Source-linked AI summary
EasyFashion: A Human-AI Co-Creation System for Personalized Fashion Design and Sewing Pattern Generation
Hong Qu, Zhaoxiang Xu, Jinbo Luo, Yujie Zhao, Jie Zhang, Yadie Yang
TL;DR
Personalized garments can reflect individual style and fit preferences, but translating those preferences into physical garments remains difficult, and many AI systems focus on visualization rather than production. EasyFashion combines multimodal design refinement, reconstructed personal avatars for try-on, and sewing-pattern generation. Technical evaluations, user studies, and a production case report support for design expression, body-specific evaluation, and production-oriented outputs.
Problem
Non-professionals can struggle to communicate garment intent, while existing workflows are costly and difficult to revise and many AI fashion systems emphasize visualization over production support.
Source: §1 · pp. 1–2Method
EasyFashion uses a VLM-based agent to turn text, reference images, and sketches into editable garment specifications, alongside personal-body reconstruction, try-on, and sewing-pattern generation.
Source: §1 · p. 3Results
Technical evaluations, user studies, and a production case report that EasyFashion supports iterative design refinement, improved fit beyond standard sizing, and production-oriented deliverables shareable with makers.
Source: §8 · p. 25Takeaways & Limitations
The reported findings highlight multimodal design expression, body-specific evaluation, and production-oriented outputs as valuable in personalized garment design.
Source: §1 · p. 3, §8 · p. 25Current body models use a small set of measurements and omit finer differences such as leg length, arm length, and thigh circumference that may affect fit.
Source: §7 · p. 25
Abstract
from arXiv · showhide
People often want garments that reflect their aesthetic preferences, fit their bodies, and meet their sizing needs, yet turning these requirements into physical garments remains difficult. Ready-to-wear options provide limited personalization, while custom tailoring is costly and time-consuming. Recent generative artificial intelligence (AI) systems can visualize garment ideas but often stop short of supporting downstream production. To address this gap, we present EasyFashion, a human-AI co-creation system that enables users to iteratively refine design intent for personalized garment style and size, evaluate designs through virtual try-on on reconstructed personal avatars, and generate sewing patterns for garment production. Using reference images, text descriptions, and body photos as input, EasyFashion translates user intent into structured garment specifications and try-on results. Technical experiments, user studies, and a real-world production case demonstrate the value of EasyFashion for multimodal design expression, body-specific evaluation, and production-oriented outputs in personalized garment design. Our project page is available at https://zoequ.github.io/easyfashion.github.io/.
1. Introduction
Personalized garments could better align style and fit with individual preferences and bodies, but non-professionals face costly, difficult-to-revise workflows and challenges communicating design intent. EasyFashion responds with multimodal, iterative design, body-specific try-on, and production-oriented outputs, informed by user research and technical and user evaluations.
1. Introduction
- Personalization spans both appearance and fit, while existing custom workflows can be costly, time-consuming, and difficult to revise after production begins.Non-professionals may also struggle to articulate design intent, leading to back-and-forth, revisions, and production rework.Source: §1 · p. 1
- Many existing AI fashion systems emphasize visual generation and style variation, with less support for confident judgments about fit and appearance on a user's own body.The introduction also identifies burdensome prompting, slow or hard-to-interpret feedback, and unreliable try-on as friction points.Source: §1 · p. 2
- A formative study with non-professionals and professionals identified low-burden input, responsive iterative feedback, and body-specific evaluation as design needs.The study used a questionnaire survey and semi-structured interviews.Source: §1 · pp. 2–3
- EasyFashion combines a VLM-based Fashion Agent, personalized body reconstruction, body-specific try-on, and downstream sewing-pattern generation in an end-to-end system.The agent turns text, reference images, and sketches into an editable JSON garment specification and updates it from feedback; body reconstruction is a separate component.Source: §1 · p. 3
- The paper evaluates body-reconstruction robustness and garment-specification generation technically, then studies usability, responsiveness, and design refinement with try-on feedback.It also reports physical prototyping of selected garments.Source: §1 · p. 3
2. Related Works
Related work establishes interaction, production, and body-reconstruction foundations for EasyFashion. The section positions the system around editable multimodal intent, pattern-compatible outputs, and more reliable body-specific try-on.
