acceptodds
Under review as a conference paper at ICLR 2027

RealSynz: Naturalizing Synthetic Images with Envelope Variation Statistics

Abstract

Generated images still show consistent shortcomings and artifacts compared with natural images. While they represent scenes well, textures such as unnaturally waxy skin and flat fabric place them in an uncanny valley. We find that the difference lies in how fine detail is distributed. Three of the five generators we study carry less fine detail than natural images, and the two that carry as much or more distribute it differently. In textured regions, every generator packs fine detail into small areas and leaves the space between them nearly empty. In flat regions, three of the five generators spread fine detail too evenly. We measure this distribution with two statistics of natural images. The moment ratio is the mean absolute value of fine detail in a patch divided by its root mean square, and envelope variation is the coefficient of variation of the local magnitude of fine detail across a patch. Edits that change only the amount of detail leave both statistics far from natural images. Based on this finding, we present RealSynz, an independent module that corrects this distribution in the output of any generator without access to the model. RealSynz corrects the moment ratio in closed form and trains a small network that decides only where to redistribute fine detail and where to add texture. On 1,000 anachronistic FLUX.2 Klein scenes, RealSynz closes 65% of the envelope variation gap, outperforms fifteen baseline methods, and transfers to four other generators without retraining.

Primary area:
generative models
Keywords:
Image Generation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_423XOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper first analyzes the differences between generated images and natural images in a statistical sense, in terms of the moment ratio and envelope variation. Then, it proposes RealSynz, a post-processing module for correcting the distribution of the generated image to be closer to that of a natural image. Specifically, the moment ratio is corrected via a closed form equation and envelope variation is corrected via redistribution of the band and adding texture. A map is estimated with a small convolutional network to act as a control field to decide where and how strongly each operation should be applied.

Critical strengths
  1. The paper is very clearly written. The motivation is convincing with supporting analyses. The method is developed with a coherent storyline. The experiments make many useful observations. I enjoyed reading the paper.

  2. The method is developed based on thorough analyses and observations. Closed-form solutions are used when possible, while deferring variable aspects to a small neural network. The paper overall provides a fresh perspective on how unnaturalness of generated images can be handled.

Critical weaknesses

My biggest concern is whether the proposed method indeed makes generated images look natural. I believe the most important experiment should be a perceptual study showing whether the generated images indeed look more closer to natural images (or become hard to distinguish) to the human eye. The paper defers this to future work, but I think it's a critical component to validate that the proposed ideas are indeed working. Frankly, the texture added by the proposed method in Fig 1, Fig 3, Fig 12-16 all look very synthetic to my eyes. Another good way to validate would be to run off-the-shelf synthetic image detectors (some are reviewed in l472-477) to check if they are more likely to classify the generated images as natural images after the proposed post-processing has been applied.

The paper mainly provides quantitative results showing envelope variation and moment ratio becoming more similar to those of natural images (Table 1 and Table 2), however, I believe this is an obvious result since the proposed method directly optimizes for lowering those two values.

Questions for authors

N/A

3 comments
AuthorsOpenReview ↗

Response to Reviewer 423X (1/2)

We thank the reviewer for the feedback. We conducted a human perceptual study and ran three AI-generated image detectors on generated images before and after RealSynz, as suggested.


W1

the most important experiment should be a perceptual study showing whether the generated images indeed look more closer to natural images (or become hard to distinguish) to the human eye.

We conducted a human perceptual study with 20 participants, each of whom evaluated the same 25 image pairs, each showing an unedited image and the same image after RealSynz, and selected the one they perceived as more natural and realistic.

ComparisonTrials choosing RealSynzParticipants favoring RealSynzImages favoring RealSynz
RealSynz vs. Unedited73.4% (367/500)90% (18/20)100% (25/25)

A participant or image favors RealSynz if RealSynz was chosen in more than half of its trials. The preference for RealSynz over the unedited images is significant and consistent. We have added the study to Section 3.2 and Table 2 of our revised version.


W2

the texture added by the proposed method in Fig 1, Fig 3, Fig 12-16 all look very synthetic to my eyes.

We agree that RealSynz is not perfect, and some of the added texture can still look synthetic under strong enlargement. Nevertheless, both our human perceptual study (W1) and the AI-generated image detectors (W3) show that RealSynz consistently makes generated images more realistic than the unedited ones: participants chose the RealSynz image in 73.4% of trials, and all three detectors classify more RealSynz images as real for all five generators.

