Follow up work - Automatic Slider Discovery

[Ranking666]

Thank you for your greate word, is there any latest related work?

[ShenZheng2000]

@Ranking666 I think "SliderSpace: Decomposing the Visual Capabilities of Diffusion Models" is a strong follow-up.

Here’s my understanding of the key differences:

ConceptSlider: Modifies a single concept (e.g., person) by enhancing one attribute (e.g., old) while suppressing its opposite (e.g., young). It focuses on binary control between two predefined attributes.

SliderSpace: Explores multiple orthogonal variations (e.g., head pose, lighting direction, and facial expression) without predefined labels. Each slider corresponds to a unique dimension discovered in the data.

[rohitgandikota]

Thanks @ShenZheng2000 for shout out!

@Ranking666, we are releasing Sliderspace code tomorrow.

The main difference is - with sliderspace you can automatically discover sliders from models knowledge. Its like asking the model to give creative sliders for a concept. So you can find 100s of automatic sliders in same time as concept sliders. You will discover so many novel and creative sliders

Project: sliderspace.baulab.info

I'll update when we release code

[rohitgandikota]

https://github.com/rohitgandikota/sliderspace

Sliderspace code is live now!

[rohitgandikota]

@ShenZheng2000 - we updated the repo. I think we released an old version. Here is the latest one - https://github.com/rohitgandikota/sliderspace