Partial-Reference Image Quality Assessment

• PR-IQA is an image quality assessment paradigm that infers quality from incomplete, non-pose-aligned reference images in novel view synthesis.
• It employs geometry-aware feature warping with DINOv2 cosine similarity to compute reliable partial quality maps and guide dense quality estimation.
• The integrated completion module and masking strategies improve 3D reconstruction fidelity by restricting supervision to high-confidence regions.

Partial-Reference Image Quality Assessment (PR-IQA) is an image quality assessment paradigm in which quality is inferred from incomplete, indirect, or non-pose-aligned reference information rather than from a pristine image at the same viewpoint. In the diffusion-based novel view synthesis (NVS) setting, PR-IQA evaluates synthesized pseudo-views by exploiting real reference images from different poses, computes a geometrically consistent partial quality map over overlapping regions, and completes that map into a dense full-image quality estimate for downstream supervision. In the formulation introduced for diffusion-augmented sparse-view NVS, PR-IQA is designed to identify photometric and geometric inconsistencies in generated views and to restrict 3D Gaussian Splatting (3DGS) supervision to high-confidence regions, thereby improving both reconstruction fidelity and novel-view rendering without requiring ground-truth views at inference (Choi et al., 6 Apr 2026).

1. Conceptual scope and problem setting

PR-IQA arises from the mismatch between conventional IQA assumptions and the supervision regime of sparse-view NVS. Diffusion models can generate pseudo-ground-truth views that densify coverage and hallucinate plausible appearance for unseen surfaces, but these outputs often contain photometric inconsistencies such as color shifts, texture smearing, de-noising artifacts, and view-dependent effects, as well as geometric inconsistencies such as misaligned structures, local perspective errors, and missing occlusions. When such pseudo-views are used naively for supervision, they can bias 3DGS parameter optimization, including splat positions, scales, and opacities, toward incorrect geometry or blurred radiance (Choi et al., 6 Apr 2026).

Within the standard IQA taxonomy, full-reference (FR) IQA requires a pixel-aligned pristine or ground-truth image at the same pose; no-reference (NR) IQA predicts quality from the assessed image alone; and reduced-/partial-reference (RR/PR) IQA traditionally uses a compact set of features or statistics derived from the pristine reference at test time. Degraded-reference IQA (DR-IQA) defines another intermediate regime in which the degraded image itself is available as a surrogate reference at test time, while pristine images are used only during training to shape a learned reference space (Zheng et al., 2021). Against that backdrop, the NVS-oriented PR-IQA formulation is distinct: it does not assume a pristine reference, does not assume pose alignment, and does not operate as a conventional side-channel RR method. Instead, it uses unaligned reference views from the same scene wherever geometric overlap can be established, and extrapolates beyond those overlap regions.

A common misconception is to equate PR-IQA in this setting with classical RR-IQA. The two share the idea of incomplete reference availability, but the operational object differs. Classical RR/PR methods usually depend on handcrafted or learned features derived from the pristine image and transmitted as side information; the NVS-oriented PR-IQA framework instead uses real reference views from different poses and explicitly couples multi-view geometry with feature similarity. A second misconception is that “full-reference-level accuracy without ground-truth supervision” implies the complete absence of ground truth. In the reported formulation, ground truth is absent at inference, but FR-quality maps computed from GT at the same pose are used during training and validation of the completion network (Choi et al., 6 Apr 2026).

2. Geometry-aware partial quality estimation

The core of PR-IQA is the construction of a partial quality map on the subset of pixels for which the synthesized view can be geometrically related to one or more real reference views. Let a synthesized view $s$ and a set of real reference views $r \in R$ have intrinsics $K_s, K_r$ and extrinsics $(R_s, t_s), (R_r, t_r)$, with homogeneous pixel coordinates denoted by tildes. For a pixel $p_s$ in the synthesized view with depth $d_s$, the standard unprojection/reprojection chain is

$$X_s = d_s K_s^{-1} \tilde{p}_s,\quad X_w = R_s^\top (X_s - t_s),\quad X_r = R_r X_w + t_r,\quad \tilde{p}_r \sim K_r X_r.$$

Visibility and occlusion are handled with z-buffering or depth consistency checks. A reprojection is valid if $p_r$ falls inside the image bounds and if the synthesized depth $d_s$ and the reference depth $d_r$ agree within a tolerance. The overlap mask $r \in R$0 marks pixels with valid reprojections across one or more references, and $r \in R$1 otherwise (Choi et al., 6 Apr 2026).

