AI in pattern making and grading: what is automated today
Research shows AI can generate usable patterns for simple garments, while complex construction, varied body shapes and production tolerances still need pattern makers.
KEY TAKEAWAYS Summary by the editors
- A June 2025 Drexel University thesis tested fine-tuned language models that turn garment images into pattern files, and found that a model trained on 150 simple jumpsuits was about 42% usable in manufacturing after minor fixes.
- In the same study, models trained on 300 pants and 3,000 diverse garments produced patterns judged unsuitable for production, so more data did not by itself improve results.
- GarmentCode, a research programming language for parametric sewing patterns, automates low-level tasks such as dart placement and lets a garment be retargeted to different body shapes by changing parameters.
- The Drexel author concludes that AI should augment human pattern makers rather than replace them, and found that application-specific models outperformed general-purpose ones.
- Artisanal skills such as pattern making and tactile toile evaluation remain essential, according to a 2025 academic paper on generative AI in fashion design.
Today, AI automates the repetitive and rule-based parts of pattern making: reading garment features from images, generating draft patterns for simple styles, applying parametric changes and proposing grade rules. It does not yet replace the pattern maker's judgement on complex construction, fit on varied bodies, fabric behaviour or factory tolerances. Published research results are promising for simple garments and weak for anything complicated.
What does pattern making involve, and what can be automated?
Pattern making converts a design into flat pieces with seam allowances, notches, grainlines and construction logic. Grading scales a base pattern across a size range. Both are rule-heavy, which makes them candidates for automation, but both also contain judgement calls that depend on fabric, fit philosophy and the factory. Automation therefore tends to start with defined, repeatable subtasks.
- Extracting garment features from photographs or technical flats.
- Generating a first-draft pattern for a simple, well-represented style.
- Changing parameters, such as length or dart position, and regenerating the pattern.
- Retargeting a pattern to a different body shape or size.
- Applying standard grading increments and flagging missing measurements.
What does the research show about AI-generated patterns?
A June 2025 Drexel University master's thesis by Junyi Chen tested a two-stage pipeline in which a vision model extracts garment features from images, structures them as JSON, and specialised language models turn that specification into SVG pattern files. Training data came from more than 10,000 patterns drawn from industry partner URBN and the public GarmentCodeData repository. Professional pattern makers from URBN evaluated the outputs.
Results were mixed. The vision component performed robustly. A model fine-tuned on 150 simple jumpsuits was about 42% usable in manufacturing, needing only minor fixes. Models trained on 300 mixed-complexity pants and on 3,000 diverse garments produced patterns judged unsuitable for production, so more data did not improve results. The author found that focused, application-specific models outperformed general-purpose ones, and that models struggled with nuanced construction details, varied body shapes and manufacturing-ready specifications. The thesis concludes that AI should augment human pattern makers.
This is a single master's thesis, not a market survey, and it describes a research prototype. It is best read as a realistic indication of where the technique stood in mid-2025.
How do parametric patterns fit in?
A different approach does not generate patterns from scratch but encodes them as programs. GarmentCode, presented by Maria Korosteleva and Olga Sorkine-Hornung in 2023, is a domain-specific language that builds sewing patterns hierarchically from interchangeable, parameterised components. The authors describe it as automating low-level tasks such as placing a dart at a chosen location, and as supporting retargeting garments to different body shapes through a configurator. The page describing it reports no quantitative evaluation, so its claims rest on the showcased examples.
Parametric systems matter for grading because a pattern defined by rules can be regenerated for a new size or body measurement rather than redrawn. They also supply structured training data for machine learning, which is why the public GarmentCodeData repository appears in the Drexel study.
What still needs a pattern maker?
| Task | Status in published research | Human role |
|---|---|---|
| Feature extraction from images | Performed robustly in the Drexel study | Verify details such as seams and trims |
| Simple garment patterns | About 42% usable after minor fixes in the best Drexel model | Check, correct and approve |
| Complex garments | Judged unsuitable for production in the larger models | Create and resolve construction |
| Parametric changes and retargeting | Supported by GarmentCode examples | Set rules and review fit |
| Fabric behaviour and fit judgement | Not solved by simulation alone, per a 2025 design paper | Toile evaluation and fit sessions |
| Factory tolerances and specifications | Models struggled to meet manufacturing-ready specifications | Author and negotiate with the factory |
What data do you need to try this?
The Drexel result suggests that quality and consistency of patterns matter more than volume. A company considering automation should first assess whether its own patterns are digital, in a consistent format and structured by garment type, size set and construction method. Patterns stored as scans or in inconsistent naming schemes are hard to learn from. Defined grade rules and size charts, linked to the style record, are the starting point for any automated grading.
It is also worth separating the use cases. An assistant that suggests a first-draft pattern for a basic, repeated style is a different proposition from one that is expected to handle a tailored jacket. The first can pay off with limited data and a review step. The second is where the research reports the most difficulty.
How should a team pilot AI pattern tools?
- Choose one simple, high-repeat category, such as basic trousers or jumpsuit-style garments.
- Digitise and standardise the existing patterns for that category.
- Define what usable means, for example the share of patterns needing only minor corrections.
- Have pattern makers rate every output and log the time they spend correcting it.
- Compare correction time with drafting from scratch before widening scope.
- Keep physical toiles and fit sessions in the process.
The broader point for planning is that automation in pattern making is uneven across categories. Basic, repeated shapes are within reach, while tailoring and complex draping are not, on the evidence cited. Teams should therefore scope projects by garment category rather than by the whole range, and expect to keep experienced pattern makers in the loop for review and for the cases the tools cannot handle.
Skills also deserve planning. If routine drafting is automated, junior pattern makers may lose the repetitive work through which they build judgement, a deskilling concern raised in the 2025 design paper. Companies that adopt these tools should decide how apprentices still learn to draft, grade and resolve construction by hand, because the same people will be needed to review what the tools produce.
Frequently asked questions
Can AI make sewing patterns automatically?
For simple garments, research prototypes can. A 2025 Drexel thesis found a model trained on simple jumpsuits was about 42% usable after minor fixes, while models trained on more complex sets produced patterns judged unsuitable for production.
Will AI replace pattern makers?
Published research points to augmentation. The Drexel thesis concludes AI should augment pattern makers, and a 2025 design paper says artisanal skills such as pattern making and toile evaluation remain essential.
Can AI grade patterns across sizes?
Rule-based and parametric systems can regenerate patterns for new sizes or body shapes, as GarmentCode demonstrates. Fit across real, varied bodies still needs human review.
What data does AI pattern making need?
Digital patterns in a consistent format, linked to garment type, size set and construction details. The Drexel study used more than 10,000 patterns, though more data alone did not improve results.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
