DualRefIQA
ACM Multimedia 2026Image Quality Assessment

Bridging the
Reference Gap.

Dual-Reference SR-IQA with Generated Image Guidance as Perceptual Anchors

A new quality metric that asks two different questions: does the super-resolved image preserve the original content—and does it look perceptually natural?

Shenghan Xu et al.
Read paper GitHub
DualRefIQA resolves the conflict between fidelity and perceptual quality using two complementary references.
One image. Two anchors. Content fidelity and perceptual naturalness are evaluated separately, then fused.

The evaluation dilemma

Sharp can be wrong.
Faithful can look blurry.

Full-reference metrics require unavailable ground-truth HR images and over-penalize plausible, misaligned details.

No-reference metrics can reward sharp-looking hallucinations because they lack a content anchor.

The idea

Two references,
two responsibilities.

DualRefIQA explicitly separates content correctness from visual quality, using the right reference—and the right matching strategy—for each.

01

Content anchor

Low-resolution input

Preserves scene identity and low-frequency structure. A local CNN path catches semantic drift and structural errors.

02

Perceptual anchor

Generated HR reference

Provides rich texture statistics. Cross-attention evaluates naturalness without requiring pixel-level alignment.

DualRefIQA architecture with Spatial De-Align module, Perceptual Path, Fidelity Path, and score regression.
Figure 2 The asymmetric dual-branch architecture. The Spatial De-Align module preserves high-frequency texture cues before shared feature extraction.

Experimental results

Closer to human judgment.

DualRefIQA delivers leading correlation among methods that do not use inaccessible ground-truth HR references.

DSRQ · SRCC

0.6913best non-GT result

DSRQ · PLCC

0.6922human score alignment

ISRGen-QA · SRCC

0.7773strong cross-dataset fit

ISRGen-QA · PLCC

0.8324without GT HR
Visual comparison of image quality scores across methods, with DualRefIQA aligning most closely with ground-truth MOS.

Qualitative comparison

Hallucinations exposed.

Conventional metrics often give high scores to generated detail even when it changes scene structure. By consulting both anchors, DualRefIQA better reproduces the ordering chosen by human raters.

170K+
LR / SR / generated-reference triplets in DSRQ
−15.4%
SRCC when the SDA module is removed

Beyond evaluation

A metric that can teach.

Used as a training objective, DualRefIQA guides SR models toward sharper, more natural textures while retaining structural fidelity. On RCAN, adding the metric improves PSNR, SSIM, LPIPS, and the predicted perceptual quality score over the corresponding baseline.

Citation

Cite DualRefIQA.

The submission PDF is anonymized; this provisional entry can be updated with the final author list and publication identifiers.

@inproceedings{xu2026dualrefiqa,
  title     = {Bridging the Reference Gap: Dual-Reference SR-IQA
               with Generated Image Guidance as Perceptual Anchors},
  author    = {Xu, Shenghan and others},
  booktitle = {Proceedings of the 34th ACM International Conference
               on Multimedia},
  year      = {2026}
}
BibTeX copied