Can We Use Saliency Maps for the Evaluation of Generative Image Models?

Eleonora Ristori; Luca Bindini; Marco Lippi; Paolo Frasconi

cs.CVOct 16, 2026

Can We Use Saliency Maps for the Evaluation of Generative Image Models?

Authors: Eleonora Ristori, Luca Bindini, Marco Lippi, Paolo Frasconi

Organizations: AI Lab, DINFO, University of Florence, Florence, Italy

Abstract

Evaluating generative image models remains challenging, as commonly used metrics depend on fixed pretrained feature extractors optimized for different tasks that do not always align with human realism perception across datasets and model families. To address this limitation, we investigate a human-inspired evaluation framework that avoids representations learned for unrelated tasks. Specifically, we train a generator-specific classifier to distinguish real images from samples produced by a given model and analyze its decisions using saliency maps. Across diverse benchmarks, we observe that these classifiers learn generator-specific artifacts rather than a universal notion of "fakeness", exhibiting strong accuracy and limited cross-generator transfer. This makes saliency analysis particularly informative: the resulting attribution maps highlight visual cues and structural irregularities characteristic of each model's outputs. By combining classification accuracy and saliency information, we obtain a composite score aligned with human judgments, without relying on large pretrained backbones and using only real and generated samples from the evaluated model.

Figures & tables

Figure 1: Framework overview. We train a generator-specific real-vs-fake classifier (GSC) to distinguish real images from those generated by a specific model; then, we combine performance on such task with XAI saliency maps to evaluate generator quality.

Figure 2: Metrics vs HER. Both FID and SAS are negatively correlated with HER.

r

denotes the Pearson correlation, shaded background indicates correlations that are not statistically significant (

p\leq 0.05

) with respect to HER. The standard deviation is calculated across different sets of images.

Figure 3: Cross/self generator recalls (CSGR). Entry $(i,j)$ shows the recall of the GSC trained on generator $i$ and tested on images from generator $j$ (fake being the positive class). The last column reports the recall of each GSC when tested on real images.

Training	Recall	Test dataset
Training	Recall	LSUN	ImageNet	FFHQ	CIFAR10
LSUN	Real	98.36	92.37	87.81	95.63
	Fake out	22.85	13.05	16.82	5.42
	Fake in	92.96	-	-	-
ImageNet	Real	94.60	95.33	93.75	91.41
	Fake out	16.90	14.42	12.29	8.88

Table 3: Cross-dataset recall. GSCs are evaluated on generated images from other datasets. Fake in denotes the recall on fake samples generated by the same model used during training, whereas Fake out refers to the recall on fake samples produced by generative models not seen during training.

Figure 4: $s(g)$ vs HER. Red points denote models for which the GSC achieves limited cross-generator generalization, indicating strong learning of model-specific artifacts. Blue points correspond to GSCs with greater generalization across generators, suggesting that they learn a more generator-agnostic notion of “fakeness”.

Aggregation Method	CIFAR10	ImageNet	FFHQ	LSUN
Median	-0.77*	-0.69*	-0.90*	-0.96*
Mean	-0.77*	-0.58	-0.86*	-0.89*
L1 norm	-0.73*	-0.65*	-0.87*	-0.94*
L2 norm	-0.57*	-0.60*	-0.81*	-0.89*
Max	-0.65*	-0.56	-0.78*	-0.87*
Min	-0.78*	-0.63*	-0.86*	-0.94*

Table 4: Pearson correlation

r

between SAS and HER by varying the aggregation method for the saliency maps in the computation of

s(g)

presented in Equation 2 . Values marked with * are statistically significant.

Figure 5: SAS vs HER, where

s(g)

is calculated using different techniques. The coefficient

r

denotes the Pearson correlation, and panels with a shaded background indicate correlations that are not statistically significant (

p\leq 0.05

) with respect to HER. GradientSHAP and SmoothGrad maintain strong correlations with HER, while CAM-based methods show weaker results, likely due to their coarse spatial resolution.

Figure 6: Qualitative comparison of saliency maps produced by different attribution methods. CAM-based approaches generate coarse maps often focusing on similar central regions across generators. Integrated Gradients (IG), SmoothGrad and GradientSHAP instead produce finer and more diverse attribution patterns, which better capture generator-specific artifacts.

Figure 7: CrossScore vs HER. Pearson

r

is always higher than

0.75

and significant (

p\leq 0.05

Figure 8: Cross-generator recall vs HER. GSCs separate models into two groups: high-quality generators, where human error exceeds

25\%

and GSCs show stronger generalization; and low-quality generators, where human accuracy is nearly perfect, and GSC generalization is minimal.