AuthorsOpenReview ↗

Response to Reviewer 423X (2/2)

W3

Another good way to validate would be to run off-the-shelf synthetic image detectors (some are reviewed in l472-477) to check if they are more likely to classify the generated images as natural images after the proposed post-processing has been applied.

We selected three AI-generated image detectors:

  1. AIDE [1] combines the highest- and lowest-frequency patches of an image, selected by DCT, with a CLIP embedding of the whole image.
  2. SSP [2] classifies the noise residual of the flattest 32-pixel patch of an image.
  3. Community Forensics [3] is a whole-image detector trained on images from thousands of generators to generalize to unseen ones.

We used the authors' released official weights and evaluation code. For each generator, we evaluate its 1,000 generated images before and after RealSynz. The table below reports the percentage of images classified as real, where higher values indicate that fewer images are detected as synthetic.

GeneratorAIDESSPCommunity Forensics
FLUX.2 Klein22.6 →\rightarrow 87.521.8 →\rightarrow 62.957.4 →\rightarrow 80.1
Qwen-Image3.40 →\rightarrow 26.839.2 →\rightarrow 56.334.6 →\rightarrow 67.7
Z-Image3.50 →\rightarrow 41.63.80 →\rightarrow 40.051.4 →\rightarrow 70.8
FLUX.1-dev10.0 →\rightarrow 18.14.70 →\rightarrow 33.48.50 →\rightarrow 33.6
SANA16.8 →\rightarrow 38.326.0 →\rightarrow 41.00.80 →\rightarrow 2.40

After applying RealSynz, all three detectors classify more images as real for all five generators. We have added these results to Section 3.2 and Table 1 of our revised version.


W4

The paper mainly provides quantitative results showing envelope variation and moment ratio becoming more similar to those of natural images (Table 1 and Table 2), however, I believe this is an obvious result since the proposed method directly optimizes for lowering those two values.

The effectiveness of RealSynz has been validated by the human perceptual study (W1) and the AI-generated image detectors (W3). In addition, we evaluated two additional texture statistics that RealSynz does not use for training: Portilla–Simoncelli (PS) statistics [4] are a few hundred wavelet statistics from which visually similar textures can be synthesized, and GLCM correlation [5] measures how strongly the gray levels of neighboring pixels co-vary. We compute both on 4,000 surface crops of size 128×128128\times128 from the 1,000 FLUX.2 Klein generated images, before and after RealSynz, and on crops from real images. We then measure the distance between the statistics of generated and real crops, normalized so that two random halves of the real images are 1.0 apart. Lower values are closer to real images.

Metric (↓\downarrow)UneditedRealSynz
PS distance1.701.42
GLCM correlation distance3.401.30

The results show that RealSynz reduces both distances, even though neither statistic is used in its training. We have added these results to Appendix D.3 and Table 17 of our revised version.


[1] Yan et al. A Sanity Check for AI-generated Image Detection. ICLR 2025.

[2] Chen et al. A Single Simple Patch is All You Need for AI-generated Image Detection. arXiv 2024.

[3] Park and Owens. Community Forensics: Using Thousands of Generators to Train Fake Image Detectors. CVPR 2025.

[4] Portilla and Simoncelli. A Parametric Texture Model Based on Joint Statistics of Complex Wavelet Coefficients. IJCV 2000.

[5] Haralick et al. Textural Features for Image Classification. IEEE Transactions on Systems, Man, and Cybernetics 1973.

Appreciate the authors providing the additional results. The user study, evaluation using AI image detectors, and the additional quantitative metrics that are independent of training are all great evidence supporting the proposed method. I believe they make the submission much stronger. The new qualitative figures (Fig 3, Fig 17) look much more convincing as well. My concerns have been addressed and I am confident in raising my score to 3 (weak acceptance).

Reviewer_z4FhOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a way to alter AI generated images to have more natural textures. They identify some reasons that AI image generators fail to produce realistic textures in images. They quantify this using a few statistical methods. Then they are able to alter the image to have the matching statistics in a real and generated image. They show that their method RealSynz can get very good numbers for these statistics, closing the real and generated image gap significantly compared to baseline methods and their ablations.