2.1. Multimodal Human-AI Co-Creation in Fashion Design
- Multimodal fashion co-creation supports ideation and image generation, but non-professional users still face difficulty translating vague inputs into concrete, editable production-relevant decisions.Prior systems often emphasize visual outputs, while production-oriented co-creation requires representations editable during interaction and compatible with pattern making.Source: §2.1 · p. 4
2.2. Sewing Pattern-Based Garment Generation
- Sewing-pattern approaches represent garments as 2D panels with stitching topology, aligning more naturally with production than direct 3D generation.EasyFashion connects user intent and body measurements to editable outputs compatible with downstream garment construction.Source: §2.2 · pp. 4–5
2.3. Human Body Reconstruction for Virtual Try-On
- Body shape and scale influence perceived fit and drape, while scanning can capture detail but requires specialized hardware, tight-fitting clothing, and post-processing.These scanning requirements limit scalability for practical personalized try-on.Source: §2.3 · p. 5
- Image-based measurement estimates can be sensitive to clothing, viewpoint, and pose, and may not preserve the scale consistency needed for SMPL-based fit evaluation.EasyFashion uses two user images plus user-provided height and gender for scale calibration, then applies an existing SMPL-compatible reshaping pipeline rather than proposing a new measurement-to-SMPL mapping.Source: §2.3 · p. 5
3. The EasyFashion System Design and Implementation
EasyFashion is a web-based system whose pipeline converts user inputs into garment specifications, personalized avatars, digital fabrics, sewing patterns, and virtual try-on. The section introduces the formative study and then the interface and technical implementation.
3. The EasyFashion System Design and Implementation
- EasyFashion is accessed through a web interface and transforms user inputs into structured garment specifications, personalized avatars, digital fabrics, sewing patterns, and virtual try-on.The internal pipeline supports personalized garment design through multimodal human-AI co-creation.Source: §3 · p. 5
- The section presents the formative study before describing the system's user interface and technical implementation.The study informed the system design.Source: §3 · p. 5
- Table 1 summarizes demographics of participants in the formative study.The supplied passage identifies the table's subject but not its participant values.Source: Table 1 · p. 5
3.1. Questionnaire Survey
A formative questionnaire and interview study examined non-professional users’ design challenges and preferences alongside professionals’ production concerns. Its findings informed requirements for expressing and refining intent, evaluating fit, and connecting design to production.
3.1. Questionnaire Survey
- The formative study combined a questionnaire and semi-structured interviews with 12 non-professional users and 6 garment-design professionals.The non-professional participants had no prior garment-design experience; the professionals had more than five years of experience.Source: §3.1 · pp. 5–6
3.1.1. Non-Professional Users’ Perspectives
- Non-professional users described customization as high-effort and high-risk, citing repeated communication and try-on, preference drift during long production cycles, and pressure in interactions with tailors.Participants said these factors could make the final garment uncertain or difficult to change.Source: §3.1.1 · p. 6
- 56% expected design to take under 10 minutes, while 65% preferred a computer, pointing to demand for efficient interaction and detailed inspection.33% expected 10–30 minutes and 11% were willing to spend more than 30 minutes.Source: §3.1.1 · p. 6
- 34% preferred reference image plus text for communicating design intent, while users reported difficulty describing construction details and found freehand sketches too abstract.Images alone and text alone were each preferred by 22% of respondents.Source: §3.1.1 · p. 7
- Participants valued on-body visualization and fabric previews, and 61% expressed interest in automatically generated sewing patterns.72% considered previewing fabric appearance important, while a majority wanted on-body visualization.Source: §3.1.1 · p. 7
3.1.2. Professional Users’ Perspectives
- Professionals said visual references can be ambiguous across body shapes and fabrics, and that try-on feedback needs body specificity and scale consistency for production-related decisions.They also emphasized alignment between try-on results, measurements, and sewing patterns.Source: §3.1.2 · p. 7
3.1.3. Design Requirements from the Formative Study
- The study distilled four design requirements: support image-and-text intent, lightweight iterative refinement, personalized fit feedback, and consistent design-to-production outputs.The system uses a persistent JSON garment specification to connect virtual try-on and sewing-pattern generation.Source: §3.1.3 · pp. 7–8
3.2. User Interface Design
The web interface organizes EasyFashion as an iterative workflow: users provide body, design, and fabric inputs, operate generation controls, and inspect intermediate and final outputs. It keeps current inputs and outputs available to support contextual refinement.
