PR-IQA: Partial-Reference Image Quality Assessment for Diffusion-Based Novel View Synthesis

Inseong Choi; Siwoo Lee; Seung-Hun Nam; Soohwan Song

cs.CVApr 6, 2026

PR-IQA: Partial-Reference Image Quality Assessment for Diffusion-Based Novel View Synthesis

Authors: Inseong Choi, Siwoo Lee, Seung-Hun Nam, Soohwan Song

Organizations: Dongguk University · NAVER WEBTOON AI

Abstract

Diffusion models are promising for sparse-view novel view synthesis (NVS), as they can generate pseudo-ground-truth views to aid 3D reconstruction pipelines like 3D Gaussian Splatting (3DGS). However, these synthesized images often contain photometric and geometric inconsistencies, and their direct use for supervision can impair reconstruction. To address this, we propose Partial-Reference Image Quality Assessment (PR-IQA), a framework that evaluates diffusion-generated views using reference images from different poses, eliminating the need for ground truth. PR-IQA first computes a geometrically consistent partial quality map in overlapping regions. It then performs quality completion to inpaint this partial map into a dense, full-image map. This completion is achieved via a cross-attention mechanism that incorporates reference-view context, ensuring cross-view consistency and enabling thorough quality assessment. When integrated into a diffusion-augmented 3DGS pipeline, PR-IQA restricts supervision to high-confidence regions identified by its quality maps. Experiments demonstrate that PR-IQA outperforms existing IQA methods, achieving full-reference-level accuracy without ground-truth supervision. Thus, our quality-aware 3DGS approach more effectively filters inconsistencies, producing superior 3D reconstructions and NVS results. The project page is available at https://kakaomacao.github.io/pr-iqa-project-page/.

Figures & tables

Figure 1: Overview of the proposed PR-IQA and quality-aware 3DGS. (a) Diffusion models generate novel views (pseudo-GTs) from sparse inputs, which often contain photometric or geometric artifacts. (b) We propose PR-IQA, a cross-reference method predicting a dense, pixel-level quality map from unaligned references. It produces a complete map correlating closely with FR-IQA metrics (e.g., DINOv2 feature-similarity map) without requiring a GT. (c) This quality map enables a dual-filtering strategy (image selection and pixel masking) for 3DGS training, reducing reconstruction errors and improving fidelity.

Figure 2: (a) Overview of the PR-IQA pipeline. The framework operates in two stages. First, we warp DINOv2 features from the reference $I_{r}$ to the query $I_{q}$ view via dense stereo, generating a partial quality map ( $\hat{Q}$ ) for overlapping regions. Next, a three-stream (query, reference, partial map) encoder-decoder predicts the full quality map $Q$ . (b) Architecture of the Dual-Gated Attention Block. The block sequentially applies two attention mechanisms: a Channel Attention Module (using max/avg pooling and MLP) recalibrates channels, and a Spatial Attention Module (using Q, K, V projections and softmax) provides spatial refinement. The block integrates both with normalization, residual connections ( $\oplus$ ), and an FFN. Each encoder and decoder is composed of this block.

Figure 3: Qualitative comparison of estimated quality maps from IQA methods. Colors encode estimated quality, where low-quality pixels are shown in blue and high-quality pixels are shown in red. Compared to baselines, our results (“Ours”) more faithfully recover object silhouettes and fine structures, closely matching the GT (DINOv2-SIM).

Figure 4: Qualitative comparison of rendered novel views from IQA-guided 3DGS. While baseline methods produce results with artifacts, blurring, or misaligned Gaussians, our PR-IQA-guided method (“Ours”) avoids these failure modes, yielding significantly cleaner and more coherent reconstructions.

Figure 5: Detailed architecture of the proposed model. The network employs an encoder–decoder design featuring cross- and self-attention modules, query fusion, and mask-aware pixel-shuffle downsampling. Key specifications, including stage-wise block counts, attention heads, and the status of component sharing (frozen vs. trainable), are explicitly annotated.

Figure 6: Impact of the number of reference views on IQA performance. We plot the PLCC and SRCC against the number of reference images used for evaluation. FR-IQA baselines are indicated by constant horizontal lines. The results are shown for (a) DINOv2 and (b) SSIM targets. In both scenarios, our model achieves the highest correlation among learned metrics (CrossScore, PuzzleSim) and demonstrates robustness even with a single reference image.

Figure 7: Impact of quality map fusion strategies on DINOv2 target evaluation. We evaluate the performance of four aggregation strategies (Max, Min, Median, and Mean) as a function of the number of reference images ( $N_{\text{ref}}\in[2,10]$ ).

