Beyond Detection: How Sightova Combines Generator ID, Watermarks and Forensic Analysis in One Scan

A single AI image scan on Sightova combines our detection model, invisible-watermark checks, content credentials, generator identification and a manipulation heatmap. Here's how they fit together, and where NSFW checks come in.

Sightova Research4 min read
Contents
  1. 01What Runs in One AI Image Scan
  2. 02Why Multiple Signals Matter
  3. 03Where NSFW Checks Fit
  4. 04For Platforms
  5. 05The Dashboard
  6. 06The Road Ahead

When we designed Sightova, we made a deliberate decision: a single scan should give you everything you need to judge whether an image is real. Not just whether it's AI-generated, but what made it, what the file says about its origin, and which regions were changed.

What Runs in One AI Image Scan

Authenticity detection — Our detection model analyzes the image's pixels to estimate whether it was generated by AI or captured by a camera. This produces the AI probability that every result is built on.

Invisible watermarks — Every scan checks for the watermarks Google (SynthID) and OpenAI hide in the images their models make. When one is found, it is close to proof and names the source. See watermark forensics.

Content credentials — We read the file's C2PA manifest and EXIF metadata. A signed manifest can declare an image AI-generated, name the tool that made it, or record that it was edited.

Generator identification — When the evidence allows, Sightova names the tool behind an image: from the content credentials, from the watermark, or from our model's own best guess when it is confident enough. Knowing which tool created an image is often as important as knowing it's synthetic.

Manipulation heatmap — The detection model also scores the image region by region, so a result can show where an image was changed, not only that it was. See pixel heatmap analysis.

Why Multiple Signals Matter

Each of these capabilities exists in isolation elsewhere, but isolated tools miss the correlations. When the model flags an image at 73% AI probability, the heatmap can reveal that only one region — a face, a background — was changed. A watermark or a signed credential can settle the question outright and name the generator. Together, these signals paint a picture no single check can provide alone.

Where NSFW Checks Fit

Content safety is a different question from authenticity: what does the image show, rather than is it real? Sightova's NSFW model runs as its own check, in the dashboard and through its own API endpoint, and returns a single probability that the image is explicit.

Platforms that need both run both. An image can be real and explicit, AI-generated and harmless, or, in the case of synthetic explicit content and deepfake nudes, both at once. For how the NSFW score behaves on everyday photos, see NSFW detection without the false alarms.

For Platforms

For content moderation teams, the API returns each check as structured JSON. Your pipeline can make nuanced decisions: quarantine images that are both AI-generated and NSFW, escalate partially manipulated photos for human review, or label AI-generated content for transparency. AI detection, NSFW and document checks each have their own endpoint, so you call only the ones you need; see the API docs.

The Dashboard

In the web dashboard, an AI image result starts with a short verification timeline (watermark, content credentials, detection model) and a single verdict. The details sit in sections below it: the AI evidence, image info, the manipulation heatmap (Starter and above) and the file's metadata, including C2PA credentials and EXIF data. NSFW, violence and document tampering checks have their own tools in the same dashboard.

The Road Ahead

We're continuing to expand each of these checks: training on emerging generators, improving the heatmap, and adding new safety classifications.

Our goal is straightforward: upload one image, understand everything about it.