3.2. User Interface Design
- EasyFashion’s web workflow links body reconstruction, garment-intent interpretation, fabric and pattern generation, and rendering into production-oriented assets and virtual try-on.The pipeline uses a reconstructed body model with measurements and an editable JSON garment specification as intermediate representations.Source: §3.2 · p. 8
- Users revise garment intent with a reference image and short text prompt while the system preserves a JSON design state across edit, regeneration, and evaluation cycles.The text can specify complementary constraints such as collar, sleeve, or garment length.Source: §3.2 · p. 8
- The Input panel collects front and side body images, basic attributes, garment reference and description, and optional fabric image or text guidance.Source: §3.2 · p. 9
- The Output panel presents patterns, fabric, estimated measurements, and avatar simulation, while the Control panel supports repeated generation, session reset, and artifact download.Previous inputs and latest outputs remain visible for comparison and revision.Source: §3.2 · p. 9
3.3. Technical Implementation
EasyFashion combines body reconstruction, a Fashion Agent, fabric generation, parametric pattern construction, and simulation in an editable workflow. Its modules connect user inputs and body measurements to personalized try-on and production-oriented artifacts.
3.3. Technical Implementation
- The implementation integrates body reconstruction, a VLM-based Fashion Agent, parametric pattern generation, fabric generation, and simulation around explicit body-model and garment-JSON representations.The editable state can be updated through user revision requests and evaluated through on-body visualization and production-oriented outputs.Source: §3.3 · p. 10
3.3.1. Human Body Reconstruction
- Body reconstruction uses front and side images plus height and gender to estimate measurements and produce a personalized SMPL model.The reshaper optimizes the SMPL shape vector while fixing pose, minimizing weighted height, chest, waist, and hip measurement losses.Source: §3.3.1 · p. 10
3.3.2. Fashion Agent and Structured Garment Specification
- The Fashion Agent converts a reference image and text into a garment specification and, when fabric input is absent, an appearance-generation prompt.It predicts a compact attribute table, maps it into a JSON template, then validates types and ranges for pattern-generation compatibility.Source: §3.3.2 · pp. 10–11
3.3.3. Digital Fabric Generation
- Fabric appearance comes from an uploaded image or text-guided generation, which produces tileable diffuse maps and corresponding roughness and normal maps.The generation path uses a fine-tuned Stable Diffusion model when no fabric image is supplied.Source: §3.3.3 · p. 11
3.3.4. Personalized Sewing Pattern Generation
- The pattern generator combines garment style attributes with reconstructed-body measurements to create customized sewing patterns and assembly metadata.Because patterns derive from the editable JSON and personalized measurements, updates preserve a link among design state, try-on, and production outputs.Source: §3.3.4 · pp. 11–12
3.3.5. Rendering and Simulation
- Rendering assembles patterns and simulates the garment on the reconstructed SMPL avatar using materials from uploaded fabric or generated fabric maps.Generated artifacts can be exported through the interface for downstream communication and garment production.Source: §3.3.5 · p. 12
4. Technical Evaluation
The technical evaluation tests body reconstruction and whether the Fashion Agent interprets garment intent across input modalities. It examines measurement accuracy, visual plausibility, pattern structure, and downstream rendering.
