DReSG: Diffusion Residuals for Stylized Gaussian Splatting

DReSG is a new stylized 3D Gaussian Splatting pipeline built around diffusion residuals rather than direct per-view diffusion guidance. The core idea is to treat attention-guided diffusion outputs as residual targets against the current render, then absorb those updates back into the shared Gaussian representation through multi-view feedback.

That matters because stylizing 3DGS scenes has a long-standing tradeoff: feature-based methods tend to be stable but can miss the stronger cues in a reference style, while diffusion guidance can push appearance harder but often introduces view drift, local artifacts, and inconsistent results across angles. DReSG is designed to sit in the middle, keeping the scene grounded in 3D while still borrowing the richer priors of diffusion models.

To keep the feedback loop controllable, the system modulates residual strength during target construction, then uses coverage-aware view selection and conflict-filtered color updates during fitting. In practical terms, that means the method is trying to avoid over-updating the same regions from conflicting viewpoints, which is exactly the kind of issue that can make stylized reconstruction unstable in production workflows.

The authors report competitive reference-guided stylization with better preservation of scene structure and cross-view stability. For teams experimenting with 3DGS-based content creation, this is another sign that the next wave of tools may come from hybrid pipelines that combine explicit 3D representation with diffusion-driven appearance editing rather than relying on either one...