Benchmarks · Updated 14 November 2026
Sightova benchmark results
How accurately Sightova tells AI-generated images and videos apart from real ones. The image results come from the model that runs on sightova.com and in our API. The video results come from our video model, which is in private testing. Both use the same 0.85 threshold. We keep test data out of training, and where a public test set overlaps with our training data, we say so next to the result.
Contents
Sightova vs commercial and research detectors
On 12 November 2026 we sent the same 100 images to Sightova, four commercial detectors and eight open research models. We removed all metadata first, so every detector had to judge the pixels alone.
How often each detector got it right
Balanced accuracy on 100 images: 30 real photos, 40 AI images and 30 real photos with an AI edit
SightovaCommercial detectorOpen research model
Sightova
99.29 (Sightova)
Competitor 1
commercial detector
98.57 (Competitor 1)
Competitor 2
commercial detector
98.33 (Competitor 2)
Competitor 3
commercial detector
97.14 (Competitor 3)
Competitor 4
commercial detector
88.57 (Competitor 4)
AIDE (GenImage)
open research model
81.43 (AIDE (GenImage))
SPAI
open research model
80.71 (SPAI)
OmniAID v1 (Mirage)
open research model
72.86 (OmniAID v1 (Mirage))
B-Free
open research model
67.86 (B-Free)
OmniAID v2 (Mirage)
open research model
66.43 (OmniAID v2 (Mirage))
Community Forensics 384
open research model
63.33 (Community Forensics 384)
CO-SPY (DRCT-2M)
open research model
56.19 (CO-SPY (DRCT-2M))
RINE (ProGAN 4-class)
open research model
53.57 (RINE (ProGAN 4-class))
The images
We picked the 100 images from our frozen hold-out set and fixed the list before any detector ran. 40 are AI images from 9 recent generators. 30 are real photos: 15 straight from a camera (the RAISE set) and 15 from the web before 2022. The other 30 are real photos where Google's Nano Banana changed one object or piece of text (from the Pico-Banana-400K set). We stripped C2PA and EXIF metadata from every file.
How it is scored
Balanced accuracy counts both kinds of mistake equally: calling a real photo AI, and missing an AI or AI-edited image. It is the average of the two pass rates. For the commercial detectors we used each vendor's own verdict. We ran each research model with its authors' released code and checkpoint, with no extra training, at its default threshold of 0.5. The brackets show what each checkpoint was trained on.
Generators: FLUX.2 max · Seedream 5.0 Pro · Grok Imagine Image 2.0 · Qwen-Image 3 Pro · Recraft v4.1 · Krea 2 Large · Microsoft MAI-Image 2.6 · Meta Muse Image · Sourceful Riverflow v2.5 Pro. How each detector was scored
We don't name the commercial detectors on this page. Contact us for the full results, including which detectors we measured against.
Contact usPublic image benchmarks
Researchers built these test sets to compare AI image detectors. On 12 November 2026 we ran the full public test set of each one through our live model and put the result next to a reference result from the research papers. For Synthbuster we ran the reference detector ourselves, because its paper only covers part of the test.
Sightova vs published results
Each benchmark is scored with its own main metric
SightovaReference result
X-AIGD
Recall (share of AI images detected)
98.71 (Sightova)
95.0 (Reference result)
Mirage-Test
Mean accuracy across five content domains
97.67 (Sightova)
88.4 (Reference result)
Community Forensics evaluation set
Mean accuracy per generator
96.07 (Sightova)
89.3 (Reference result)
Synthbuster + RAISE-1k
Balanced accuracy
96.37 (Sightova)
95.2 (Reference result)
| Benchmark | Images | Metric | Sightova | Reference | Reference model |
|---|---|---|---|---|---|
X-AIGD ICLR 2026 · Recall (share of AI images detected) · 27,057 AI images | 27,057 AI | Recall (share of AI images detected) | 98.71 | 95.0 Gram-Net (X-AIGD paper) | Gram-Net (X-AIGD paper) |
Mirage-Test ICML 2026 · Mean accuracy across five content domains · 49,000 images | 49,000 | Mean accuracy across five content domains | 97.67 | 88.4 OmniAID-Mirage | OmniAID-Mirage |
Community Forensics evaluation set CVPR 2025 · Mean accuracy per generator · 51,794 images | 51,794 | Mean accuracy per generator | 96.07 | 89.3 Community Forensics ViT-S/384 | Community Forensics ViT-S/384 |
Synthbuster + RAISE-1k IEEE OJSP 2023 · Balanced accuracy · 9,999 images | 9,999 | Balanced accuracy | 96.37 | 95.2 B-Free, measured by Sightova | B-Free, measured by Sightova |
| Real photos called real (999) | 100.00 | 94.6 | |||
| AI images detected (9,000) | 92.73 | 95.8 |
Where the reference results come from: Gram-Net as reported in the X-AIGD paper (Xiao et al., ICLR 2026); OmniAID-Mirage (Guo et al., ICML 2026); Community Forensics ViT-S/384 (Park & Owens, CVPR 2025) on the released evaluation set; B-Free (Guillaro et al., CVPR 2025): measured by Sightova with the authors' released code and checkpoint on all nine Synthbuster generators and the RAISE-1k photos.