4.1.1. Dataset Preparation
- The body-reconstruction test used 42 participants, with scan-derived measurements as ground truth and front- and side-view images captured for evaluation.Participants wore tight-fitting clothing or no upper clothing during scanning; predicted measurements were aligned to the scan software’s circumference definitions.Source: §4.1.1 · p. 12
4.1.2. Qualitative Evaluation
- Reconstructed female and male bodies were visually consistent with their corresponding scans, suggesting plausible 3D geometry from lightweight image inputs.Source: Figure 6 · p. 13, §4.1.2 · p. 13
4.1.3. Quantitative Evaluation
- 99.17% for female height and 98.91% for male height; male waist was lowest at 85.60%.Female accuracies were 99.17% for height, 91.55% for chest, 91.62% for waist, and 95.18% for hip; male accuracies were 98.91% for height, 90.95% for chest, 85.60% for waist, and 96.63% for hip.Source: Table 2 · p. 13, §4.1.3 · p. 13
4.2.1. Dataset Preparation
- The Fashion Agent’s garment specifications were evaluated through sewing-pattern structure and visual consistency of rendered results with intended designs.The evaluation used 64 GarmentCode samples, with an experienced designer providing ground-truth patterns and descriptions.Source: §4.2 · p. 13, §4.2.1 · p. 14, §4.2.3 · p. 15
4.2.2. Qualitative Evaluation
- Across hoodie, pants, and dress examples, reference image + text most consistently aligned pattern structures and rendered appearances with the intended designs.Other conditions omitted or distorted garment structure, proportions, or design details in some examples; text-only could encode a wavy skirt but yield dimensional inconsistencies that prevented successful try-on rendering.Source: Figure 7 · p. 14, §4.2.2 · p. 14
4.2.3. Quantitative Evaluation
- 0.83 panel accuracy and 0.57 edge accuracy make reference image + text the best-performing input condition.The evaluation compares generated sewing-pattern structure across five input conditions.Source: Figure 8 · p. 15, §4.2.3 · p. 15
4.3. System Generalization Capability
EasyFashion’s examples span diverse body shapes and garment styles, and its personalized garments are compared with standard sizes. The comparison reports fit and pressure-map differences for two representative body types.
4.3. System Generalization Capability
- Across examples, outputs remained visually coherent and aesthetically appropriate across body proportions, style preferences, and garment categories, though this evidence is qualitative and preliminary.Source: §4.3 · p. 15, Figure 9 · p. 16
4.4. Design Method Comparison
- The comparison adapted the Chinese women’s ready-to-wear size designation standard to the garment categories in the study, whose standard labels ranged from S to XXL.Source: §4.4 · p. 16, Figure 10 · p. 16
- For a user facing hip-fit and length trade-offs in XL or XXL, the personalized garment jointly addressed width and length requirements and showed fewer localized high-pressure regions.Source: §4.4 · p. 16, Figure 11 · p. 17
- For a user whose main issue was garment length, larger ready-to-wear sizes could still introduce tightness or silhouette distortion; personalization followed body proportions instead.Pressure maps indicated reduced tightness alongside improved length.Source: §4.4 · p. 17
5. User Evaluation
The user evaluation tested EasyFashion with non-professional users through two sequential studies and a production case, assessing input conditions, iterative use, and real-world fabrication.