Critical strengths

The major strength of this paper is identifying a mathematically quantifiable measure of texture naturalness. This is an interesting and novel way to quantify why AI images look different from real ones. With this quantifiable measure, the paper presents a way to minimize the difference between the real and AI generated measures to pull AI images to have more natural textures.

The paper is pretty well written. I am not an expert, but I was able to understand the paper with not a lot of effort.

Critical weaknesses

The only critical weakness of the paper is the circular evaluation metrics. The method is designed to optimize C_β envelope variation and C_r moment ratio, but these are also the metrics used to compare the proposed approach to baseline approaches. It should be noted that naively optimizing these is not a good approach as the paper shows that this can be faulty. The proposed optimization with the neural network works very smoothly. I think that having some more qualitative data in the paper could really help readers assess the quality of results. A human evaluation study would be really helpful, but I understand that this is challenging to conduct.

Questions for authors

Is it possible to validate with some additional metrics? Does the method hurt the image quality in any other way? I am currently borderline leaning towards accept for this paper, but I need some more information about the quality of the results with metrics that are not explicitly used for optimization in the methods.

2 comments
AuthorsOpenReview ↗

Response to Reviewer z4Fh (1/2)

We thank the reviewer for the comments. To address the concerns, we provide two independent texture statistics that RealSynz does not optimize, along with classification results from off-the-shelf AI-generated image detectors. We also include additional qualitative results in the revised version and report the human perceptual study as suggested.


W1

The only critical weakness of the paper is the circular evaluation metrics. The method is designed to optimize CβC_\beta envelope variation and CrC_r moment ratio, but these are also the metrics used to compare the proposed approach to baseline approaches. It should be noted that naively optimizing these is not a good approach as the paper shows that this can be faulty. The proposed optimization with the neural network works very smoothly.

To address this, we evaluate RealSynz with metrics it does not optimize: two independent texture statistics and three off-the-shelf AI-generated image detectors.

Independent texture statistics. For a more comprehensive evaluation, we introduced two additional independent metrics: Portilla–Simoncelli (PS) statistics [1] and gray-level co-occurrence matrix (GLCM) correlation [2], neither of which is used during training or inference.

PS statistics are a few hundred wavelet statistics from which visually similar textures can be synthesized, and GLCM correlation measures how strongly the gray levels of neighboring pixels co-vary. We compute both on 4,000 surface crops of size 128×128128\times128 from the 1,000 FLUX.2 Klein generated images, before and after RealSynz, and on crops from real images. We then measure the distance between the statistics of generated and real crops, normalized so that two random halves of the real images are 1.0 apart. Lower values are closer to real images.

Metric (↓\downarrow)UneditedRealSynz
PS distance1.701.42
GLCM correlation distance3.401.30

RealSynz reduces both distances. Note that neither statistic is used in the training of RealSynz. We have added these results to Appendix D.3 and Table 17 of our revised version.

Off-the-shelf AI-generated image detectors. We also ran modern learning-based AI-generated image detectors on generated images before and after RealSynz. Specifically, we used three pre-trained detectors: AIDE [3], SSP [4], and Community Forensics [5]. The table reports the percentage of the same 1,000 generated images per generator classified as real, before →\rightarrow after RealSynz.

GeneratorAIDESSPCommunity Forensics
FLUX.2 Klein22.6 →\rightarrow 87.521.8 →\rightarrow 62.957.4 →\rightarrow 80.1
Qwen-Image3.40 →\rightarrow 26.839.2 →\rightarrow 56.334.6 →\rightarrow 67.7
Z-Image3.50 →\rightarrow 41.63.80 →\rightarrow 40.051.4 →\rightarrow 70.8
FLUX.1-dev10.0 →\rightarrow 18.14.70 →\rightarrow 33.48.50 →\rightarrow 33.6
SANA16.8 →\rightarrow 38.326.0 →\rightarrow 41.00.80 →\rightarrow 2.40

All three detectors classify more images as real for all five generators, so the improvement is not limited to the statistics we optimize. We have added these results to Section 3.2 and Table 1 of our revised version.


W2

I think that having some more qualitative data in the paper could really help readers assess the quality of results.

