Reproduce “Stylistic Attribute Control in Latent Diffusion Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Max Reimann, Benito Buchheim, Jürgen Döllner
Researchers from Hasso-Plattner-Institute developed a method for fine-grained, parametric control over stylistic attributes within latent diffusion models, allowing continuous adjustment of elements like contour width and local contrast. The approach leverages specialized ControlNet fine-tuning and guidance composition to achieve smooth, disentangled stylistic edits, with user studies showing a 50% preference for its outputs over instruction-based methods.
Key insights
- Parametric control over specific stylistic attributes can be effectively learned and disentangled from content by integrating a fine-tuned ControlNet and composing its guidance.
- Using an additional attention transformer block for parameter embedding and regularization during training is crucial for achieving stable, continuous, and localized stylistic changes.
- Adapting null-text inversion to optimize null-parameter embeddings enables robust and content-preserving stylistic editing of real images.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.