5.1. Participants
- The study recruited 30 participants, most of whom reported no prior garment design or production experience.Participants were evenly split by reported sex; 70% reported no prior experience, 13% relevant experience, and 17% selected neutral.Source: §5.1 · p. 18, Figure 12 · p. 18
5.2. Experimental Design
- Study A compared five input modalities using SYSUSE and PEU, while Study B evaluated the reference image + text setting using five rating dimensions.Study A used a within-subjects design; Study B used the same questionnaire in a more open-ended iterative task, with lower scores indicating better ratings.Source: §5.2 · pp. 18–19
5.3. User Experimental Procedure
- Study A allowed one generation per condition, whereas Study B let participants revise text and images iteratively until results reasonably matched their intent.Participants typically reached a satisfactory garment in 2 to 3 rounds and rarely exceeded five; generation took approximately 117 seconds per run.Source: §5.3 · p. 19, Figure 13 · p. 19
5.4. Data Analysis
- Quantitative ratings used 7-point Likert scales, with Friedman tests and Holm-corrected pairwise comparisons for Study A and descriptive statistics for Study B.Qualitative findings came from open-ended responses and brief post-task interviews.Source: §5.4 · pp. 19–20
5.5. Experimental Results
- Reference image + text achieved the lowest median scores for SYSUSE (1.67) and PEU (2.00) in Study A; Study B ratings were favorable across all five dimensions.Study B means were INFOQUAL 1.84, PEU 1.96, OBAU 1.93, SYSUSE 1.65, and Overall Satisfaction 1.73; lower scores indicate more favorable ratings.Source: §5.5 · p. 20, Figure 14 · p. 20
- Interview feedback described refinement as inspecting try-on outputs, revising intent, and regenerating, with participants reporting satisfactory results within few rounds.Participants especially valued combining a reference image with text for localized modifications; some reported weaker performance on layered garments, less realistic fabric appearance, or a desire for more avatar customization.Source: §5.5 · p. 20
- Participants valued production-oriented outputs, and the case study illustrates a handoff from virtual try-on and patterns to a sewn garment.The interview feedback included a sewing hobbyist who said patterns could help compensate for limited pattern-drafting skills.Source: §5.5 · p. 21, Figure 15 · p. 21
5.6. Production Case Study
- A production case showed that one participant moved from a single-round design to a dress made from EasyFashion’s generated patterns and accompanying materials.An independent tailor proceeded from the A4-tiled patterns and measurement sheet without further verbal explanation; the participant reported good fit and preferred length, while noting that her stiff fabric choice was not ideal for the style.Source: §5.6 · pp. 21–22, Figure 15 · p. 21
6. Discussion
The discussion interprets EasyFashion’s findings across measurement reliability, multimodal expression, fit, iteration, and production readiness. It also identifies boundaries where verification and broader validation remain important.
6.1. Measurement Reliability for Fit Decisions
- Image-based measurements are more dependable for broad sizing than for waist-sensitive shaping, where posture, clothing, and camera angle can affect estimates.The authors suggest verifying key circumferences and distinguishing stable from uncertain dimensions for users.Source: §6.1 · p. 22
6.2. Multimodal Constraints for Expressive Control
- Reference images anchor silhouette and style, while text clarifies what to preserve, change, or prioritize; conflicting signals can still yield unexpected interpretations.The discussion recommends structuring the image as a stable reference and text as concise change instructions.Source: §6.2 · p. 22
6.3. Robustness for Body Diversity
- Coherent outputs across demonstrated body shapes and garment types do not establish robustness to everyday variation in photos or unusual garment designs.The authors identify lighting, pose, clothing, seam placement, and extreme silhouettes as potential edge cases requiring verification or conservative handling.Source: §6.3 · p. 22
6.4. Personalization Benefits for Better Fit
- Measurement-conditioned patterns shift fit decisions from choosing a nearest standard size toward fitting an individual body, but virtual try-on and pressure maps do not fully establish comfort.The authors characterize personalization as improving fit plausibility and reducing obvious mismatches, while recommending refinement and broader wear testing.Source: §6.4 · pp. 22–23
6.5. Ease of Use for Idea Expression
- Reference image + text received the best Study A usability and ease-of-use ratings, while unclear edit effects or controllable parameters can undermine expression.The discussion recommends keeping the action space small, showing effective short edits, and making changes across iterations visible.Source: §6.5 · p. 23
