renue Implements 'Drawing Cleanup' Feature Using GPT-image-2 in Drawing SaaS 'Drawing Agent', Achieving Precise Noise Removal and High-Accuracy Drawing Creation
renue Inc. has added a new 'Drawing Cleanup' feature to its 2D-to-3D SaaS 'Drawing Agent', utilizing OpenAI's gpt-image-2 to automatically remove auxiliary lines and dimensions, enhancing 3D model generation accuracy.
📋 Article Processing Timeline
- 📰 Published: April 24, 2026 at 18:00
- 🔍 Collected: April 24, 2026 at 09:31
- 🤖 AI Analyzed: April 24, 2026 at 09:57 (25 min after Collected)
renue Inc. (Headquarters: Minato-ku, Tokyo; CEO: Yusuke Yamamoto) has added a new 'Drawing Cleanup' feature utilizing OpenAI's image generation model 'gpt-image-2' to its web application 'Drawing Agent', which autonomously generates 3D models from 2D drawings. It removes noise such as auxiliary lines, dimension annotations, hatching, and leader lines all at once, extracting only the pure outline and material boundaries.
Executive Summary
The newly added 'Drawing Cleanup' is a new feature that uses OpenAI's image generation model 'gpt-image-2' to remove noise elements prior to drawing reading.
It removes auxiliary lines, dimension annotations, hatching, leader lines, cross-section labels, and title blocks from the input drawing, extracting only the pure shape. It automates processes that previously relied on manual correction.
Through verification, we have confirmed that shape extraction is successful even for complex drawings containing curved surfaces and composite shapes. We will progressively apply it to areas with high symbol density, such as manufacturing prototype drawings, construction component drawings, and design drawings. We plan to gradually roll it out to existing customers' operational environments and reflect the knowledge gained from actual operations into improving filter accuracy.
About renue's Drawing SaaS 'Drawing Agent'
We provide 'Drawing Agent', a drawing SaaS that automatically generates 3D models simply by uploading 2D drawing images.
Even without CAD software operational skills, designers themselves can convert 2D drawings into 3D data in minutes. We transform the conversion work, which conventionally took CAD operators hours, into an experience completed just by uploading a file.
Recently, we added a feature to 'Drawing Agent' to reference parts information in advance.
Premises of the Target Domain
In manufacturing and construction sites, CAD-native PDFs, scanned PDFs, paper drawings, and FAX images are mixed. In paper-derived drawings, dimension lines, hatching, and handwritten notes are factors that lower reading accuracy. While existing AI-OCR is strong in text extraction, it struggles with noise removal in the graphic layer.
In product drawings that heavily use curved surfaces, distinguishing between auxiliary lines and solid lines is difficult even for humans. As the retirement of skilled workers progresses, dependency on individuals for reading drawings and cost estimation continues.
In April 2026, OpenAI released the image generation model 'gpt-image-2', reporting improvements in text drawing, editing, and multilingual accuracy. The model has reached a level applicable even to generating process drawings including dimension annotations. Taking this evolution into account, renue reviewed the configuration of its drawing reading preprocessing.
Goal
With this update, we will resolve the structure where the noise removal accuracy in the pre-reading stage was the bottleneck for the entire product. It is an approach to break through the plateau in reading accuracy by reviewing the model configuration. By automating preprocessing, we aim to simultaneously expand the accuracy and applicable range of the downstream engine. We will increase the processing volume per cost estimator, leading to a shortened lead time for quote responses.
We will proceed in parallel with expanding supported drawing patterns and reducing the customer's burden of drawing preparation. We will transition from an operation centered on CAD-native PDFs to a configuration that can withstand field inputs including paper, scans, and handwriting. Regarding automatic 3D CAD generation from 2D drawings, we will also increase the stability of deliverables by raising the base input quality.
In the medium to long term, we will build an environment where drawings handled in each process of design, cost estimation, and construction can be processed in a single workflow. We aim to lower the customer's drawing preparation costs and expand the situations where AI agents are used in actual operations. Our policy is to absorb the transitional period of sites where paper and electronic drawings coexist on the product side.
Challenges
Auxiliary lines and hatching become reading noise
Auxiliary lines, dimension lines, hatching, and leader lines on drawings are background information for humans, but they are inputs equivalent to solid lines for image recognition models. While there are operations to correct cross-sectional drawings as originals, the manual preprocessing cost remains. In conventional rule-based processing, the generalization of line types did not work, making it necessary to rebuild preprocessing for each project.