X-AIGD only publishes its AI-generated test images, so its score is the share of AI images caught. We left out 42 Community Forensics images that are a single pixel in size, and one RAISE photo in Synthbuster that cannot be opened. Both detectors were scored without them.
None of these test images were used to train Sightova. Like the Community Forensics reference model, Sightova was trained on the Community Forensics training split, which is kept separate from its evaluation set. X-AIGD, Mirage-Test and Synthbuster are not in our training data at all.
Robustness to edits
We took one AI-generated photo and changed it in ten ways that happen when images are shared online or deliberately disguised: heavy JPEG and WebP compression, shrinking it to 320 pixels, blur, motion blur, camera-style noise, a screenshot with a colour shift, an invisible watermark, and saving the file twice. We then scanned each edited file with Sightova's live model. All ten versions of the AI image were still detected as AI, at 95 to 96%, and all ten versions of a real phone photo stayed real, with an AI probability of 0.03 to 0.04%.
AI-generated photo
- OriginalSightovaAI 95.8%
- JPEG quality 25SightovaAI 95.8%
- WebP quality 20SightovaAI 96.1%
- Downscaled to 320 pxSightovaAI 95.3%
- Gaussian blurSightovaAI 95.8%
- Motion blurSightovaAI 95.8%
- Sensor noiseSightovaAI 96.1%
- Screenshot: colour cast, rescale, JPEGSightovaAI 95.9%
- Invisible watermarkSightovaAI 95.9%
- Re-encoded twice (JPEG 80, then JPEG 60)SightovaAI 96.2%
Real photo
- OriginalSightovaReal · AI 0.03%
- JPEG quality 25SightovaReal · AI 0.03%
- WebP quality 20SightovaReal · AI 0.03%
- Downscaled to 320 pxSightovaReal · AI 0.03%
- Gaussian blurSightovaReal · AI 0.03%
- Motion blurSightovaReal · AI 0.04%
- Sensor noiseSightovaReal · AI 0.04%
- Screenshot: colour cast, rescale, JPEGSightovaReal · AI 0.03%
- Invisible watermarkSightovaReal · AI 0.03%
- Re-encoded twice (JPEG 80, then JPEG 60)SightovaReal · AI 0.03%
AI-generated example: created by Sightova with OpenAI GPT Image 2.6, November 2026.
Real photo: taken by Sightova on a Samsung Galaxy S25 Ultra, November 2026; not part of any training data.
How the edits were made: blur with a radius of 0.2% of the longest side, motion blur over 1% of it, and camera-style noise with a strength (sigma) of 8. The screenshot adds a slight colour shift, scales the image to 85% and saves it as JPEG 85. The invisible watermark is a standard DWT-DCT watermark. Every edit was made on the full-size file and scanned with our live model and threshold.
Detection by image size
We shrank about 5,500 test images, both AI-generated and real, to smaller and smaller sizes, from 192 pixels on the longest side up to the original size, and scanned every version with Sightova's live model. Smaller images hold less detail, so detection gets harder as they shrink. Sightova still scores 91% at 192 pixels and 93% at 256 pixels, rising to 98% at full size.Every line in the chart is Sightova. The test images come from three public test sets that researchers built to compare AI detectors: Synthbuster + RAISE, X-AIGD and Mirage-Test. Each coloured line shows how Sightova scores on one of them, and the blue line is the average of the three. X-AIGD scores highest because it holds only AI images, so there are no real photos to get wrong.
