UniFixer: Refiner for Diffusion View Synthesis

• UniFixer is a universal refiner that post-processes degraded novel views using a reference-guided, coarse-to-fine pipeline.
• It employs a three-stage framework—Reference Pre-Alignment, Global Structure Anchoring, and Local Detail Injection—to correct spatial, temporal, and backbone-induced degradations.
• Empirical results demonstrate state-of-the-art improvements in novel view synthesis and stereo conversion over existing fixers without altering upstream synthesis models.

Searching arXiv for "UniFixer" and related papers. UniFixer is a universal, reference-guided refiner for diffusion-based view synthesis that post-processes degraded novel views rather than replacing the upstream synthesis model. It is designed to fix diverse degradation artifacts—particularly blurred details and distorted structures—by using a high-quality reference view from the same scene in a coarse-to-fine pipeline. The framework is described as plug-and-play, achieves zero-shot fixing across different types of diffusion degradation, and is evaluated on novel view synthesis and stereo conversion tasks (Chen et al., 12 May 2026).

1. Problem setting and degradation model

UniFixer is motivated by the observation that diffusion-based view synthesis has become mainstream in two major forms: explicit Depth-Warp-Inpaint pipelines and implicit end-to-end generation. The paper identifies two main causes of output degradation in these systems: pixel-to-latent compression, which suppresses fine textures because diffusion models usually operate in a VAE latent space, and diffusion hallucination, which can invent details inconsistent with the scene and thereby introduce blurry textures, patch inconsistency, or structural distortion (Chen et al., 12 May 2026).

The degradation is organized along three dimensions. Spatial degradation is associated with image resolution and latent compression; lower-resolution inputs produce over-smoothed textures and weaker edges, while higher resolution does not eliminate hallucination artifacts. Temporal degradation arises in video-based or multi-frame synthesis when large camera motion or large frame stride weakens inter-frame correspondence, causing over-smoothing or inconsistent hallucinated details across frames. Backbone-related degradation depends on architectural choice: UNet backbones tend to over-smooth after repeated down/up-sampling, whereas DiT backbones can produce patch inconsistency or hallucinated structures, especially under domain shift. The authors further report that these degradations occupy distinct clusters in a t-SNE feature space.

This formulation situates UniFixer as a degradation-correction stage rather than a primary renderer. A plausible implication is that the method is intended to be architecture-agnostic at the synthesis stage, provided that a degraded target view and a suitable reference view are available.

2. Coarse-to-fine framework

UniFixer takes a degraded novel view $I_{deg}$ and a reference image $I_{ref}$ and predicts a refined output $\hat{I}$. Its pipeline has three stages: Reference Pre-Alignment (RPA), Global Structure Anchoring (GSA), and Local Detail Injection (LDI). The design is explicitly coarse-to-fine: first reduce large spatial mismatch, then fix structural layout, then restore local details (Chen et al., 12 May 2026).

Stage	Function	Core mechanism
RPA	Coarse alignment	Warping the reference toward the degraded view
GSA	Structural correction	Attention over degraded and warped-reference latents
LDI	Texture restoration	Multi-scale reference feature injection with deformable alignment and gated fusion

The framework is implemented as a refiner applied after an upstream diffusion-based view synthesis model. The upstream model can be explicit NVS, explicit stereo conversion, implicit NVS, or implicit stereo conversion. The paper emphasizes that UniFixer does not require altering the upstream model or retraining it.

This organization is central to the method’s claim of universality. The paper’s conceptual explanation is that existing fixers often work only in latent or low-resolution feature space, which creates an information bottleneck, whereas UniFixer combines pixel-space alignment, latent-space structure anchoring, and multi-scale detail transfer.

3. Reference Pre-Alignment

The first stage addresses the viewpoint gap between the reference and the degraded output. RPA computes a coarse warp of the reference toward the target according to

$I_{warped} = \mathcal{W}(I_{ref}, T),$

where $\mathcal{W}(\cdot)$ is a warping operator and $T$ is the view transformation map (Chen et al., 12 May 2026).

The implementation depends on the upstream synthesis paradigm. For explicit methods, geometry is already estimated, so the system uses depth or disparity together with camera pose or perspective projection to warp the reference to the target view. For implicit methods, no explicit geometry exists; instead, the paper estimates optical flow between the source/reference and synthesized target using SEA-RAFT and applies flow-based warping.

RPA is presented as a way to transform a difficult global correspondence search into a more local refinement problem. The paper also notes that this stage preserves raw pixel details better than latent-space reference fusion. In effect, RPA establishes the coarse geometric support on which the subsequent stages operate.

