About TagFake

Image type classification

TagFake uses CLIP zero-shot classification to estimate whether an image is closer to PHOTO or ART. The result is shown as the PHOTO/ART indicator on the analysis page.

AI Probability detection

The AI probability is computed from evaluations by multiple detectors.

The following detectors are used in analysis, including candidates under evaluation.

• Frequency Analysis (FFT)
• Noise Analysis
• Error Level Analysis (ELA)
• DIRE (Diffusion Reconstruction Error)
• UniversalFakeDetect (CSID)
• DeepFakeDetectorV2
• AI Image Detector
• TruFor

NSFW Detection

TagFake uses its own method to detect whether posted images contain sensitive (NSFW) content. Images flagged as NSFW are hidden by default on list pages, and are shown only in pixelated form even when NSFW display is enabled. NSFW detection does not affect the AI generation or manipulation probability scores.

AI Probability

Analysis example

The two images below were actually analyzed by this service. On list pages, images with an AI-generated probability below 50% appear in the left column, images at 50% or higher appear in the right column, and the purple bar at the bottom of each card shows the AI-generated probability.

AI-generated probability < 50%

ART

AI0.0%

A photo of Vermeer's Girl with a Pearl Earring

AI-generated probability ≥ 50%

ART

AI94.7%

An image generated with ChatGPT

How to read heatmaps

The analysis page shows an AI basis heatmap, which layers the evidence from the detectors used for AI detection according to each detector's weight. Darker red areas indicate regions with stronger evidence of AI generation. The heatmap is a reasoning hint from the detectors - it does not prove that a specific region was AI-generated.

Cases where results are less reliable

The following kinds of images can fall outside what the detectors were designed for, so the AI Probability should be read with extra caution.

• Photos that contain an embedded illustration, anime poster, or painting

  The embedded artwork can resemble AI-generation traits and raise the AI Probability.

• Real photos with stickers, emoji, or text overlays added

  The overlaid elements can introduce artificial edges and textures that distort the AI Probability.

• Images with heavy color filters, skin-smoothing, stylization, generative fill, or composite edits applied

  Strong edits can mask or mimic the pixel-level traces the detectors rely on, making the AI Probability less reliable. Light adjustments that only re-encode the image - color or exposure correction, resizing, format conversion - normally have little effect.

• Screenshots of phone or PC screens, user interfaces, web pages, or code

  Screen capture and display artifacts are outside the detectors' expected input, so the AI Probability becomes less reliable.

• Images that have been re-saved or re-encoded many times, especially as low-quality JPEG

  Compression noise can bury or resemble the traces the detectors look for, so the AI Probability can swing in either direction.

• Charts, graphs, slides, and other diagrammatic images

  These images fall outside the photo and illustration ranges the detectors are designed for, so the AI Probability becomes less reliable.

• 3D-rendered images (both stylized and photorealistic)

  Both stylized and photorealistic renders can fall outside the detectors' training distribution, making the AI Probability less reliable.

• Collages that mix photos and illustrations in one frame

  Mixed regions with different visual properties can make detector judgments unstable.

• Memes and other images dominated by overlaid text

  Large text areas can cover the visual traces the detectors rely on, making the AI Probability less reliable.

TagFake does not detect these cases automatically, so the analysis page still shows a score for such images. Read results for images like these with extra caution rather than at face value.