DALL·E image
detector

DALL·E, developed by OpenAI, has two major generations in active use: DALL·E 2 (launched 2022) and DALL·E 3 (2023, integrated into ChatGPT). Both use diffusion models but with different architectures that leave distinct forensic fingerprints.

DALL·E 2 generates 1024x1024 images with a characteristic smooth texture. Low-frequency areas show a softening pattern absent in real photographs. Backgrounds tend to be abstract or diffuse, and human faces are lower quality than Midjourney.

DALL·E 3 significantly improves visual quality and text-prompt adherence. Its images are more diverse in composition but maintain distinctive DCT patterns. Rendered text areas — a strength of DALL·E 3 — leave specific artifacts in the frequency spectrum that serve as detection signals.

Three independent forensic layers run in parallel:

1. Pixel-level spectral analysis. DALL·E images show anomalies in high-frequency coefficients. The residual noise is uniform and correlated — characteristic of diffusion models — in contrast to the random thermal noise from real camera sensors (PRNU).

2. EXIF metadata inspection. DALL·E images downloaded from ChatGPT or the API lack camera metadata. OpenAI has started including C2PA credentials on some DALL·E 3 outputs, which ScanTrace detects automatically.

3. Contextual reasoning. A language model receives numeric data from layers 1 and 2 and generates a human-readable explanation. It never sees the image — only interprets forensic signals.

ScanTrace detection rates as of April 2026:

Stable Diffusion XL: 97% — the most pronounced artifacts, easiest to detect.

DALL·E 3: 95% — distinctive DCT patterns and strong signal in text-rendering areas.

Midjourney v7: 93% — consistent spectral signature despite extreme visual realism.

Flux 1.1 Pro: 91% — the hardest. Weaker signal, requires sufficient resolution (minimum 768x768 recommended).

DALL·E 3 sits in a middle ground: easier than Midjourney or Flux to detect forensically, but its integration into ChatGPT means it produces the highest volume of synthetic images globally.

With over 100 million active ChatGPT users, DALL·E 3 is by far the most widely used image generator by volume. Europol warns that AI-generated images — including those from DALL·E — are increasingly used in identity fraud, political disinformation, and fake social media profiles.

The European Broadcasting Union reported in 2025 that 77% of European newsrooms have inadvertently published synthetic content. For journalists and fact-checkers, verifying every unattributed image is now a professional obligation.

Suspect DALL·E when an image shows: uniform quality without sensor noise, perfectly integrated readable text in the scene (a DALL·E 3 strength), square resolution (1024x1024 or 1792x1024 — DALL·E's native formats), complete absence of EXIF metadata, and compositions that follow very specific textual instructions.

ScanTrace returns a verdict in under 15 seconds — verification that fits any editorial deadline.

Can ScanTrace distinguish between DALL·E 2 and DALL·E 3?

Yes. DALL·E 2 and DALL·E 3 produce different spectral signatures. DALL·E 2 generates 1024x1024 images with a characteristic texture in low-frequency areas. DALL·E 3 improves instruction following and reduces artifacts but maintains DCT patterns unique to OpenAI's architecture.

Are images generated through ChatGPT (which uses DALL·E) detectable?

Yes. ChatGPT uses DALL·E 3 internally. The resulting images carry the same forensic signatures as those generated directly through OpenAI's API. Some may also include C2PA credentials that ScanTrace reads automatically.

Does DALL·E include detectable watermarks?

OpenAI has experimented with invisible watermarking on DALL·E 3 outputs. ScanTrace detects these when present but does not depend solely on them — the spectral analysis works even if the watermark has been stripped.

Does it work on cropped or resized DALL·E images?

Yes. Cropping and resizing reduce the available signal but do not eliminate it. As long as the image retains at least 512x512 pixels from the original, detection accuracy stays above 90%.

Is the DALL·E detector free?

Yes. 10 free analyses per month with no credit card required. Each analysis returns a verdict, 0–1 confidence score and forensic explanation.