FixAnything improves 3D scene renders via video generative models

Researchers have introduced FixAnything — a unified model for fixing rendering artifacts across different 3D representations, including Gaussian Splatting, NeRF, mesh, and point clouds. Instead of specialized architectures for each format, the authors use a pretrained video generation model with minimal fine-tuning, framing the task as video-to-video translation. A binary mask indicates which pixels are considered clean, and the quality of the camera pose recovered via structure-from-motion serves as a reward signal for direct preference optimization. Details are in the arXiv paper.

What it means

For Gaussian Splatting (3DGS), this is a direct target: the method is claimed to improve render quality specifically for 3DGS among other representations. In our 3DGS fact sheet, it is an open-source 3D generator from Inria and MPII, requiring 24 GB of VRAM and CUDA. Artifacts with sparse input views are a known weakness of this approach, and FixAnything offers an external cleanup layer on top of an already trained scene without changing the 3DGS pipeline itself. Important: the authors have not released code, so it is not yet possible to verify compatibility with the reference 3DGS implementation or assess real memory requirements. The 3DGS fact sheet has no data on such post-processing methods — this is a new niche not covered by the current tool description. If the method proves itself in practice, it could reduce the density requirements for input images in 3DGS and related representations. For now, this is a research work without public code, and it is too early to assess its applicability in production pipelines.