Sightova, averageSynthbuster + RAISEX-AIGDMirage-Test
| Longest side of the image (px) | Sightova, average | Synthbuster + RAISE | X-AIGD | Mirage-Test |
|---|---|---|---|---|
| 192 | 91.1 | 90.9 | 91.6 | 90.8 |
| 256 | 92.9 | 93.0 | 93.3 | 92.4 |
| 384 | 95.4 | 95.0 | 96.1 | 95.1 |
| 512 | 96.5 | 95.5 | 97.5 | 96.4 |
| 768 | 97.7 | 96.8 | 98.5 | 97.8 |
| 1024 | 98.0 | 96.5 | 99.0 | 98.5 |
| Full size | 97.9 | 96.5 | 99.0 | 98.0 |
Detection by frame and cropping
Images are often cropped, framed or placed on a phone-sized canvas before they are shared. We cropped and framed the same 5,549 test images in 11 ways and scanned every version with Sightova's live model.Sightova catches 97% of the AI images when they are untouched. With most crops, a white border or a meme caption it still catches 95 to 97%. Black bars, a blurred background and a 50% corner crop land at 94 to 95%, and keeping only the centre quarter of the image, the hardest edit, still catches 91%. Fewer than 3% of real photos are wrongly flagged after any of the edits.
AI-generated photo
- OriginalSightovaAI 95.8%
- Cropped to 75%SightovaAI 95.5%
- Cropped to 50%SightovaAI 95.6%
- Cropped to 25%SightovaAI 95.8%
- Corner crop, 50%SightovaAI 95.6%
- Square cropSightovaAI 95.3%
- Tall crop, 9:16SightovaAI 96.1%
- Wide crop, 16:9SightovaAI 95.3%
- White borderSightovaAI 96.1%
- Black barsSightovaAI 96.4%
- Blurred backgroundSightovaAI 96.1%
- Meme captionSightovaAI 95.3%
Real photo
- OriginalSightovaReal · AI 0.03%
- Cropped to 75%SightovaReal · AI 0.03%
- Cropped to 50%SightovaReal · AI 0.03%
- Cropped to 25%SightovaReal · AI 0.04%
- Corner crop, 50%SightovaReal · AI 0.03%
- Square cropSightovaReal · AI 0.02%
- Tall crop, 9:16SightovaReal · AI 0.03%
- Wide crop, 16:9SightovaReal · AI 0.02%
- White borderSightovaReal · AI 0.05%
- Black barsSightovaReal · AI 0.17%
- Blurred backgroundSightovaReal · AI 0.13%
- Meme captionSightovaReal · AI 0.06%
AI-generated example: created by Sightova with OpenAI GPT Image 2.6, November 2026.
Real photo: taken by Sightova on a Samsung Galaxy S25 Ultra, November 2026; not part of any training data.
AI images caught
3,550 AI-generated images, scanned after each edit. Higher is better.
Original
no edit
97.1%
Cropped to 75%
96.1%
Cropped to 50%
95.0%
Cropped to 25%
90.9%
Corner crop, 50%
94.6%
Square crop
96.4%
Tall crop, 9:16
96.1%
Wide crop, 16:9
96.6%
White border
95.5%
Black bars
94.3%
Blurred background
93.8%
Meme caption
96.6%
Real photos wrongly flagged
1,999 real photos, scanned after each edit. Lower is better.
Original
no edit
1.4%
Cropped to 75%
1.2%
Cropped to 50%
1.0%
Cropped to 25%
1.0%
Corner crop, 50%
1.0%
Square crop
0.9%
Tall crop, 9:16
1.1%
Wide crop, 16:9
1.7%
White border
0.7%
Black bars
2.5%
Blurred background
2.9%
Meme caption
0.7%
The test used 5,549 images from Synthbuster + RAISE, X-AIGD and Mirage-Test: 3,550 AI-generated and 1,999 real. The crops keep the middle 75%, 50% or 25% of the image, and the corner crop keeps 50% from one corner. Black bars and the blurred background put a landscape image on a tall 9:16 canvas, or a portrait image on a wide 16:9 canvas. No edited file was compressed again.