Quality on overlapping pixels is computed from geometric reprojection and photometric or feature consistency. With synthesized-view features $r \in R$2 and reference features $r \in R$3, the principal feature-space measure is DINOv2 cosine similarity normalized to $r \in R$4:

$r \in R$5

The formulation also allows photometric errors such as $r \in R$6, SSIM maps, or gradient consistency to be converted to quality scores via monotonic mappings, while geometric reliability can be modulated by reprojection error and depth agreement. Visibility or confidence weights $r \in R$7 are derived from VGGT stereo confidence for the 3D point, reprojection validity from z-buffering, and per-view reliability such as pose uncertainty. Across multiple references, aggregation can be a weighted average, or robust operators such as min or median, but the reported implementation finds that a pixel-wise max operator performs best for multi-reference fusion by selecting the most reliable evidence per pixel. The final partial map is masked as $r \in R$8.

In the reported implementation, dense 3D correspondences are obtained from VGGT, and LoftUp-upsampled DINOv2 features are warped from reference to query to compute $r \in R$9. Overlap determination uses depth confidence filtering and z-buffering for occlusion, with 20% depth-confidence filtering identified as optimal in the ablation. This combination makes the partial map explicitly geometry-aware: quality is only measured where correspondence is supportable, rather than inferred from purely image-local statistics (Choi et al., 6 Apr 2026).

3. Completion from partial to dense quality maps

Because overlap masks leave blind spots, PR-IQA includes a completion module that predicts a dense quality map $K_s, K_r$0 over all pixels by inpainting the sparse or partial quality evidence. The reported network is a three-stream encoder-decoder with inputs consisting of the synthesized image $K_s, K_r$1, a reference image $K_s, K_r$2, the partial quality map $K_s, K_r$3, overlap masks $K_s, K_r$4, and positional encodings. Its encoders are $K_s, K_r$5 for the reference branch, $K_s, K_r$6 for the synthesized branch, and $K_s, K_r$7 for the partial-map branch; all are multi-scale pyramids with dual-gated attention blocks in which channel attention is followed by spatial attention (Choi et al., 6 Apr 2026).

The dual gating has a specific intended role. The channel gate selects relevant features, while the spatial gate propagates quality spatially. This design is described as decoupling “what” to propagate from “where,” so that reliable overlap-supported evidence can extend into non-overlapping regions without excessive ghosting or drift. Cross-attention injects reference-view context at every stage. Queries from synthesized or partial-map features attend to keys and values from reference features at the same scale, according to

$K_s, K_r$8

The implementation applies mask-aware attention by zeroing out $K_s, K_r$9 at invalid positions based on $(R_s, t_s), (R_r, t_r)$0 and occlusion, and adds 2D positional embeddings and camera embeddings to stabilize alignment. After each stage $(R_s, t_s), (R_r, t_r)$1, synthesized and partial streams are fused via

$(R_s, t_s), (R_r, t_r)$2

and the decoder upsamples fused features to full resolution to predict $(R_s, t_s), (R_r, t_r)$3 (Choi et al., 6 Apr 2026).

Architecturally, the completion module is the point at which PR-IQA differs most sharply from geometry-only overlap scoring. Geometry alone can establish reliable evidence in visible overlap, but cannot score unsupported regions. The completion network is therefore not an auxiliary post-processing step; it is the mechanism that turns overlap-conditioned evidence into a full-image quality assessment. The ablation results reinforce this interpretation: removing the partial-map branch or the reference branch degrades PLCC and SRCC, and the ordering of the dual-gated attention matters, with channel then spatial attention yielding the best ranking consistency (Choi et al., 6 Apr 2026).

4. Training objectives, targets, and reproducibility

The completion network is trained to approximate FR-quality maps $(R_s, t_s), (R_r, t_r)$4 computed with ground truth at the same pose, but ground truth is used only for training and validation and not for inference. Two target variants are instantiated: a DINOv2 feature similarity map defined by pixel-wise cosine similarity between the synthesized image $(R_s, t_s), (R_r, t_r)$5 and the GT image $(R_s, t_s), (R_r, t_r)$6, and an SSIM map between $(R_s, t_s), (R_r, t_r)$7 and $(R_s, t_s), (R_r, t_r)$8. The optimization objective blends local accuracy, distribution alignment, and ranking consistency:

$(R_s, t_s), (R_r, t_r)$9

$p_s$0

and the total loss is

$p_s$1

with $p_s$2 (Choi et al., 6 Apr 2026).

Although a masked consistency loss against $p_s$3 on $p_s$4 can be added, the reported ablation finds the architectural anchoring to $p_s$5 sufficient; removing the partial-map branch leads to significant drops. This is significant because it locates the essential supervisory signal not merely in the overlap mask, but in the explicit partial quality stream that remains coupled to the reference-conditioned completion network.

