Announcing "SliderSpace": An automated way to unlock diffusion model's creativity and more...

Today we are happy to announce SliderSpace: Decomposing Visual Capabilities of Diffusion Model (https://arxiv.org/abs/2502.01639)

Concept Sliders enables the users to unlock their creativity by providing slider controls, but the user needs to provide the edit/training ideas. Did you ever wonder, what does the diffusion model think of a concept? Can we unlock the creativity of the model?

With SliderSpace you can automatically decompose and extract creative 100s of sliders for any concept using literally a ``single prompt''. Lets see what a model thinks of when prompted for "toy"

We also decomposed the knowledge of "art styles" in the model and found something super cool!!

We found that the model creates its own styles!! We discovered new styles that potentially do not exist! We are not 100% sure about this as we might have to exhaustively search through the database of all artists on earth. But we can use your help. Let us know if these styles are something you recognize.

Paper: https://arxiv.org/abs/2502.01639
Project: https://sliderspace.baulab.info
Code: https://github.com/rohitgandikota/sliderspace

Edit: We are excited to announce the official code implementation of SliderSpace (https://github.com/rohitgandikota/sliderspace)