Papers

Łukasz Staniszewski, Łukasz Kuciński, Kamil Deja

Noise approximations produced with inversion methods exhibit input image patterns. We find the root cause and provide simple fix, improving inversion-based image editing, interpolation and style transfer.

Łukasz Staniszewski, Katarzyna Zaleska, Mateusz Modrzejewski, Kamil Deja

We show that various musical concepts are controlled by subset of layers in audio diffusion models and steer activations within those layers for controllable audio generation.

Łukasz Staniszewski*, Bartosz Cywiński*, Franziska Boenisch, Kamil Deja, Adam Dziedzic

By employing Activation Patching, we discover that textual content within images is controlled by less than 1% of diffusion models' parameters. We show that using those layers can improve fine-tuning, text editing and removing unsafe words from images.

Łukasz Staniszewski*, Katarzyna Zaleska*, Kamil Deja

We investigate diverse low-rank adapters initialization and merging methods in the tasks of continual object and style personalization with Diffusion Models.