instantX-research/InstantStyle
A diffusion model framework for style-preserving text-to-image generation using CLIP feature subtraction to separate style and content.

Velocity · 7d
+2.2
★ / day
Trend
→steady
star history
InstantStyle is a text-to-image generation framework that achieves effective disentanglement of style and content from reference images. It uses two techniques: subtracting content text features from image features to decouple style, and injecting style information into specific attention layers. The method aims to mitigate content leakage while preserving the artistic style of reference images.