Loss variants	Mip-NeRF 360		Tanks and Temples
Loss variants	PLCC	SRCC	PLCC	SRCC
w/o $\mathcal{L}_{\text{JSD}}$	-0.181	-0.202	-0.242	-0.274
w/o $\mathcal{L}_{\text{PLCC}}$	-0.119	-0.134	-0.147	-0.150
Full Model	0.555	0.622	0.573	0.649

Table 9: Ablation study on the contribution of loss components. We compare the full model with variants trained without the JSD loss (w/o

\mathcal{L}_{\text{JSD}}

) or without the PLCC loss (w/o

\mathcal{L}_{\text{PLCC}}

). All metrics are evaluated on the Mip-NeRF 360 and Tanks and Temples datasets using PLCC and SRCC for the target of DINOv2. Bold indicates the best performance.

Figure 8: Low-overlap qualitative results. Red region in the generated image shows overlaps of 16% (Family) and 22% (Ignatius) .

Figure 9: Quality estimation results on hallucinated non-overlapping regions (boxed) from the Barn and Garden scenes. The dashed boxes highlight unsupported areas that are visible in the generated image but not reliably matched to the reference view.

Method	Stage	Runtime (s)	Memory (MB)
PR-IQA	Feature Ext.	0.303	5509.250
	VGGT	0.207	2448.317
	Inference	0.510	530.950
3DGS	-	25.210	749.780

Table 16: Computational cost analysis. We report the averaged runtime (seconds) and memory usage (MB) for individual components of the PR-IQA pipeline and the 3DGS optimization process.

Figure 16: Qualitative comparison of 3DGS reconstruction quality. Our IQA-Guided 3DGS produces sharper geometry and more accurate textures compared to baselines by focusing computational resources on high-quality regions. Red boxes highlight representative areas where our method demonstrates superior reconstruction quality.

Explore similar work

GeoQuery: Geometry-Query Diffusion for Sparse-View Reconstruction
May 12, 2026Xiao Cao, Yuze Li, Youmin Zhang +43D Gaussian Splatting Diffusion Models for Image Restoration
University of Electronic Science and Technology of China & Rawmantic AI, China · Rawmantic AI, China · Tianjin University, China +1
BetterScene: 3D Scene Synthesis with Representation-Aligned Generative Model
Feb 26, 2026Yuci Han, Charles Toth, John E. Anderson +2Novel View Synthesis Sparse-View 3D Reconstruction
Dept. of Electrical and Computer Engineering, The Ohio State University · USACE ERDC GRL
DualDiff3D: Dual Structure-Appearance Diffusion Priors for Reliability-Enhanced 3D Gaussian Splatting
Sep 1, 2026Qian Wang, Yu Wang, Weiqi Li +43D Gaussian Splatting Novel View Synthesis
School of Electronic and Computer Engineering, Peking University, Shenzhen, China · Pengcheng Laboratory, Shenzhen, China · Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media2026 Technology, Shenzhen, China

May 12, 2026J/K move · Enter open · S save

GeoQuery: Geometry-Query Diffusion for Sparse-View Reconstruction

Xiao Cao, Yuze Li, Youmin Zhang, Jiayu Song+3

University of Electronic Science and Technology of China & Rawmantic AI, China · Rawmantic AI, China · Tianjin University, China+1

3D Gaussian Splatting Diffusion Models for Image Restoration Novel View Synthesis Sparse-View 3D Reconstruction

Open paper AskPDF arXiv

3D Gaussian Splatting (3DGS) has emerged as a prominent paradigm for 3D reconstruction and novel view synthesis. However, it remains vulnerable to severe artifacts when trained under sparse-view constraints. While recent methods attempt to rectify artifacts in rendered views using image diffusion models, they typically rely on multi-view self-attention to retrieve information from reference images. We observe that this mechanism often fails when the rendered novel views output by 3DGS are heavily corrupted: damaged query features lead to erroneous cross-view retrieval, resulting in inconsistent rendering refinement. To address this, we propose GeoQuery, a geometry-guided diffusion framework that integrates generative priors with explicit geometric cues via a novel Geometry-guided Cross-view Attention (GCA) mechanism. First, by leveraging predicted depth maps and camera poses, we construct a geometry-induced correspondence field to sample reference features, forming a geometry-aligned proxy query that replaces the corrupted rendering features. Furthermore, we design a new cross-view feature aggregation pipeline, in which we restrict the cross-view attention to a local window around each proxy query to effectively retrieve useful features while suppressing spurious matches. GeoQuery can be seamlessly integrated into existing diffusion-based pipelines, enabling robust reconstruction even under extreme view sparsity. Extensive experiments on sparse-view novel view synthesis and rendering artifact removal demonstrate the effectiveness of our approach.

Abstract

Explore similar work

GeoQuery: Geometry-Query Diffusion for Sparse-View Reconstruction

BetterScene: 3D Scene Synthesis with Representation-Aligned Generative Model

DualDiff3D: Dual Structure-Appearance Diffusion Priors for Reliability-Enhanced 3D Gaussian Splatting