	LSUN Bedroom	FFHQ	ImageNet	CIFAR10
FID	-0.11	-0.59	0.37	-0.96*
$\text{FD}_{\text{DINOv2}}$	-0.94*	-0.50	-0.76*	-0.98*
$\text{FD}_{\text{GSC}}$	-0.77*	-0.49	-0.32	-0.83*
CMMD	-0.98*	-0.20	-0.67*	-0.96*
$\text{MMD}_{\text{DINOv2}}$	-0.95*	-0.79*	-0.46	-0.93*
$\text{MMD}_{\text{GSC}}$	-0.97*	-0.58	-0.54	-0.71*

Table 5: Pearson correlation coefficient between HER and different evaluation metrics: the traditional FID [ 17 ] , Fréchet Distance and MMD computed with DINOv2 features, the CMMD score computed with CLIP features [ 19 ] , and the corresponding Fréchet Distance and MMD scores obtained using GSCs as feature extractors. * indicates significant correlations (

p\leq 0.05

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9: Pearson correlation

r

between SAS and HER as the scaling parameter

\alpha

varies. Across all datasets, the correlation remains statistically significant (

p\leq 0.05

), showing that SAS is robust to the choice of

\alpha

Training	Recall	Test dataset
Training	Recall	LSUN	ImageNet	FFHQ	CIFAR10
LSUN	Real	99.74	84.21	74.01	66.19
	Fake out	51.68	38.71	38.35	68.25
	Fake in	99.84	-	-	-
ImageNet	Real	96.12	99.45	98.47	44.25
	Fake out	30.96	24.26	17.96	80.94

Appendix

Table 6: Cross-dataset recall for ViT-based GSCs. Classifiers trained on one dataset are evaluated on real and generated images from other datasets. Fake in denotes the recall on fake samples generated by the same model used during training, whereas Fake out refers to the recall on fake samples produced by generative models not seen during training.

Figure 10: Cross/self generator recalls (CSGR) for ViT backbones, using a pool of $M$ generators. Specifically, the value in the $i$ -th row and $j$ -th column represents the cross-generator (self-generator when $i=j$ ) recall of the GSC trained on images from generator $i$ and tested on images from generator $j$ (with $i,j\in{1,\dots,M}$ ), considering fake images as the positive class. The $(M+1)$ -th column shows the recall of each GSC when evaluated on real images.

Figure 11: SAS vs HER, using a ViT as our GSC. Panels with a shaded background do not have significant (

p\leq 0.05

) correlations with HER. Compared with CNN-based classifiers (see Figure 2 ), ViTs show weaker correlations with HER, likely because their stronger generalization and global feature representations reduce sensitivity to generator-specific artifacts.

Dataset	Recall
Dataset	Real	Fake-in	Fake-out
LSUN Bedroom	98.44	98.14	84.87
ImageNet	95.41	95.39	72.91
FFHQ	98.24	98.51	77.72
CIFAR10	94.33	95.36	70.41

Appendix

Table 7: Recall for GSCs trained on DINOv2 features rather than on raw images. Fake-in denotes the recall on fake samples generated by the same model used during training, whereas Fake-out refers to the recall on fake samples produced by generative models not seen during training.

Figure 12: On the left: UMAP of the feature space of our GSC; the samples from the generator seen during training are well separated, while the other generated samples and the real ones are perfectly overlapped. On the right: UMAP projection of GSC representations learned from DINOv2 features; real samples occupy a largely distinct region, indicating effective real-versus-fake discrimination, while samples from different generators substantially overlap, suggesting limited sensitivity to generator-specific fingerprints.

Image Generation Evaluation

Image generation evaluation measures the quality, fidelity, diversity, and instruction-following ability of AI-generated images.

See all 112 papers

RA-ClipScore: Making Generative Model Evaluation More Interpretable
Aug 12, 2026Yifan Lu, Taras Kucherenko, Hedvig Kjellström +1Image Generation Image Generation Evaluation
KTH Royal Institute of Technology · National Library of Sweden · Most of the work was performed while Taras was at Electronic Arts (EA). +3
TRIG: Toward Measuring Texture Realism in Image Generation
Oct 10, 2026Zhiwei Xu, Jiali Cui, Zhiqiang Lao +2Image Generation Evaluation Image Quality Assessment
Northwestern University · Futurewei Technologies
Scalable Black-Box Model Attribution for Images
Aug 16, 2026Asaf Livne, Amir Jevnisek, Shai AvidanDigital Image Forensics
Tel Aviv University

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

RA-ClipScore: Making Generative Model Evaluation More Interpretable

Yifan Lu, Taras Kucherenko, Hedvig Kjellström, Judith Bütepage

KTH Royal Institute of Technology · National Library of Sweden · Most of the work was performed while Taras was at Electronic Arts (EA).+3

Image Generation Image Generation Evaluation

Open paper AskPDF arXiv

Generative models can produce images nearly indistinguishable from real data, yet rigorous and interpretable evaluation remains challenging. Conventional metrics such as FID provide only scalar scores with limited diagnostic insight. Widely adopted CLIP-based metrics enable semantic evaluation beyond simple training class labels, but inherit limitations from CLIP's training paradigm that restrict attribute-wise analysis. We propose RA-CLIPScore, a novel metric that mitigates these issues and extends CLIP-based evaluation to spatial distribution alignment, measuring whether generated objects adhere to the positional priors found in the training data. RA-CLIPScore introduces dual prompts to decouple competing attributes and leverages local patch tokens to capture fine-grained regional semantics. We evaluate image generative models on their ability to match both attribute and spatial distributions of the training data. Extensive experiments show that RA-CLIPScore provides more robust and interpretable evaluations than prior methods, particularly under distribution misalignment or partially irrelevant textual attributes. We further demonstrate how it reveals spatial biases in generative models. User evaluations confirm that Regional Single Attribute Divergence based on our RA-CLIPScore aligns more closely with human perception of visual diversity than existing semantic metrics.

Abstract

Image Generation Evaluation

Explore similar work

RA-ClipScore: Making Generative Model Evaluation More Interpretable

TRIG: Toward Measuring Texture Realism in Image Generation

Scalable Black-Box Model Attribution for Images