This repository contains the official implementation of LouvreSAE, a lightweight, interpretable method for style representation and style transfer that leverages art-specific Sparse Autoencoders (SAEs) on CLIP embeddings. The approach operationalizes artistic style as a sparse, decomposable set of concepts, constructing "style profiles" from recurring activations across an artist's body of work. By injecting style residuals directly into the latent space of pre-trained diffusion models, the method enables precise, steerable generation without fine-tuning. Validations on ArtBench10 show it outperforms existing methods on style metrics while being 1.7–20x faster and offering granular control over specific aesthetic attributes.
Website — Paper — Models — Contact us
Pre-print. Under review.
To run LouvreSAE, set up a Python 3.10 environment and install the dependencies listed in requirements.txt. You can follow the LourveSAE_demo.ipynb.ipynb with a full demonstration of our pipeline, from SAE inference and style profile construction to style transfer.
Discovering Style Concepts in CLIP Embeddings. Our collection of LouvreSAE's (which can be downloaded from Hugging Face) decomposes dense CLIP embeddings into sparse concepts pertaining to style and art (see illustration below). We provide baseline autointerpretability in the form of VLM-generated labels, exemplars (images where the concept has a high activation), and prototypes (white spheres transformed in the direction of this concept).
Packaging Concepts into Style Profiles. Style profiles are operationalized as sets of recurring activation patterns across semantically diverse images, creating compact representations that capture an aesthetic signature while abstracting away content. To build a style profile, gather a collection of diverse images with a shared style, and perform style filtering, as outlined in LourveSAE_demo.ipynb.ipynb. We generally find the inclusion threshold of C=0.6 to be a good starting point, but you might want to experiment with this to achieve optimal results.
Performing Style Transfer. Style transfer is performed by injecting style profile residuals directly into the latent space of a pre-trained diffusion model (we use Kandinsky 2.2, but the method generalizes to any diffusion models utilizing compatible embedding priors), enabling zero-shot generation without fine-tuning. You can generate stylized images by applying a style profile to your content, or you can granularly steer specific aesthetic attributes by manually adjusting the scaling of the SAE feature activations.
Reproducing Our Paper Results. For calculation of the eval metrics (VGG Loss, LPIPS, and CLIP Score) and average runtimes, as reported in the paper, refer to the metrics and runtimes modules, respectively.
If you find our work useful, please consider citing our paper.
@article{panda2025louvresae,
title = {LouvreSAE: Sparse Autoencoders for Interpretable and Controllable Style Transfer},
author = {Panda, Raina and Fein, Daniel and Singhal, Arpita and Fiore, Mark and Agrawala, Maneesh and Bohacek, Matyas},
journal = {arXiv preprint arXiv:2512.18930},
year = {2025},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
doi = {10.48550/arXiv.2512.18930},
url = {https://arxiv.org/abs/2512.18930}
}