Video detection
Private testingVideo has its own model, separate from the image model. It is in private testing and not yet available in the dashboard or the API. The results below were measured on 14 November 2026. A video counts as AI when at least 75% of the frames we sample from it score 0.85 or higher.
99.80
AIGVDBench score (AUC)
16 video generators against real video
97.45%
accuracy on GenBuster-Bench++
a public test we never trained on
0.3%
of real videos wrongly flagged
2,360 real test clips, originals and chat-app copies
83.7%
of AI-labelled TikToks caught
0.1% of real TikToks wrongly flagged
Public video benchmarks
We scored the original files from each benchmark's test clips, which we never trained on. For AIGVDBench and DF26 we report AUC, which measures how well the model ranks AI videos above real ones, regardless of the threshold. For Ivy-Fake and GenBuster-Bench++ we report accuracy, using the 75% rule above.
Public video benchmarks
Each benchmark is scored with its own main metric
AIGVDBench
AUC, average over 16 generators
99.80 (Sightova video model)
DF26
AUC
96.68 (Sightova video model)
Ivy-Fake
accuracy
95.96 (Sightova video model)
GenBuster-Bench++
accuracy
97.45 (Sightova video model)
The test clips
AIGVDBench: 300 clips from each of 16 generators, and 600 real clips. DF26: 966 AI clips and 108 real. Ivy-Fake: 1,250 AI clips from 12 generators and 1,250 real. GenBuster-Bench++: 1,000 AI clips and 1,000 real.
Overlap with our training data
We trained on other parts of AIGVDBench and DF26, but never on their test clips, and we have not trained on any GenBuster data. 306 of Ivy-Fake's real clips are also in our training data. Without them, balanced accuracy on Ivy-Fake is 95.95% and AUC is 97.70.
| Benchmark | Metric | Result |
|---|---|---|
| AIGVDBench | AUC, average over 16 generators | 99.80 |
| DF26 | AUC | 96.68 |
| Ivy-Fake | Accuracy | 95.96% |
| Ivy-Fake | AI videos caught | 92.0% |
| Ivy-Fake | Real videos wrongly flagged | 0.1% |
| Ivy-Fake | AUC | 97.70 |
| Ivy-Fake, without the 306 real clips we trained on | Balanced accuracy | 95.95% |
| Ivy-Fake, without the 306 real clips we trained on | AUC | 97.70 |
| GenBuster-Bench++ | Accuracy | 97.45% |
| GenBuster-Bench++ | AI videos caught | 96.1% |
| GenBuster-Bench++ | Real videos wrongly flagged | 1.2% |
| GenBuster-Bench++, chat-app copies | AI videos caught | 88.9% |
| GenBuster-Bench++, chat-app copies | Real videos wrongly flagged | 1.6% |
Video generators
These results come from our frozen video test: 19,124 clip versions that we set aside before training. We score every clip twice, once as the original file and once as a chat-app copy, because messaging apps re-compress video and strip out detail.
AI videos caught, by generator
Share of each generator's clips that the model flags as AI
Original fileChat-app copy
Wan 2.6
79 clips · DF26
100.0% (Original file)
100.0% (Chat-app copy)
Grok Imagine
80 clips · DF26
98.8% (Original file)
100.0% (Chat-app copy)
Veo 3.1
80 clips · DF26
98.8% (Original file)
97.5% (Chat-app copy)
Veo 3
300 clips
99.0% (Original file)
91.3% (Chat-app copy)
Seedance 2.0
280 clips
93.6% (Original file)
88.2% (Chat-app copy)
Seedance 2.5
180 clips
92.2% (Original file)
83.3% (Chat-app copy)
Kling 3.0
79 clips · DF26
100.0% (Original file)
81.0% (Chat-app copy)
Sora 2
291 clips
91.1% (Original file)
75.9% (Chat-app copy)
Across every AI set in the frozen test, including older face swaps and open-source image-to-video clips, the model catches 85.3% on average.