Diversity in notation and resolution
Even for the same material, notation methods vary by company and site. The meaning of line types, types of hatching, and symbol legends are not standardized. In drawings via scan or FAX, the resolution drops, and thin dimension lines or small symbols disappear. Even if training data for multiple patterns is prepared, it is difficult to keep up with the variations in field inputs.
Handling curved products and handwritten notes
Drawings that heavily use curved surfaces and composite shapes express the shape through a combination of cross-sections and developments. The number of auxiliary lines is also large, making it an area with high reading difficulty. At construction sites, last-minute design changes are often handwritten.
Executive Summary
The newly added 'Drawing Cleanup' is a new feature that uses OpenAI's image generation model 'gpt-image-2' to remove noise elements prior to drawing reading.
It removes auxiliary lines, dimension annotations, hatching, leader lines, cross-section labels, and title blocks from the input drawing, extracting only the pure shape. It automates processes that previously relied on manual correction.
Through verification, we have confirmed that shape extraction is successful even for complex drawings containing curved surfaces and composite shapes. We will progressively apply it to areas with high symbol density, such as manufacturing prototype drawings, construction component drawings, and design drawings. We plan to gradually roll it out to existing customers' operational environments and reflect the knowledge gained from actual operations into improving filter accuracy.
About renue's Drawing SaaS 'Drawing Agent'
We provide 'Drawing Agent', a drawing SaaS that automatically generates 3D models simply by uploading 2D drawing images.
Even without CAD software operational skills, designers themselves can convert 2D drawings into 3D data in minutes. We transform the conversion work, which conventionally took CAD operators hours, into an experience completed just by uploading a file.
Recently, we added a feature to 'Drawing Agent' to reference parts information in advance.
Premises of the Target Domain
In manufacturing and construction sites, CAD-native PDFs, scanned PDFs, paper drawings, and FAX images are mixed. In paper-derived drawings, dimension lines, hatching, and handwritten notes are factors that lower reading accuracy. While existing AI-OCR is strong in text extraction, it struggles with noise removal in the graphic layer.
In product drawings that heavily use curved surfaces, distinguishing between auxiliary lines and solid lines is difficult even for humans. As the retirement of skilled workers progresses, dependency on individuals for reading drawings and cost estimation continues.
In April 2026, OpenAI released the image generation model 'gpt-image-2', reporting improvements in text drawing, editing, and multilingual accuracy. The model has reached a level applicable even to generating process drawings including dimension annotations. Taking this evolution into account, renue reviewed the configuration of its drawing reading preprocessing.
Goal
With this update, we will resolve the structure where the noise removal accuracy in the pre-reading stage was the bottleneck for the entire product. It is an approach to break through the plateau in reading accuracy by reviewing the model configuration. By automating preprocessing, we aim to simultaneously expand the accuracy and applicable range of the downstream engine. We will increase the processing volume per cost estimator, leading to a shortened lead time for quote responses.
We will proceed in parallel with expanding supported drawing patterns and reducing the customer's burden of drawing preparation. We will transition from an operation centered on CAD-native PDFs to a configuration that can withstand field inputs including paper, scans, and handwriting. Regarding automatic 3D CAD generation from 2D drawings, we will also increase the stability of deliverables by raising the base input quality.
In the medium to long term, we will build an environment where drawings handled in each process of design, cost estimation, and construction can be processed in a single workflow. We aim to lower the customer's drawing preparation costs and expand the situations where AI agents are used in actual operations. Our policy is to absorb the transitional period of sites where paper and electronic drawings coexist on the product side.
Challenges
Auxiliary lines and hatching become reading noise
Auxiliary lines, dimension lines, hatching, and leader lines on drawings are background information for humans, but they are inputs equivalent to solid lines for image recognition models. While there are operations to correct cross-sectional drawings as originals, the manual preprocessing cost remains. In conventional rule-based processing, the generalization of line types did not work, making it necessary to rebuild preprocessing for each project.
Diversity in notation and resolution
Even for the same material, notation methods vary by company and site. The meaning of line types, types of hatching, and symbol legends are not standardized. In drawings via scan or FAX, the resolution drops, and thin dimension lines or small symbols disappear. Even if training data for multiple patterns is prepared, it is difficult to keep up with the variations in field inputs.
Handling curved products and handwritten notes
Drawings that heavily use curved surfaces and composite shapes express the shape through a combination of cross-sections and developments. The number of auxiliary lines is also large, making it an area with high reading difficulty. At construction sites, last-minute design changes are often handwritten.