The original submission includes a gallery in Appendix F showing qualitative results. In our revised version, we have added qualitative results to the main paper in Section 3.2 and Figure 3, and in Appendix F and Figure 17.


[1] Portilla and Simoncelli. A Parametric Texture Model Based on Joint Statistics of Complex Wavelet Coefficients. IJCV 2000.

[2] Haralick et al. Textural Features for Image Classification. IEEE Transactions on Systems, Man, and Cybernetics 1973.

[3] Yan et al. A Sanity Check for AI-generated Image Detection. ICLR 2025.

[4] Chen et al. A Single Simple Patch is All You Need for AI-generated Image Detection. arXiv 2024.

[5] Park and Owens. Community Forensics: Using Thousands of Generators to Train Fake Image Detectors. CVPR 2025.

AuthorsOpenReview ↗

Response to Reviewer z4Fh (2/2)

W3

A human evaluation study would be really helpful, but I understand that this is challenging to conduct.

We conducted a human perceptual study with 20 participants, each of whom evaluated the same 25 image pairs, each showing an unedited image and the same image after RealSynz, and selected the one they perceived as more natural and realistic.

ComparisonTrials choosing RealSynzParticipants favoring RealSynzImages favoring RealSynz
RealSynz vs. Unedited73.4% (367/500)90% (18/20)100% (25/25)

A participant or image favors RealSynz if RealSynz was chosen in more than half of its trials. The preference for RealSynz over the unedited images is significant and consistent. We have added the study to Section 3.2 and Table 2 of our revised version.


Q1

Is it possible to validate with some additional metrics?

Please refer to W1 and W3 for new metrics and human perceptual study results.


Q2

Does the method hurt the image quality in any other way?

By design, RealSynz edits only the finest texture band and leaves everything else in the image untouched. Any effect on image quality is therefore confined to the fine texture itself. To verify that the rest of the image is preserved, we compare each of the 1,000 FLUX.2 Klein generated images before and after editing, both on the original images and after blurring them with 8 blur passes, which removes the texture band and keeps the coarser structure. The table reports the mean absolute pixel difference in gray levels (0–255) and the share of it that remains after blurring. We compare with SDEdit [6], which re-runs the last few denoising steps of FLUX.2 Klein with each image's original generation prompt, so it aims to keep the same scene; we set its strength so that its total pixel change matches that of RealSynz.

MethodPixel changePixel change after blurKept after blur
SDEdit, strength 0.203.452.5975%
SDEdit, strength 0.353.812.7472%
RealSynz3.920.9023%

With a similar total change, SDEdit keeps over 70% of it after blurring, while RealSynz keeps only 23%. This indicates that RealSynz confines its edit to the fine texture and leaves the rest of the image largely unchanged. We have added this analysis to Appendix D.1 and Table 13 of our revised version.


[6] Meng et al. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. ICLR 2022.

Reviewer_4ivZOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper analyzes the fine detail differences between generated images and natural images, and observes that it's not the difference of amount of detail, but distribution that cause the perceptual gap. To fill in the gap, the paper first proposes two metrics, moment ratio and envelope variation to quantify distribution of details within a patch. Then, the paper proposes a convlutional neural network to learn the amount of detail changes to apply for a generated image, in order to make it align with natural images. The experimental results show that interms of the defined metrics, and some qualitative examples, the proposed method indeed improved fine detail perceptual quality comparing to baseline methods. However, there lacks a holistic metric like FID to give an overall measure of the realness of the prost processed results.

Critical strengths
  1. The analysis conducted by the paper is sound and convincing, revealing that it is the distribution, not amount of details, that causes the perceptual difference between generated image and natural image
  2. The efficacy of the proposed learning method to improve the distribution gap also supported by experimental results
  3. The paper further proposes a new benchmark for measuring texture naturalism, which is helpful for future research in this direction
Critical weaknesses
  1. Although the reported quantitative and qualitative results support the claims and proposed method for the paper, there isn't a holistic metric like FID to measure texture naturalness. The current results are given by moment ratio and envelope variation, which is not very straight forward to directly map them to perceptual qualities.
Questions for authors

please see weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer 4ivZ (1/2)

We thank the reviewer for the suggestions. Below we provide FID, a human perceptual study, and AI-generated image detector results, each comparing generated images before and after RealSynz.