The reported reproducibility details are unusually explicit. Inputs are resized to $p_s$6. Optimization uses AdamW with $p_s$7, initial learning rate $p_s$8, and cosine annealing with warm restarts of period $p_s$9k down to $d_s$0, for a total of $d_s$1k steps on $d_s$2 RTX 3090 GPUs with total batch size 12. The encoders use stages with blocks $d_s$3, heads $d_s$4, and channels $d_s$5; decoder stages use blocks $d_s$6, heads $d_s$7, and channels $d_s$8. The model has approximately $d_s$9M parameters, with memory of approximately $$X_s = d_s K_s^{-1} \tilde{p}_s,\quad X_w = R_s^\top (X_s - t_s),\quad X_r = R_r X_w + t_r,\quad \tilde{p}_r \sim K_r X_r.$$0 GB for single-image inference and approximately $$X_s = d_s K_s^{-1} \tilde{p}_s,\quad X_w = R_s^\top (X_s - t_s),\quad X_r = R_r X_w + t_r,\quad \tilde{p}_r \sim K_r X_r.$$1 GB for training with batch size 1. Training uses 120k pairs synthesized from MFR using ViewCrafter’s VDM, and references are selected at $$X_s = d_s K_s^{-1} \tilde{p}_s,\quad X_w = R_s^\top (X_s - t_s),\quad X_r = R_r X_w + t_r,\quad \tilde{p}_r \sim K_r X_r.$$2 and $$X_s = d_s K_s^{-1} \tilde{p}_s,\quad X_w = R_s^\top (X_s - t_s),\quad X_r = R_r X_w + t_r,\quad \tilde{p}_r \sim K_r X_r.$$3 frames from queries (Choi et al., 6 Apr 2026).

5. Quality-aware integration into diffusion-augmented 3DGS

PR-IQA is not presented solely as an assessor; it is integrated into a quality-aware 3DGS training pipeline. For each sampled viewpoint $$X_s = d_s K_s^{-1} \tilde{p}_s,\quad X_w = R_s^\top (X_s - t_s),\quad X_r = R_r X_w + t_r,\quad \tilde{p}_r \sim K_r X_r.$$4 between two nearby inputs $$X_s = d_s K_s^{-1} \tilde{p}_s,\quad X_w = R_s^\top (X_s - t_s),\quad X_r = R_r X_w + t_r,\quad \tilde{p}_r \sim K_r X_r.$$5, multiple diffusion candidates $$X_s = d_s K_s^{-1} \tilde{p}_s,\quad X_w = R_s^\top (X_s - t_s),\quad X_r = R_r X_w + t_r,\quad \tilde{p}_r \sim K_r X_r.$$6 are generated. For each candidate, PR-IQA maps are computed with $$X_s = d_s K_s^{-1} \tilde{p}_s,\quad X_w = R_s^\top (X_s - t_s),\quad X_r = R_r X_w + t_r,\quad \tilde{p}_r \sim K_r X_r.$$7 and $$X_s = d_s K_s^{-1} \tilde{p}_s,\quad X_w = R_s^\top (X_s - t_s),\quad X_r = R_r X_w + t_r,\quad \tilde{p}_r \sim K_r X_r.$$8, then fused by pixel-wise max to form $$X_s = d_s K_s^{-1} \tilde{p}_s,\quad X_w = R_s^\top (X_s - t_s),\quad X_r = R_r X_w + t_r,\quad \tilde{p}_r \sim K_r X_r.$$9. An image-level score $p_r$0 is defined as the spatial mean of $p_r$1, and the top candidate is selected as

$p_r$2

This establishes a first filtering stage at the whole-image level, before pixel-level masking is applied (Choi et al., 6 Apr 2026).

The second filtering stage operates during 3DGS optimization. Pseudo-view supervision is converted into a binary confidence mask by thresholding the quality map at percentile $p_r$3:

$p_r$4

with $p_r$5 by default. The masked reconstruction loss is

$p_r$6

and the total training loss combines this term with D-SSIM weighted by $p_r$7:

$p_r$8

The framework also permits continuous weighting, for example with $p_r$9, and reports comparable performance, though binary masking gives slightly better perceptual metrics in some cases. Threshold selection is treated as a balance between filtering and data retention: $d_s$0 performs best across datasets, $d_s$1 is too aggressive, and $d_s$2 is slightly more lenient (Choi et al., 6 Apr 2026).

The overall pipeline comprises six stages: recovering camera parameters for sparse inputs; sampling candidate viewpoints and generating diffusion pseudo-views; computing partial maps via geometry-aware warping of DINOv2 features and completing them through the PR-IQA network; selecting the best candidate per viewpoint by image-level score; training 3DGS on real inputs and selected pseudo-views with masking from $d_s$3; and optionally iterating by retraining PR-IQA with improved reconstructions or adding more references. This suggests that PR-IQA functions as a supervisory governor for diffusion augmentation rather than as an isolated evaluator.

