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. Our key insight is that even noisily-rendered sequences preserve camera motion and coarse scene structure, allowing cleanup to be formulated as video-to-video translation. To control what scene structure should be preserved, we introduce a binary mask denoting the clean pixels, enabling the model to anchor its output to high-quality inputs (e.g. training views) while refining the rest. To encourage FixAnything to produce 3D-consistent renderings that support downstream reconstruction, we use camera pose accuracy (recovered via structure-from-motion) as a reward signal for direct preference optimization (DPO). Across four distinct 3D representations, FixAnything consistently improves rendering quality with lightweight finetuning, demonstrating that a single generalist video prior can replace multiple specialist refinement pipelines. The simplicity of the framework enables immediate adoption of stronger future video models without architectural redesign.
Figures & tables
Figure 1: FixAnything takes a rendering video from any 3D representation — 3DGS, NeRF, mesh, or sparse point cloud — and produces a photorealistic, 3D-consistent video that preserves the camera trajectory and scene content, using a single model finetuned from a pretrained video diffusion model with minimal adaptation. Top: One frame selected from each of the four videos rendered using different 3D representations. Clean frames are the input training views (they do not need to be at the start/end). Bottom: Corresponding clean output frames showing quality, consistency & generality.
Figure 2: Overview of FixAnything. Given a rendered video from any 3D representation (3DGS, NeRF, mesh, sparse point cloud) and a binary mask indicating clean frames (e.g., training views), FixAnything produces a cleaned video that preserves scene content and camera trajectory. Top: the rendered video is encoded into latent space via a frozen VAE, which is channel-concatenated with the binary mask and noise. These inputs are denoised with a Wan2.1 DiT, adapted only through LoRA [ 10 ] , without any architectural modification. Bottom: after supervised finetuning (SFT), we further optimize the model by generating multiple candidate outputs per video with different random seeds and evaluate their geometric consistency by using COLMAP to recover camera poses. The resulting pose accuracy is used to construct preference pairs for Flow-DPO, which steers the model toward outputs that are geometrically consistent.
Figure 3: Different types of degraded renderings from the same scene, paired with ground truth (last row). First and last columns are clean training views, while intermediate frames show artifacts of varying severity. NeRF produces blur and fog (row 1), 3DGS produces floaters (row 2), meshes have holes at ambiguous depth regions (row 3), and sparse point clouds render as scattered patches at detected keypoints (row 4). Despite their visual differences, all representations preserve the underlying camera trajectory and coarse scene layout, and FixAnything trains on all four types jointly.
Figure 4: Effect of mask-aware conditioning in cleaning up 3DGS inputs . The input video (row 1) has clean frames at positions 0 and 30. Without the mask (row 2, m30=0 ), the model treats frame 30 as degraded and hallucinates over it: objects like the wet floor sign disappear (red arrows). This corruption propagates to nearby frames, causing frame 40 to lose content that should be preserved. With the mask (row 3, m30=1 ), the model preserves frame 30 and uses it as context for neighboring ones: the wet floor sign and details are retained (green arrows), matching ground truth (row 4).
Figure 5: Why geometric-aware preference optimization helps. The SFT-only model (row 2) hallucinates structures (red boxes) that shift across frames, causing SfM to recover wrong poses. After Flow-DPO (row 3), the model avoids such hallucinations and closely matches the ground truth (row 4), improving pose AUC@5° by 7.2%.
3 Views
6 Views
9 Views
Method
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Sparse-view reconstruction
3DGS [ 13 ] †
10.97
0.248
0.567
13.34
0.332
0.498
14.99
0.403
0.446
RegNeRF [ 24 ] †
11.46
0.214
0.600
12.69
0.236
0.579
12.33
0.219
0.598
FreeNeRF [ 38 ] †
10.91
0.211
0.595
12.13
0.230
0.576
12.85
0.241
0.573
DNGaussian [ 15 ] †
11.10
0.273
0.579
12.67
0.329
0.547
13.44
0.365
0.539
Table 1: Quantitative comparison on DL3DV. FixAnything uses a single model to clean up renderings from four different input representations (NeRF, 3DGS, mesh, sparse SfM points), achieving comparable quality across all four. Notably, sparse SfM point renderings perform on par with or better than 3DGS and NeRF inputs, despite providing only scattered keypoints and the sole dense visual information comes from the few clean training views that the trajectory passes through. We color each cell as best and second best . ( † indicates numbers reported by [ 20 , 37 ] )
Figure 6: FixAnything generalizes across input representations. The same model cleans up sparse COLMAP point clouds (top) and meshes (bottom), producing photorealistic output despite minimal visual input except for the clean first and last views.
