Pixel Heatmap Analysis: How Sightova Shows Exactly Where AI Edited an Image

Binary 'real or fake' verdicts aren't always enough. Sightova's heatmap shows which regions of an image were most likely changed by AI — a critical tool for partial edits and deepfakes.

Sightova Research4 min read
Contents
  1. 01How It Works
  2. 02Why This Matters
  3. 03How It Learned
  4. 04Viewing the Heatmap
  5. 05When It's Most Useful

Most AI detection tools give you a single answer: real or fake. That works for fully synthetic images — a Midjourney landscape, a DALL-E product shot. But what about the harder cases? A real photograph where someone's face has been swapped. A genuine product image with AI-enhanced backgrounds. A news photo where a critical detail has been digitally altered.

These partial manipulations are where the real damage occurs, and a binary classifier can't tell you what's been changed. That's why we built pixel-level heatmap analysis into Sightova.

How It Works

Sightova's detection model doesn't only score the image as a whole. It also scores it region by region and turns those scores into a map: each cell carries a value between 0 and 1, the model's confidence that this part of the image was generated or changed by AI.

The result is a visual overlay that maps directly onto the original image. Regions the model identifies as unchanged appear in green. Regions flagged as AI-edited appear in red. The intensity of the color reflects the model's confidence level.

Why This Matters

Consider a real-world scenario: an insurance company receives a damage claim with photographic evidence. The image looks authentic at a glance. Our primary model flags it with moderate AI probability. The heatmap reveals exactly what triggered the detection — a specific region of the image shows clear signs of generative manipulation, while the rest of the photograph is genuine.

This is information that a simple "87% AI-generated" score cannot convey. The heatmap transforms abstract probability into spatial, actionable intelligence.

How It Learned

The model learned from pairs of images: originals alongside AI-edited versions of the same photo, covering a wide range of editing techniques such as face swaps, background replacements, object insertion and texture regeneration. That is how it learned the boundary and texture traces that appear where a generative model has changed part of a real photograph.

Small edits inside real photos are still the hardest case for any detector. Our testing methodology reports how the model performs on public benchmarks of AI edits, including the edit types it still misses.

Viewing the Heatmap

In the Sightova dashboard, the heatmap appears in the Manipulation Heatmap section of an AI image result, on Starter and above. The visualization overlays directly on your uploaded image, giving you immediate spatial context. Summary statistics — the share of the image flagged, mean confidence and peak confidence — provide a quantitative complement to the visual overlay.

For documents, the forgery heatmap is also available through the API: add include_heatmap=true to a document tampering request to get the grid back with the verdict.

When It's Most Useful

Heatmap analysis is particularly valuable for:

  • Deepfake detection — identifying face-swapped regions in otherwise authentic photos
  • Forensic investigation — pinpointing exactly which elements of an image were altered
  • Content moderation — distinguishing between fully synthetic content and manipulated real photos
  • Insurance and legal — providing visual evidence of where manipulation occurred

A binary verdict tells you if something is wrong. The heatmap tells you where.