In this work, we propose a diffusion-based approach for generating
Thai textile designs from user-provided partial sketches. Our method
incorporates two key innovations: (1) training augmentation with
synthetically generated partial sketches that mimic human drawings,
and (2) text-based color control through user descriptions. To enhance
robustness and generalization, we employ a multi-stage training pipeline.
First, we leverage a pre-trained Stable Diffusion model. We then finetune the model on a larger dataset of Indonesian fabrics (which shares
stylistic similarities with Thai textiles) before specializing in a dataset of
Thai textiles. Additionally, we implement curriculum learning, where
the model starts with complete sketches and gradually progresses to
more challenging partial sketches. Ablation studies demonstrate the
effectiveness of our approach, yielding robust and colorful designs that
adapt to various sketch styles and color instructions. This work opens
exciting avenues for future research in AI-assisted textile design, fostering
a seamless blend of human creativity and machine intelligence.
Keywords: diffusion, pattern generation, sketch-based Image synthesis,
image to image translation
Thai textile designs from user-provided partial sketches. Our method
incorporates two key innovations: (1) training augmentation with
synthetically generated partial sketches that mimic human drawings,
and (2) text-based color control through user descriptions. To enhance
robustness and generalization, we employ a multi-stage training pipeline.
First, we leverage a pre-trained Stable Diffusion model. We then finetune the model on a larger dataset of Indonesian fabrics (which shares
stylistic similarities with Thai textiles) before specializing in a dataset of
Thai textiles. Additionally, we implement curriculum learning, where
the model starts with complete sketches and gradually progresses to
more challenging partial sketches. Ablation studies demonstrate the
effectiveness of our approach, yielding robust and colorful designs that
adapt to various sketch styles and color instructions. This work opens
exciting avenues for future research in AI-assisted textile design, fostering
a seamless blend of human creativity and machine intelligence.
Keywords: diffusion, pattern generation, sketch-based Image synthesis,
image to image translation