W1

Although the reported quantitative and qualitative results support the claims and proposed method for the paper, there isn't a holistic metric like FID to measure texture naturalness.

We computed FID [1] as suggested on all 1,000 FLUX.2 Klein images generated from held-out test prompts against 1,000 real images. We apply RealSynz both directly to the generated images and on top of two diffusion restorers (HYPIR [2] and PiSA-SR [3]).

InputFID, inputFID, + RealSynzFID change
Unedited271.6260.3-11.3
HYPIR286.7281.2-5.5
PiSA-SR281.9275.4-6.5

RealSynz lowers FID in every setting. On its own, it lowers FID by 11.3 points. On top of each restorer, it lowers FID by 5.5 and 6.5 points. We have added the FID results to Appendix D.3 and Table 18 of our revised version.


[1] Heusel et al. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. NeurIPS 2017.

[2] Lin et al. Harnessing Diffusion-Yielded Score Priors for Image Restoration. ACM Transactions on Graphics 2025.

[3] Sun et al. Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach. CVPR 2025.

AuthorsOpenReview ↗

Response to Reviewer 4ivZ (2/2)

W2

The current results are given by moment ratio and envelope variation, which is not very straight forward to directly map them to perceptual qualities.

To map our results more directly to perceptual qualities, we conducted two additional evaluations. The results show:

1. Human perceptual study: participants prefer RealSynz. We conducted a human perceptual study with 20 participants, each of whom evaluated the same 25 image pairs, each showing an unedited image and the same image after RealSynz, and selected the one they perceived as more natural and realistic.

ComparisonTrials choosing RealSynzParticipants favoring RealSynzImages favoring RealSynz
RealSynz vs. Unedited73.4% (367/500)90% (18/20)100% (25/25)

A participant or image favors RealSynz if RealSynz was chosen in more than half of its trials. The preference for RealSynz over the unedited images is significant and consistent. We have added the study to Section 3.2 and Table 2 of our revised version.

2. AI-generated image detectors classify more RealSynz images as real. We ran modern learning-based AI-generated image detectors on generated images before and after RealSynz. Specifically, we used three pre-trained detectors: AIDE [4], SSP [5], and Community Forensics [6]. The table reports the percentage of the same 1,000 generated images per generator classified as real, before →\rightarrow after RealSynz.

GeneratorAIDESSPCommunity Forensics
FLUX.2 Klein22.6 →\rightarrow 87.521.8 →\rightarrow 62.957.4 →\rightarrow 80.1
Qwen-Image3.40 →\rightarrow 26.839.2 →\rightarrow 56.334.6 →\rightarrow 67.7
Z-Image3.50 →\rightarrow 41.63.80 →\rightarrow 40.051.4 →\rightarrow 70.8
FLUX.1-dev10.0 →\rightarrow 18.14.70 →\rightarrow 33.48.50 →\rightarrow 33.6
SANA16.8 →\rightarrow 38.326.0 →\rightarrow 41.00.80 →\rightarrow 2.40

After applying RealSynz, all three detectors classify more images as real for all five generators. We have added these results to Section 3.2 and Table 1 of our revised version.


[4] Yan et al. A Sanity Check for AI-generated Image Detection. ICLR 2025.

[5] Chen et al. A Single Simple Patch is All You Need for AI-generated Image Detection. arXiv 2024.

[6] Park and Owens. Community Forensics: Using Thousands of Generators to Train Fake Image Detectors. CVPR 2025.

Thanks the authors for the satisfying results. I therefore will raise my score to 4

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response (1/2)

We sincerely thank all reviewers for their time and constructive feedback. We are encouraged that the reviewers found our texture measures novel (z4Fh), the analysis sound and thorough (4ivZ, 423X), the benchmark useful for future work (4ivZ), and the paper clearly written (z4Fh, 423X).

A shared concern was whether the gains on CβC_\beta and CrC_r, which RealSynz optimizes, translate into images that look more natural (4ivZ, z4Fh, 423X). We therefore added evaluations that RealSynz does not optimize:

Human perceptual study (423X, z4Fh, 4ivZ). 20 participants compared 25 image pairs before and after RealSynz and chose the RealSynz image in 73.4% of 500 trials; 18 of 20 participants and all 25 images favored RealSynz (Section 3.2, Table 2).