Real videos wrongly flagged
The share of real videos that the model wrongly calls AI. Lower is better.
| Real video set | Original | Chat-app copy |
|---|---|---|
| Web video (OpenVid): 600 clips | 0.5% | 0.5% |
| People talking to camera (DF26): 108 clips | 0.0% | 0.9% |
| YouTube video, light compression (FaceForensics++ c23): 140 clips | 0.7% | 0.0% |
| YouTube video, heavy compression (FaceForensics++ c40): 140 clips | 0.0% | 0.0% |
| Phone video (FloreView): 108 clips | 0.0% | 0.0% |
| Phone video (VISION): 84 clips | 0.0% | 0.0% |
| All real clips, originals and copies: 2,360 | 0.3% |
How we test video
- The test sets
- We set our frozen video test (19,124 clip versions) aside before training and never train on it. The AI videos come from AIGVDBench (16 generators), DF26 (Grok Imagine, HunyuanVideo 1.5, Kling 3.0, LTX 2.3, Veo 3.1, Wan 2.2, Wan 2.6), FaceForensics++ face swaps and reenactments, Sora 2, Veo 3 and 3.1, Seedance 1.0 to 2.5, Kling and Sora. The real videos come from OpenVid, DF26, FaceForensics++ and phone-camera video (FloreView, VISION).
- How a video is judged
- The model scores frames sampled evenly through the video. A video counts as AI when at least 75% of those frames score 0.85 or higher. Every test on this page uses the same rule.
- Original file and chat-app copy
- The original is the file exactly as the generator, camera or dataset produced it. The chat-app copy is the same clip re-encoded with H.264 at 848 px, CRF 28, which is close to what messaging apps do to video.
- TikTok videos
- We split public TikTok videos by account, so no video from a test account was used in training. The AI videos are ones TikTok itself labels as AI-generated. The real videos were uploaded before July 2022.
- The scores
- AUC measures how well the model ranks AI videos above real ones by their average frame score, and does not depend on the 0.85 threshold. Accuracy, catch rates and false alarm rates use the 75% rule.
- The model
- Sightova video model, November 2026. It is trained for video only and does not change the image model.
Cite these results
Sightova. "Sightova benchmark results: AI image and video detection." Updated 14 November 2026. https://sightova.com/benchmark
Researchers, analysts and enterprise teams can contact us to test our image model on their own data or to discuss video detection.
About the detector comparisonHow big the test is, what we trained on, and how we read each detector's answer
- The test has 100 images, so one wrong answer moves a detector's score by 0.7 points (an AI or edited image) to 1.7 points (a real photo).
- We never trained on these exact images. Our training data does include other images from the same generators, made with similar prompts, and other images from the Pico-Banana-400K edit set. The RAISE camera photos are the only part that comes from a source we have never trained on.
- How each detector was scored (run on 12 November 2026):
- Sightova: Our live image model at the 0.85 threshold. This is the same verdict sightova.com and the API return.
- Competitors 1 to 4: we sent each image to the detector's own API and used its own verdict. Answers meaning AI-generated or partly AI-generated counted as AI. Answers meaning real, unsure, or edited with ordinary photo software counted as not AI. One detector only returns a score from 0 to 1, so for that one a score of 0.5 or higher counted as AI. Every verdict came from the detectors' models, not from metadata or watermarks.
- We ran each research model with its authors' released code and checkpoint, with no extra training, at its default threshold of 0.5. The brackets show what each checkpoint was trained on.
- AIDE (GenImage): Yan et al., "A Sanity Check for AI-generated Image Detection", ICLR 2025.
- SPAI: Karageorgiou et al., "Any-Resolution AI-Generated Image Detection by Spectral Learning", CVPR 2025.
- OmniAID v1 (Mirage): Guo et al., "OmniAID: Decoupling Semantics and Artifacts for Universal AI-Generated Image Detection in the Wild", ICML 2026 (trained on Mirage-Train, CLIP ViT-L/14-336).
- B-Free: Guillaro et al., "A Bias-Free Training Paradigm for More General AI-generated Image Detection", CVPR 2025.
- OmniAID v2 (Mirage): Guo et al., same paper as OmniAID v1, ICML 2026: the authors' newer recommended release.
- Community Forensics 384: Park & Owens, "Community Forensics: Using Thousands of Generators to Train Fake Image Detectors", CVPR 2025 (ViT-S at 384 px, trained on 4,803 generators).
- CO-SPY (DRCT-2M): Cheng et al., CVPR 2025 (checkpoint trained on DRCT-2M, Stable Diffusion 1.4).
- RINE (ProGAN 4-class): Koutlis & Papadopoulos, "Leveraging Representations from Intermediate Encoder-blocks for Synthetic Image Detection", ECCV 2024.