6.6. Rapid Convergence for Iterative Design
- Most participants reached an acceptable design within few refinement rounds, but rapid convergence may indicate premature stopping as well as efficient progress.The discussion recommends keeping iteration lightweight while supporting continued refinement and verification of critical fit or construction details.Source: §6.6 · p. 23
6.7. Pattern Outputs for Production Readiness
- The production case suggests sewing patterns and measurement summaries can serve as production inputs, although one successful fabrication does not establish broad production readiness.Broader validation would involve more garments, materials, makers, and production contexts.Source: §6.7 · p. 23
6.8. Design Implications for Human-AI Co-Creation
- The design implications emphasize transferable artifacts, complementary multimodal constraints, and iterative decisions supported by feedback users can act on.Outputs such as structured specifications, measurements, and production templates can support communication beyond the system; fit or constraint-satisfaction cues can inform refinement.Source: §6.8 · p. 24
7. Limitations and Future Work
EasyFashion’s limitations span interaction and technical scope: visual outputs and AI-generated proposals may be misread, multimodal input can add cognitive load, and fit and garment coverage remain constrained. Future work includes clearer uncertainty communication, input scaffolding, richer body representations, and broader garment data.
Interaction and user-centered limitations
- Visual try-on may create false certainty because subtle fit issues and construction constraints are not fully captured.The authors propose communicating uncertainty and constraint violations, and supporting comparisons across design iterations.Source: §7 · p. 24
- Users may mistake generated results for definitive recommendations, although body-derived measurements are more stable than reference-inferred style details and rendering or simulation artifacts.The authors propose summaries of dimensions, structure, and design assumptions so users can validate intent and detect mismatches early.Source: §7 · p. 24
- Multimodal input may increase cognitive load when users are unsure what to provide in each modality or how modalities are combined.Proposed scaffolding includes guided templates, effective-prompt examples, and progressive disclosure when ambiguity or conflicting constraints are detected.Source: §7 · pp. 24–25
Technical and dataset limitations
- Body models use a small set of measurements, omitting finer-grained differences that may affect fit, and robustness under varied lighting, clothing thickness, and pose remains to be examined.The authors suggest additional measurements or richer body-shape representations to improve reconstruction and fit quality.Source: §7 · p. 25
- Garment styles are limited by GarmentCode dataset coverage, leaving underrepresented features such as some collar variations and advanced pattern details less supported.The authors propose broader style categories, finer-grained pattern annotations, and constraint-aware editing within manufacturable boundaries.Source: §7 · p. 25
8. Conclusion
EasyFashion supports personalized garment design from multimodal inputs, combining body reconstruction and parametric sewing patterns with visual previews and production-oriented outputs. Technical evaluations, user studies, and a production case study report iterative design refinement, fit improvements beyond standard sizing, and remote workflows using deliverables shared with makers.
8. Conclusion
- EasyFashion combines multimodal intent understanding, personalized body reconstruction, and parametric sewing-pattern generation for personalized garment design.The system provides visual previews and production-oriented outputs for non-professional users.Source: §8 · p. 25
- Technical evaluations, user studies, and an end-to-end production case study show support for iterative design refinement, improved fit beyond standard sizing, and remote workflows through deliverables shared with makers.The authors highlight multimodal inputs as complementary constraints and editable intermediate representations as important for connecting body-specific evaluation with manufacturable outcomes.Source: §8 · p. 25
Declaration of generative AI and AI-assisted technologies
The authors report using AI tools for language and submission preparation, explanatory illustrations, and the Fashion Agent’s vision-language model.
Declaration of generative AI and AI-assisted technologies
- The authors used ChatGPT/Codex for language polishing and submission tasks, generative AI tools for explanatory illustrations, and Gemini-2.5-Flash as the Fashion Agent’s vision-language model.They state that they reviewed and edited the content and take responsibility for the publication.Source: Declaration of generative AI and… · p. 25