See differently
Expert 4 extracts one pooled representation and teacher margin from the image.
Robust AI-image detection / NTIRE 2026
From pixelsto evidence.
One frozen upstream teacher and three project-trained residual heads turn an image into one inspectable review signal.
See the signal
Signal capture 01
Illustrative outputselected graph / v1
Fake probability
0.78
01
frozen Expert 4
03
residual heads
02
native-size routes
0–1
review signal
01The problem
AI images rarely arrive lab-clean.
Screenshots, compression, resizing, noise, and everyday edits can blur the visual traces that detectors learn to recognize.
SynthFlag answers with a fixed native-size route over one upstream representation and a consistent score interface.
Fuse carefully
A native-size route selects the specialist or fixed two-head margin stack.
Report responsibly
The score supports investigation; it does not prove an image’s origin.
02The method
Two routes.
One score.
A frozen Tu et al. Expert 4 supplies a 1,152-dimensional feature and teacher margin. Three compact residual heads correct that margin through a deterministic native-size route before sigmoid produces a signal in the range 0–1.
Component / 011
Expert 4
Frozen SigLIP teacher
- Input
- 384 px
- Head
- 1152-D + 2 logits
Component / 022
Low-res
CIFAKE specialist
- Input
- ≤ 64 px native
- Head
- 1152 → 256 → 1
Component / 033
Epoch 05
Large-image residual
- Input
- 65% stack
- Head
- 1152 → 256 → 1
Component / 044
Epoch 08
Large-image residual
- Input
- 35% stack
- Head
- 1152 → 256 → 1
Routed detector interface
One representation enters. One review signal leaves.
Output range
0–1
03Read the evidence
A score,
not a sentence.
SynthFlag returns a continuous fake-image probability. How that score becomes a decision depends on the use case, operating point, and consequences of being wrong.
- Preserve the raw probability for review.
- Choose thresholds outside protected evaluation data.
- For creator operations, constrain false positives first.
- Pair model output with provenance and human judgment.
Illustrative score anatomy
0.78
Example only.
Not a benchmark result.
0.00 / lower1.00 / higher
Value
0.78
Type
Probability
Source
3 residual heads
TEST1 · public development benchmark
15,000 sources. Clean and damaged.
One deterministic composite corruption per public source produced 30,000 paired predictions for the selected Expert 4 plus three-head runtime shown above.
- unique public images
- 15,000
- paired predictions
- 30,000
- macro clean ROC-AUC
- 0.9324
- macro augmented ROC-AUC
- 0.8773
CIFAKE
Clean → augmented ROC-AUC
SID-Set
Clean → augmented ROC-AUC
WildFake
Clean → augmented ROC-AUC
For TikTok-like creator operations, constrain false positives first: a wrong flag can question authentic work, interrupt distribution or monetization, and create an appeal. TEST1 reports TPR at 1% and 5% FPR. Its fixed 0.5 point is diagnostic, not a universal moderation cutoff.
Inspect TEST1
Day 3 · research interview
A promising clue. Not a final result.
Professor Ng Teck Khim introduced us to camera-forensics signals created by Bayer sampling and demosaicing. We briefly explored local variance and cross-channel statistics, but the interview happened on Day 3 and the early prototype was not strong or stable enough for final use. It did not enter the released model, TEST1, or final claims.
Interview evidence and future direction—not detector performance or researcher endorsement.
Designed for the wild
Real-world inputs.
Repeatable output.
The inference workflow recursively reads common image formats, applies deterministic preprocessing, and writes resumable predictions to a portable CSV with run metadata.
JPEGPNGBMPWebPTIFFrecursive foldersresumable CSVrun metadata
04Use responsibly
Evidence deserves context.
SynthFlag is a research detector. Its scores should support a broader review process—never stand alone as conclusive proof about an image’s origin. For creator operations, false positives are constrained first because authentic work, reach, monetization, and appeals can be affected by an incorrect flag.
01
Retain the original image and metadata.
02
Calibrate a low-FPR threshold for the actual use case.
03
Escalate consequential decisions to human review.
Verdict sequence / 05
Flag for
review.
Turn the detector signal into a reviewable result—then pair it with provenance, context, and human judgment.