Overview
Research area: Computer Vision Authors: Khiem Vuong, Deva Ramanan, Srinivasa Narasimhan Published: 2025-08-26 arXiv: 2508.17625
Summary
Rendering views using 3D scene representations such as Gaussian Splatting (3DGS), Neural Radiance Fields (NeRF), meshes, or even point clouds produces artifacts when input views are sparse or target views lie far from the input. Recent work mitigates these artifacts using diffusion-based generative priors, but is specialized to individual representations and requires custom architectures or extensive retraining.
FixAnything is a single model for fixing a wide range of rendering artifacts. It repurposes a pretrained video generative model, leveraging its implicit multi-view priors with only minimal modification and lightweight finetuning.
Key ideas
- Video-to-video cleanup: Even noisily rendered sequences preserve camera motion and coarse scene structure, allowing artifact removal to be formulated as video-to-video translation.
- Masked anchoring: A binary mask indicates clean pixels, anchoring the model's output to high-quality input regions (e.g., training views) while refining the rest.
- 3D-consistency via DPO: To encourage renders that support downstream reconstruction, camera-pose accuracy (obtained via structure-from-motion) is used as a reward signal for direct preference optimization (DPO).
Results
Across four different 3D representations, FixAnything consistently improves rendering quality through lightweight fine-tuning, demonstrating that a single general video prior can replace multiple specialized refinement pipelines.
Original Abstract
Rendering views using 3D scene representations such as Gaussian Splatting (3DGS), Neural Radiance Fields (NeRF), meshes, or even point clouds produces artifacts when input views are sparse or target views lie far from the input. Recent work mitigates these artifacts using diffusion-based generative priors, but is specialized to individual representations and require custom architectures or extensive retraining. We present FixAnything, a single model for fixing a wide range of rendering artifacts. It does so by repurposing a pretrained video generative model, leveraging its implicit multi-view priors with only minimal modification and lightweight finetuning. [...]
Paper: https://arxiv.org/abs/2508.17625