Figure 7: Qualitative comparison with Difix3D and Difix3D+ [ 35 ] . Difix3D enhances each frame independently using the nearest training view then distills results into 3DGS, introducing inconsistencies. Difix3D+ applies an additional enhancement step at render time, sharpening details but amplifying these inconsistencies. By processing the full video at once, FixAnything produces temporally coherent output.
Table 9Table 10
Figure 8: Uncertainty estimation. Uncertainty maps of mesh rendering cleanup correlates well with reconstruction error and with unobserved (black) input regions.
Per-scene optimization methods such as 3D Gaussian Splatting provide state-of-the-art novel view synthesis quality but extrapolate poorly to under-observed areas. Methods that leverage generative priors to correct artifacts in these areas hold promise but currently suffer from two shortcomings. The first is scalability, as existing methods use image diffusion models or bidirectional video models that are limited in the number of views they can generate in a single pass (and thus require a costly iterative distillation process for consistency). The second is quality itself, as generators used in prior work tend to produce outputs that are inconsistent with existing scene content and fail entirely in completely unobserved regions. To solve these, we propose a two-stage pipeline that leverages two key insights. First, we train a powerful bidirectional generative model with a novel opacity mixing strategy that encourages consistency with existing observations while retaining the model's ability to extrapolate novel content in unseen areas. Second, we distill it into a causal auto-regressive model that generates hundreds of frames in a single pass. This model can directly produce novel views or serve as pseudo-supervision to improve the underlying 3D representation in a simple and highly efficient manner. We evaluate our method extensively and demonstrate that it can generate plausible reconstructions in scenarios where existing approaches fail completely. When measured on commonly benchmarked datasets, we outperform all existing baselines by a wide margin, exceeding prior state-of-the-art methods by 1-3 dB PSNR.
Riccardo de Lutio, Tobias Fischer, Yen-Yu Chang +7
NVIDIA, USA · NVIDIA, Switzerland · ETHZ, Switzerland +4
3D Gaussian Splatting (3DGS) delivers high-fidelity real-time rendering but suffers from geometric and photometric degradations under sparse-view constraints. Current generative restoration approaches are often limited by insufficient temporal coherence, a lack of explicit spatial constraints, and a lack of large-scale training data, resulting in multi-view inconsistencies, erroneous geometric hallucinations, and limited generalization to diverse real-world artifact distributions. In this paper, we present ArtifactWorld, a framework that resolves 3DGS artifact repair through systematic data expansion and a homogeneous dual-model paradigm. To address the data bottleneck, we establish a fine-grained phenomenological taxonomy of 3DGS artifacts and construct a comprehensive training set of 107.5K diverse paired video clips to enhance model robustness. Architecturally, we unify the restoration process within a video diffusion backbone, utilizing an isomorphic predictor to localize structural defects via an artifact heatmap. This heatmap then guides the restoration through an Artifact-Aware Triplet Fusion mechanism, enabling precise, intensity-guided spatio-temporal repair within native self-attention. Extensive experiments demonstrate that ArtifactWorld achieves state-of-the-art performance in sparse novel view synthesis and robust 3D reconstruction. Code available at: https://github.com/fyting/ArtifactWorld.
3D scene representations like NeRF and 3D Gaussian Splatting (3DGS) suffer severe artifacts in sparse-view settings. Recent generative 3D artifact fixers attempt to address this, but rely on paired corrupted and clean renders requiring costly, per-scene reconstructions across varying view configurations. While 2D image augmentations act as instant regularizers, no explicit equivalents exist for 3D representations to preserve spatial consistency across views, an essential property for 3D-aware training. We propose 3D Morphological Perturbations as an optimization-free regularizer that preserves spatial consistency. Leveraging explicit 3DGS, we treat each Gaussian as a fundamental building block - analogous to a 2D pixel - and apply perturbations across its morphological parameter space via scale, rotation, and pruning. Our method eliminates per-scene 3DGS optimization loops from dataset curation while enabling models to learn stronger geometric priors than sparse-view baselines in diagnostic ablations conducted on a lightweight video diffusion sandbox. Scaled to a 14B-parameter video model via ControlNet, our approach maintains visual fidelity while reducing mean depth error by 12.5% over state-of-the-art image-to-image 3D artifact refiners, ultimately boosting downstream robotics policy success rates by up to 8.0% across 3 of 4 manipulation tasks.
Onat Şahin, Mohammad Altillawi, George Eskandar +2
Technical University of Munich · Huawei Heisenberg Research Center · Tavus