Reproduce “OmniStyle-INR: Universal and Multimodal Style Transfer for INRs” 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
Rafał Kajca, Michał Miziołek, Kornel Howil, Rafał Tobiasz, Przemysław Spurek
OmniStyle-INR provides a shared CLIP- and VGG-guided stylization procedure for images, videos, static 3D scenes, and dynamic 4D neural representations. By combining global and patch-level style objectives with content, background, and optical-flow consistency losses, it preserves scene structure while enabling image- or text-conditioned editing; on DAVIS videos, it improves long-range temporal consistency over CLIPGaussian, with RMSE of 0.036 versus 0.095 under text conditioning.
Key insights
- Patch-level CLIP supervision is central to distributing style across the scene: in the 4D ablation, removing it reduced CLIP-SIM from 19.14 to 1.79.
- The method trades some style strength for structural fidelity. On DAVIS, it slightly trails CLIPGaussian on direct text-style similarity but achieves stronger directional style and content-fidelity scores, while producing lower flow-based temporal errors.
- Temporal regularization is not uniformly beneficial at high weight: overly strong temporal penalties can suppress stylization, so consistency and appearance quality depend on careful loss balancing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.