Shih-Chieh Su
AttnMod is a novel technique for generating unique, unpromptable art styles by modifying the attention mechanisms within existing diffusion models.
By adjusting attention multipliers in specific UNet attention blocks during the denoising process, AttnMod enables the creation of diverse visual outputs without altering the original model weights or prompts.
This approach allows for the exploration of new artistic styles that transcend traditional influences, offering a tool for artists and researchers to experiment with style generation in a controlled and reproducible manner.
[video width="2048" height="2048" mp4="https://thecvf-art.com/wp-content/uploads/SlideOut_2048_f0.5.mp4" autoplay=true muted=true loop=true][/video]
The project page provides comparisons across different seeds and styles, and demonstrates the effects of both single and multi-block attention modifications. Comprehensive documentation, along with links to the associated paper and code, are also available.
