Concept sliders: LoRA adaptors for precise control in diffusion models

This web content introduces Concept Sliders, a method that allows artists to have precise control over attributes in generated images produced by text-to-image models. Artists often struggle to fine-tune attributes like lighting or eye size without disrupting the overall structure of the image. Concept Sliders address this issue by training low-rank adaptors on top of pretrained models. These sliders can be used to control the strength of specific concepts in the image, enabling nuanced tweaking. The content also mentions the limitations of current generative technologies, such as blurry or distorted images. The ability to control concepts during image generation and editing opens up new possibilities for creative expression. The article provides detailed information on training and using Concept Sliders, including textual and visual concepts, as well as the benefits of low-rank and disentangled sliders. It also demonstrates the improvement in image quality and the composable nature of Concept Sliders.

https://sliders.baulab.info/

To top