AI-generated image detectors (423X, z4Fh, 4ivZ). Three off-the-shelf detectors (AIDE, SSP, Community Forensics) classify more images as real after RealSynz for all five generators (Section 3.2, Table 1).

Independent texture statistics (z4Fh, 423X). RealSynz moves Portilla–Simoncelli statistics and GLCM correlation closer to real images, although neither is used in its training (Appendix D.3, Table 17).

Below we detail the three results shared across reviewers. All new results have been included in our revised version, at the sections and tables indicated above. The new changes are marked in blue.


Human perceptual study (423X, z4Fh, 4ivZ)

We conducted a human perceptual study with 20 participants, each of whom evaluated the same 25 image pairs, each showing an unedited image and the same image after RealSynz, and selected the one they perceived as more natural and realistic.

ComparisonTrials choosing RealSynzParticipants favoring RealSynzImages favoring RealSynz
RealSynz vs. Unedited73.4% (367/500)90% (18/20)100% (25/25)

A participant or image favors RealSynz if RealSynz was chosen in more than half of its trials. The preference for RealSynz over the unedited images is significant and consistent. We have added the study to Section 3.2 and Table 2 of our revised version.

AuthorsOpenReview ↗

General Response (2/2)

AI-generated image detectors (423X, z4Fh, 4ivZ)

We selected three AI-generated image detectors:

  1. AIDE [1] combines the highest- and lowest-frequency patches of an image, selected by DCT, with a CLIP embedding of the whole image.
  2. SSP [2] classifies the noise residual of the flattest 32-pixel patch of an image.
  3. Community Forensics [3] is a whole-image detector trained on images from thousands of generators to generalize to unseen ones.

We used the authors' released official weights and evaluation code. For each generator, we evaluate its 1,000 generated images before and after RealSynz. The table below reports the percentage of images classified as real, where higher values indicate that fewer images are detected as synthetic.

GeneratorAIDESSPCommunity Forensics
FLUX.2 Klein22.6 →\rightarrow 87.521.8 →\rightarrow 62.957.4 →\rightarrow 80.1
Qwen-Image3.40 →\rightarrow 26.839.2 →\rightarrow 56.334.6 →\rightarrow 67.7
Z-Image3.50 →\rightarrow 41.63.80 →\rightarrow 40.051.4 →\rightarrow 70.8
FLUX.1-dev10.0 →\rightarrow 18.14.70 →\rightarrow 33.48.50 →\rightarrow 33.6
SANA16.8 →\rightarrow 38.326.0 →\rightarrow 41.00.80 →\rightarrow 2.40

After applying RealSynz, all three detectors classify more images as real for all five generators. We have added these results to Section 3.2 and Table 1 of our revised version.


Independent texture statistics (z4Fh, 423X)

For a more comprehensive evaluation, we introduced two additional independent metrics: Portilla–Simoncelli (PS) statistics [4] and gray-level co-occurrence matrix (GLCM) correlation [5], neither of which is used during training or inference.

PS statistics are a few hundred wavelet statistics from which visually similar textures can be synthesized, and GLCM correlation measures how strongly the gray levels of neighboring pixels co-vary. We compute both on 4,000 surface crops of size 128×128128\times128 from the 1,000 FLUX.2 Klein generated images, before and after RealSynz, and on crops from real images. We then measure the distance between the statistics of generated and real crops, normalized so that two random halves of the real images are 1.0 apart. Lower values are closer to real images.

Metric (↓\downarrow)UneditedRealSynz
PS distance1.701.42
GLCM correlation distance3.401.30

RealSynz reduces both distances. Note that neither statistic is used in the training of RealSynz. We have added these results to Appendix D.3 and Table 17 of our revised version.


[1] Yan et al. A Sanity Check for AI-generated Image Detection. ICLR 2025.

[2] Chen et al. A Single Simple Patch is All You Need for AI-generated Image Detection. arXiv 2024.

[3] Park and Owens. Community Forensics: Using Thousands of Generators to Train Fake Image Detectors. CVPR 2025.

[4] Portilla and Simoncelli. A Parametric Texture Model Based on Joint Statistics of Complex Wavelet Coefficients. IJCV 2000.

[5] Haralick et al. Textural Features for Image Classification. IEEE Transactions on Systems, Man, and Cybernetics 1973.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.