Concept Steering

Once we've identified interpretable concepts, we can steer them: increasing or decreasing their intensity to control specific artistic attributes in generated images. This steering is performed by scaling the activation of individual SAE features during image generation, allowing for fine-grained, interpretable control over style without any model fine-tuning.

Below we demonstrate concept steering with interactive sliders. For each concept, you can adjust the steering intensity (coefficient $\alpha$) and see how it affects the generated image. The prototype image shows the concept applied to a white sphere.

Packaging Concepts into Style Profiles

Through our operational definition of style as recurring activation patterns across semantically diverse images, we construct style profiles: compact representations that capture an artist's unique aesthetic. These profiles enable efficient style representation and style transfer. Given a content image and a reference style profile, LouvreSAE transfers the style without any model fine-tuning, additional training, or optimization.

Below we compare LouvreSAE against state-of-the-art baselines including B-LoRA, InstantStyle, and StyleShot. For each artist style, we show a reference sample, a content image, and the results from each method.

Quantitative Results

We evaluate LouvreSAE against state-of-the-art baselines on the ArtBench10 dataset, measuring both style fidelity and content preservation. Our method achieves the best VGG Style Loss and competitive CLIP Score for style, while maintaining strong content preservation. Critically, LouvreSAE is dramatically faster: style extraction takes only 6 seconds (compared to 660 seconds for B-LoRA), and image generation is comparable to other methods.

Quantitative Comparison with Baselines. LouvreSAE achieves the best style transfer performance (lowest VGG Style Loss, highest CLIP Score for style) while being dramatically faster. Style extraction assumes 40 training images. Runtimes measured on NVIDIA A100 SXM 40GB.

More Results & Discussion

LouvreSAE demonstrates that interpretable, efficient style transfer is achievable through careful decomposition of latent representations. Our operational definition of style—as recurring activation patterns across semantically diverse content—provides a principled foundation for understanding and manipulating artistic aesthetics in generative models.

Beyond the quantitative improvements, the interpretability of our approach offers unique advantages. Artists and researchers can inspect individual concepts, understand which stylistic elements are being transferred, and selectively enable or disable specific attributes. This level of control and transparency is difficult to achieve with end-to-end fine-tuning approaches.

Below we show additional qualitative results across various artists, demonstrating the versatility and robustness of LouvreSAE across different artistic styles and content types.

Future Directions

Several exciting directions remain for future work. First, extending our approach to other generative architectures (e.g., Stable Diffusion, DALL-E) could broaden its applicability. Second, developing user interfaces that allow artists to directly manipulate style profiles could enable new creative workflows. Third, investigating how style concepts compose and interact could lead to even finer-grained control. Finally, studying how our operational definition of style relates to perceptual and art historical theories could deepen our understanding of artistic representation in neural networks.

We believe LouvreSAE represents a step toward more interpretable, efficient, and controllable style transfer—one that respects both the technical constraints of modern AI systems and the nuanced, multifaceted nature of artistic style itself.

Citation

@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}
  }

Acknowledgments

This work was supported by a Brown Institute Magic Grant. The authors thank Yael Vinker for helpful discussions.