4. Global Structure Anchoring and Local Detail Injection

After pre-alignment, UniFixer corrects global structure inside a pre-trained one-step diffusion model, SD-Turbo. The latent encoder $\mathcal{E}$ produces encoded representations for the degraded and warped images:

$$\mathbf{z} = [\mathcal{E}(I_{deg}), \mathcal{E}(I_{warped})] \in \mathbb{R}^{2C \times H \times W}.$$

These latents are reshaped into a token sequence of length $2HW$, and the anchoring operation is

$$\mathbf{z} := \mathbf{z} + \text{Softmax}\left(\frac{QK^T}{\sqrt{d}}\right)V,\quad Q,K,V=\text{Linear}(\mathbf{z}).$$

The degraded view acts as an anchor through the residual connection, while joint attention over degraded and warped-reference tokens allows the model to aggregate shared structural information, reduce drift in layout and color, correct geometry-like distortions, and preserve scene semantics (Chen et al., 12 May 2026).

The LDI module addresses the loss of fine texture in the latent bottleneck by injecting multi-scale encoder features into the decoder pathway. With frozen encoder features

$I_{ref}$0

LDI applies two subcomponents.

First, Adaptive Warping Deformation (AWD) predicts offsets and a modulation mask:

$I_{ref}$1

followed by deformable refinement

$I_{ref}$2

This is used to correct small residual misregistrations, especially around occlusion boundaries, thin structures, and small pose discrepancies.

Second, Uncertainty-Aware Gated Fusion (UAGF) predicts a confidence map

$I_{ref}$3

and fuses features through

$I_{ref}$4

The fused features are then injected into the decoder:

$I_{ref}$5

where $I_{ref}$6.

The intended interpretation is explicit in the paper: if the degraded view is more trustworthy, the gate favors $I_{ref}$7; if the reference details are reliable, it favors $I_{ref}$8. This prevents unreliable warped regions from being blindly injected.

5. Training protocol and empirical evaluation

The training set is curated from the first 1K scenes of DL3DV. For each scene sequence, the procedure is: estimate geometry and camera poses using DA3, select the middle frame as reference, warp the reference to all other views, and generate degraded outputs by applying diffusion synthesis with VACE. The training tuples are

$I_{ref}$9

and the loss is

$\hat{I}$0

with $\hat{I}$1 (Chen et al., 12 May 2026).

Optimization uses Adam with learning rate $\hat{I}$2, batch size 1, patch size $\hat{I}$3, training time of approximately 50K iterations on one NVIDIA A100 40GB, with training time of approximately 14 hours. Although training is conducted under one fixed degradation setting—resolution $\hat{I}$4, temporal stride 1, and DiT backbone—the paper reports generalization to different resolutions, motion strides, backbones, and tasks.

Evaluation covers the DL3DV test benchmark, the Mono2Stereo dataset, the Spring dataset, and the Spatial Video Dataset. For explicit methods the reported metrics are PSNR, SSIM, LPIPS, DISTS, and FID; for implicit methods they are CLIP-IQA, MUSIQ, MANIQA, and FID. The main fixer baselines are W/O fixer, DIFIX3D+, and MaRINeR. The upstream synthesis models include VACE, ViewCrafter, Mono2Stereo, StereoCrafter, ReCamMaster, and StereoPilot. The paper states that DIFIX3D+ and MaRINeR are retrained on the same training data for fair comparison.

Quantitatively, UniFixer is reported to improve over the no-fixer baseline, DIFIX3D+, and MaRINeR across spatial, temporal, and backbone-related degradation on DL3DV. It also improves stereo conversion on Mono2Stereo and Spring, and improves implicit view synthesis and stereo conversion on Spring and SVD. The abstract characterizes the overall empirical outcome as state-of-the-art performance on novel view synthesis and stereo conversion.

6. Ablations, limitations, and interpretation

The ablation studies evaluate variants with only GSA, GSA + RPA, RPA + LDI, and the full model. The full model performs best. The paper’s summarized conclusions are that RPA helps because it makes correspondence search easier, GSA helps because it anchors global structure, and LDI helps because it restores local texture details and suppresses warping artifacts. A separate referencing-mechanism ablation compares warping-based transfer with skip connection, cross attention, and feature matching; the reported result is that the warping-based design performs best. Supplementary ablations further divide LDI into AWD only and AWD + UAGF, with both improving performance and the combination performing best (Chen et al., 12 May 2026).

Two limitations are emphasized. For implicit methods, flow-based warping can fail when viewpoint changes are large, because SEA-RAFT may estimate poor optical flow and yield bad pre-alignment. For explicit DWI systems, errors in depth or visibility reasoning can propagate through warping and inpainting, especially around occlusions, reflective or refractive regions, and multi-depth or semi-transparent structures. These cases can still produce corrupted warps.

The paper also addresses a common misunderstanding about reference-guided refinement: UniFixer is not described as blindly warping reference content into the target. Its gating mechanism can preserve plausible view-dependent effects already present in the upstream synthesis when the warped reference is unreliable. This suggests that the method’s contribution lies not only in transferring detail from a reference image, but in selectively arbitrating between degraded target evidence and warped-reference evidence across scales.