YUNWEN · Cloud Motif Generation

⌄

OVERVIEW

A generative design study exploring how cloud motifs can retain recognizable forms while reflecting intended cultural associations such as auspiciousness. I prepared training images and tags, trained LoRA models, generated motif studies and co-wrote the research manuscript as second author.

FOCUS

Cultural motif research · Training data and tags · Generative image evaluation

TOOLS & METHODS

CLIP-assisted tagging · Stable Diffusion 1.5 · SDXL · FLUX LoRA · Liblib cloud training · Visual evaluation

STATUS

Published

THE CHALLENGE

What needed to become clear.

General image-generation tools often lost the cloud-motif structure. Early training retained unwanted backgrounds and produced crowded compositions. Cultural intent needed to be translated into explicit annotations and assessed through visual judgement, without treating generated imagery as proof of cultural understanding.

THE APPROACH

How the idea was tested.

I prepared 100 training images at 1024 × 1024, cleaned backgrounds and contours, reviewed CLIP-assisted tags, removed irrelevant labels and added form, meaning and context. Iterative trials with Stable Diffusion 1.5, SDXL and FLUX led to a FLUX LoRA workflow. Generated examples were organized by cultural intent, directional composition and symmetry for visual review.

CURRENT OUTCOME

Where the project stands.

The motifs were applied to garment visualizations and a physical printed garment. Faculty review gave positive feedback on their cultural associations and forms. The research paper was accepted by DSIE 2025, with Wu Yihao as second author. Other team members designed and implemented the Yunshang clothing-customization platform; this portfolio focuses on my model-training and image-generation contribution.

PROJECT MATERIALS

Documents &
moving images.

PDF

YUNWEN 云纹生成研究作品集

5 pages

PDF

Content loads when it approaches the viewport

This file is not downloaded on the first screen, keeping the page fast.

NEXT PROJECT