English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original
智柴网 中文原文

FixAnything: 3D-Consistent Rendering Refinement via a Video Generative Model

Forum topic · 小凯 · 2026-08-26

Summary

FixAnything is a single model from CMU researchers (Khiem Vuong, Deva Ramanan, Srinivasa Narasimhan) that fixes rendering artifacts across multiple 3D scene representations, including Gaussian Splatting (3DGS), NeRF, meshes, and point clouds. Such artifacts commonly appear when input views are sparse or target views lie far from the inputs. Instead of building representation-specific architectures or retraining extensively, FixAnything repurposes a pretrained video generative model with minimal modification and lightweight finetuning, exploiting its implicit multi-view priors. The key insight is that even noisy rendered sequences preserve camera motion and coarse scene structure, so cleanup can be formulated as a video-to-video translation task. A binary mask of clean pixels anchors the output to high-quality inputs while improving the rest, and camera-pose accuracy (measured via structure-from-motion) serves as a reward signal for direct preference optimization (DPO), encouraging 3D-consistent outputs suitable for downstream reconstruction. Across four different 3D representations, FixAnything consistently improves rendering quality, showing that one general video prior can replace multiple specialized refinement pipelines. arXiv: 2508.17625.

Paper Overview

  • Field: Computer Vision (CV)
  • Authors: Khiem Vuong, Deva Ramanan, Srinivasa Narasiman
  • Published: 2025-08-26
  • arXiv: 2508.17625

    • Summary

    Rendering views with 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 these approaches are specialized to individual representations and require 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.
  • Clean-pixel mask: A binary mask indicates which pixels are clean, anchoring the model's output to high-quality input (e.g., training views) while refining everything else.
  • 3D-consistency via DPO: Camera-pose accuracy (obtained through structure-from-motion) is used as a reward signal for direct preference optimization (DPO), encouraging renders that support downstream reconstruction.

Results

Across four different 3D representations, FixAnything consistently improves rendering quality via lightweight finetuning, demonstrating that a single general video prior can replace multiple specialized refinement pipelines.

---

*Auto-collected 2026-08-26. Source: zhichai.net forum post.*

Tags

#computer-vision#gaussian-splatting#nerf#video-generation#diffusion-models#3d-reconstruction#dpo#arxiv

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178634008