6. Empirical behavior, ablations, and operating characteristics

On IQA evaluation against FR GT maps on Tanks and Temples, Mip-NeRF 360, and RealEstate10K, the reported DINOv2-target variant achieves PLCC/SRCC of $d_s$4 on Mip-NeRF 360, $d_s$5 on Tanks and Temples, and $d_s$6 on RealEstate10K. Against SSIM targets, the SSIM-target variant achieves $d_s$7, $d_s$8, and $d_s$9 on the same datasets, respectively. These results are reported as outperforming Cross-Reference baselines such as PuzzleSim and CrossScore, and as approaching FR-level correlations without GT supervision at inference (Choi et al., 6 Apr 2026).

When embedded in 3DGS, PR-IQA improves NVS quality over both ViewCrafter and Vanilla 3DGS. The reported metrics are summarized below.

Dataset	Ours_DINOv2 PSNR/SSIM/LPIPS	Baselines
Mip-NeRF 360	16.76 / 0.493 / 0.414	ViewCrafter: 16.18 / 0.474 / 0.453; Vanilla: 16.08 / 0.461 / 0.415
Tanks and Temples	16.24 / 0.551 / 0.403	ViewCrafter: 15.77 / 0.523 / 0.455; Vanilla: 15.30 / 0.509 / 0.406
RealEstate10K	17.72 / 0.632 / 0.327	ViewCrafter: 16.94 / 0.620 / 0.327; Vanilla: 16.39 / 0.625 / 0.345

The ablations clarify which components are structurally important. Removing the partial-map branch or the reference branch degrades PLCC and SRCC; performance scales with the number of references; and max fusion is consistently best among max, min, median, and mean. Binary and soft masking both work, with similar PSNR and SSIM, but binary masking slightly improves LPIPS in some cases. Geometric robustness is reported under 5–10% camera noise, where performance degrades gracefully but still beats CR baselines. PR-IQA also maintains higher correlations than baselines for overlap ratios down to 5–10%, yields lower FPR@Top-X% in non-overlapping regions, and generalizes across unseen generators such as GEN3C and SEVA. Runtime overhead is approximately 1.02 s per image for feature extraction, VGGT, and inference, compared with approximately 25 s per scene for 3DGS optimization, which is presented as practical in the overall pipeline (Choi et al., 6 Apr 2026).

7. Relations to adjacent paradigms, limitations, and prospective directions

PR-IQA in the NVS formulation occupies a specific position relative to other IQA families. Relative to FR-IQA, it aims to reach FR-level accuracy without requiring pose-aligned GT at inference, while still being trained with proxy GT maps. Relative to NR-IQA, it leverages multi-view geometry and semantics, whereas NR metrics are described as weak for subtle multi-view inconsistencies and semantics in generated views. Relative to patch-based CR-IQA methods such as CrossScore and PuzzleSim, it adds explicit geometry and dense completion to unseen regions. Relative to geometry-only CR-IQA such as MEt3R, it retains reliability in overlap while avoiding blind spots through dense completion (Choi et al., 6 Apr 2026).

Broader adjacent work shows that “partial reference” can be instantiated in other ways. DR-IQA uses the degraded image as the test-time reference and learns a deep reference space by conditional knowledge distillation from pristine images during training; its distinction from RR/PR-IQA is that no pristine-derived side information is available at test time, and the reference representation is learned end-to-end rather than handcrafted (Zheng et al., 2021). A different line, MQAF, combines a reference-guided score with a memory-driven score and an adaptive weighting parameter $d_r$0, and an overview of that framework describes masked or patch-level computation of the reference score as a practical adaptation to partial-reference operation. This suggests an alternative design philosophy for PR-IQA when reference information is incomplete or unreliable: rather than completing geometry-supported evidence densely, the system can reduce reliance on the reference and back off toward a learned memory of distortion patterns (Lan et al., 20 Feb 2026).

The limitations reported for the NVS-oriented PR-IQA are specific and consequential. Its perceptual ceiling reflects the FR targets used during training, namely DINOv2-SIM or SSIM. Dependence on geometric cues such as VGGT stereo and pose estimation can degrade behavior in textureless regions or under extreme lighting changes. The method has been validated on 3DGS with ViewCrafter-like diffusion, and broader integration across emerging diffusion and reconstruction systems remains a promising direction. The paper also identifies end-to-end joint optimization of geometry and quality, and the incorporation of human preference data, as plausible routes to improve perceptual alignment and robustness (Choi et al., 6 Apr 2026).

In that sense, PR-IQA is best understood not as a single metric family but as a broader operational principle: use whatever reference support is geometrically or semantically defensible, infer quality where correspondence is reliable, and propagate that evidence in a way that remains aware of missing support. In sparse-view NVS, the reported formulation turns that principle into a concrete supervision mechanism for diffusion-augmented 3D reconstruction (Choi et al., 6 Apr 2026).