0.28	0.06	0.04
InternVL3.5 14B	0.43	0.68	0.42	0.66	0.14	0.17

Appendix

Table 16: Task 1 accuracy by render viewpoint, with DIM, under the forced-choice Task 1 prompt of Sec. 5.2 (weighting by

n

recovers the positive and negative accuracies of Tab. 6 ). Bold: best viewpoint per model; the fine-tuned (FT) row is not marked.

Model	front	right	back	left	top	bottom
Positive pairs ( $n$ = 86, 81, 92, 79, 78, 84/83 † )
Claude Sonnet 5	0.77	0.98	0.68	0.97	0.46	0.51
Gemini 3.6 Flash	0.31	0.47	0.18	0.47	0.42	0.31
GPT-5.6 Luna	0.74	0.88	0.67	0.85	0.51	0.57
Mistral-S 3.2 24B	0.12	0.12	0.10	0.11	0.06	0.00
InternVL3.5 14B	0.41	0.65	0.39	0.71	0.12	0.13

Appendix

Table 17: Task 1 accuracy by render viewpoint, without DIM, under the forced-choice Task 1 prompt of Sec. 5.2 (weighting by

n

recovers the positive and negative accuracies of Tab. 6 ). Bold: best viewpoint per model; the fine-tuned (FT) row is not marked.

Figure 7: Examples of 2D drawing retrieval tasks with four candidate CAD renders. Green border: ground truth; blue border: model selection.

	Model	Acc.	F1	Rec.	Prec.
Overall	Claude Sonnet 5	0.374	0.372	0.379	0.425
	Gemini 3.6 Flash	0.690	0.689	0.689	0.691
	GPT-5.6 Luna	0.637	0.636	0.635	0.643
	Mistral-S 3.2 24B	0.274	0.235	0.267	0.282
	InternVL3.5 14B	0.266	0.274	0.269	0.310
	Qwen3-VL 8B	0.317	0.305	0.321	0.322

Appendix

Table 18: Task 2 VLM results by model (four-way drawing-to-render retrieval). Bold: best zero-shot model per column; the fine-tuned (FT) row is excluded

Panel of Figure 8 — Figure 8: Task 3 predictions across models (green = correct, red = incorrect).

Group	Tier	Result	Count (%)
Complex	Any-cl. neg. ( $n{=}122$ )	All success	50 (41.0%)
	Any-cl. neg. ( $n{=}122$ )	Mixed	72 (59.0%)
	Same-cl. neg. ( $n{=}133$ )	All failed	1 (0.8%)
		All success	53 (39.8%)
		Mixed	79 (59.4%)
	Positive ( $n{=}245$ )	All failed	25 (10.2%)

Appendix

Table 19: Task 1 failure analysis by surface-type class, negative type, and prediction outcome.

Figure 9: Task 1 examples where all evaluated VLMs misclassified positive pairs.

Figure 10: Task 1 examples where all evaluated VLMs correctly matched pairs.

Figure 11: Task 1 qualitative examples showing mixed model predictions. Rows labelled Hard Negative are same-class negatives (same prismatic/freeform class) and Easy Negative rows are any-class negatives; neither is mined for visual similarity.

Split	Models	Pairs	Pos.	Hard neg.	Same view
Train	27,479	109,916	54,958	38,483	0.166 / 0.166
Validation	1,446	5,784	2,892	2,058	0.169 / 0.162

Appendix

Table 20: Fine-tuning pairs. Same view: share of positive / negative pairs whose render shows the drawing’s view.

Setting	Value
LoRA rank / $\alpha$ / dropout	16 / 32 / 0.05
Trainable parameters	52.5M (0.60%)
LoRA targets (LM)	q,k,v,o,gate,up,down_proj
LoRA targets (vision)	qkv,proj,linear_fc1,linear_fc2
Optimizer	AdamW, lr $10^{-4}$ , cosine schedule, 3% warmup
Effective batch	16 (1 pair $\times$ 16 accumulation)

Appendix

Table 21: Fine-tuning prompt and configuration (Qwen3-VL 8B). The prompt is used for training only; evaluation uses the Task 1 prompt of the main text.

VLM Evaluation

Vision-language model evaluation measures multimodal systems’ perception, reasoning, grounding, robustness, and reliability across controlled and real-world tasks.

See all 1,017 papers

LED-BENCH: A Benchmark for Large-format Engineering Drawing Understanding
Oct 14, 2026Ke Li, Tianyi Wang, Enning You +4Document Understanding Document Image Understanding
Institute of AI for Engineering, Tongji University
CrossProjection: Geometric Grounding Beyond Viewpoint Change in Architectural Drawings
Aug 1, 2026Kaho Li, Pengyu Zeng, Yuqin Dai +3Multi-View Consistency Vision-Language Grounding
Shenzhen International Graduate School, Tsinghua University, Shenzhen 518055, China · UCL Institute for Environmental Design and Engineering, The Bartlett School of Environment, Energy and Resources, University College London, London, UK
MechVQA: Benchmarking and Enhancing Multimodal LLMs on Comprehensive Mechanical Drawing Understanding
May 29, 2026Qian Kou, Xiaofeng Shi, Yulin Li +4Visual Spatial Reasoning Multimodal Large Language Models
Beijing Academy of Artificial Intelligence (BAAI), China · Institute of Information Engineering, Chinese Academy of Sciences, China · Beijing University of Technology, China

Oct 14, 2026J/K move · Enter open · S save

LED-BENCH: A Benchmark for Large-format Engineering Drawing Understanding

Ke Li, Tianyi Wang, Enning You, Chaoqun Wan+3

Institute of AI for Engineering, Tongji University

Document Understanding Document Image Understanding Multimodal Model Evaluation Spatial Reasoning Benchmarks

Open paper AskPDF arXiv

Real-world engineering drawings are often large-format and information-dense. Answering practical engineering questions therefore requires both fine-grained understanding of engineering information and complex reasoning over objects, relations, and evidence distributed across different regions or views. Existing benchmarks cover a range of engineering-drawing understanding tasks involving local regions, individual views, specific elements, or more constrained visual settings. However, systematic evaluation of understanding over realistic, large-format engineering drawings remains limited. To address this gap, we introduce LED-BENCH, a multidisciplinary benchmark designed to evaluate fine-grained understanding and complex reasoning over large-format engineering drawings, spanning the mechanical, architectural, and transportation domains. Extensive evaluation of 11 representative multimodal large language models, including 8 proprietary models and 3 open-source models, reveals substantial remaining difficulty: GPT-5.6 Sol achieves 52.65% average accuracy, performance varies across engineering domains, and all evaluated models perform worse on questions requiring broader evidence integration. These results highlight fine-grained understanding, evidence integration, and complex reasoning over large-format engineering drawings as persistent challenges for current multimodal models.

Abstract

VLM Evaluation

Explore similar work

LED-BENCH: A Benchmark for Large-format Engineering Drawing Understanding

CrossProjection: Geometric Grounding Beyond Viewpoint Change in Architectural Drawings

MechVQA: Benchmarking and Enhancing Multimodal LLMs on Comprehensive Mechanical Drawing Understanding