SliderSpace realizes CLIP-PCA directions as causally independent LoRA sliders

measured in 1 paper

Gandikota et al. run PCA on CLIP embeddings of a diffusion model's own output samples, yielding orthogonal principal directions of its semantic variation [gandikota-etal-2025-sliderspace-diffusion-decomposition] Per direction, a low-rank LoRA adapter is trained via a cosine-alignment loss to make its induced CLIP-embedding shift align with that PCA direction, turning passive geometry into a reusable slider [gandikota-etal-2025-sliderspace-diffusion-decomposition] Demonstrated on SDXL-DMD, SDXL, SDXL-Turbo, and FLUX Schnell, sliders increase diversity at similar CLIP-Score with user-study win-rates 66-72% [gandikota-etal-2025-sliderspace-diffusion-decomposition] An ablation shows the cosine-alignment loss and the CLIP-space choice are both causally necessary, else the sliders collapse to junk or color/shape-only directions [gandikota-etal-2025-sliderspace-diffusion-decomposition]

Structure

Linear Direction

Context

PCA on CLIP embeddings of diffusion-model output samples (orthogonal semantic directions), LoRA weight-space adapters trained per-direction via cosine-alignment loss, ablation showing the alignment loss and CLIP-space choice are both causally necessary, cross-model demonstration (SDXL-DMD, SDXL, SDXL-Turbo, FLUX Schnell)

Confirmed in models

SDXL-DMD (4-step distilled) · Stable Diffusion, Stable Diffusion XL · Stable Diffusion, SDXL Turbo · Stable Diffusion, FLUX.1 [schnell] · FLUX

Method

PCA, LoRA-trained direction realization

Papers

SliderSpace: Decomposing the Visual Capabilities of Diffusion Models — Gandikota, Rohit, Wu, Zongze, Zhang, Richard, Bau, David, Shechtman, Eli, Kolkin, Nick2025 · arXiv:2502.01639