SliderSpace: Decomposing the Visual Capabilities of Diffusion Models

Project Website | Paper | Trained Weights | Huggingface Demo

Official implementation of "SliderSpace: Decomposing the Visual Capabilities of Diffusion Models" Unlock the creativity of diffusion models and get sliders to control the creative concepts!

Setup

To set up your python environment:

conda create -n sliderspace python=3.10
conda activate sliderspace

git clone https://github.com/rohitgandikota/sliderspace.git
cd sliderspace
pip install -r requirements.txt

If you want to use Claude to expand your training prompts, you need to export your ANTHROPIC_API_KEY (ideally to your bashrc or windows environment variables)

SDXL Sliderspace

To discover SliderSpace directions for a concept inside SDXL (and distilled versions) use the following script

python train_sdxl.py \
    --concept_prompts "picture of a spaceship" \
    --exp_name "spaceship" \
    --distilled_ckpt "dmd2" \
    --num_sliders 10 \
    --clip_total_samples 1000 \

You can train 32 sliders under 90 mins on A6000 GPU! For comparison training 16 Concept Sliders can take upto 480 mins. For more stable discovery increase the --clip_total_samples 10000 this will slow down the discovery process, but you will get far robust directions.

FLUX Sliderspace

To discover SliderSpace directions for a concept inside FLUX models use the following script

python train_flux.py \
    --concept_prompts "picture of a wizard" \
    --exp_name "wizard" \
    --num_sliders 10 \
    --clip_total_samples 1000 \
    --clip_batch_size 1 \ 

FLUX discovery takes around ~120 mins on A100 GPU to discover 32 directions. Increase clip_batch_size if you have enough VRAM. For more stable discovery increase the --clip_total_samples 10000 this will slow down the discovery process, but you will get far robust directions.

Evaluation and Inference with the Trained SliderSpace

Once you train your SliderSpace using about training scripts, you can run inference on the discovered sliders and discover the directions you trained:

• notebooks/sdxl-inference.ipynb: This is a simple and clean notebook for inference on SDXL sliderspace. You can make GIFs too!
• notebooks/flux-inference.ipynb: This is a simple and clean notebook for inference on FLUX sliderspace. You can make GIFs too!

Citing our work

The preprint can be cited as follows

@inproceedings{gandikota2025sliderspace,
  title={SliderSpace: Decomposing the Visual Capabilities of Diffusion Models},
  author={Gandikota, Rohit and Wu, Zongze and Zhang, Richard and Bau, David and Shechtman, Eli and Kolkin, Nick},
  journal={arXiv preprint arXiv:2502.01639},
  year={2024}
}