Research Demo · ECICE 2026

MetaFake: Residual Late Fusion of Frozen Backbones for AI-Generated Image Detection

MetaFake: a frozen patch detector, a frequency expert and CLIP, fused by a 288K-parameter head.

  • WildRF Benchmark
  • LaDeDa Backbone
  • 16×16 DCT Spectrum
  • CLIP ViT-L/14
  • Residual Fusion Head
Section 01

Upload Photo for Inspection

Drag and drop, click to browse, or paste an image directly from your clipboard. Images are processed in-memory and never saved to disk.

Input Photo Quota: 20 / 20 remaining
Drop image here or click to browse
Supports JPEG, PNG, WebP · Max 10 MB · Up to 40 MP
Inspection Preview No image loaded
Select or drop a photo to begin analysis
File Name
—
Dimensions
—
File Size
—
Input Shape
256 × 256 RGB
Section 02

Verdict & Confidence

The decision is produced by MetaFake's residual fusion head evaluating spatial, frequency, and semantic representations.

Model Decision Device: MPS
Awaiting Input
No analysis yet
— logit: —

Upload an image and run analysis to evaluate deepfake probability.

Threshold: 0.383
0.0 (Authentic / Real) Decision Boundary 1.0 (AI-Generated / Fake)
LaDeDa Backbone
— ms
CLIP ViT-L/14
— ms
DCT Transform
— ms
Fusion Heads
— ms
Total Latency: — ms
Feature extraction runs once and is shared across all 15 variant heads.
Section 03

End-to-End Pipeline Architecture

Follow the flow of information: image input feeds the three frozen experts, converging into 5 fusion inputs, evaluated by the residual MLP head.

Pipeline Execution Graph
LaDeDa Spatial DCT Frequency CLIP Semantics
Input Photo eval_transform 256 × 256 ImageNet Norm Expert: LaDeDa P(fake): — Patch Std σ: — Patch Max: — Expert: DCT Freq 16×16 Block Spectrum P(fake): — Expert: CLIP Semantics ViT-L-14 Linear Probe P(fake): — 5 Fusion Inputs (v) LaDeDa z: — patch std σ: — patch max: — DCT d: — CLIP c: — Residual Fusion s = z + g(v) g = — s = — 288K MLP Head Sigmoid σ(s) — P(fake) Thresh: —
LaDeDa Patch Heatmap (~30×30 Grid) Std: — · Max: —
70%
Higher anomaly values (yellow/red) highlight localized synthesis artifacts detected by LaDeDa's patch discriminator.
16×16 Frequency Spectrum (|DCT|) Hover cell for value
Top-Left: Low Freq (DC) Bottom-Right: Nyquist (High)
Mean absolute DCT coefficients computed across 16×16 grayscale blocks. Shown for transparency: in the paper's ablations the DCT expert adds no measurable gain.
Section 04

Who Decided: Attribution & Gradient Dynamics

Understand how each expert influenced the final decision. In MetaFake, residual fusion anchors on LaDeDa, with CLIP and DCT providing targeted corrections.

Decision Contribution Correction g: —
Input Feature Value Sensitivity (∂s/∂v) Relative Weight
∂s/∂v represents the local gradient of the fused score with respect to each input feature.
Section 05

Model Ablations & Architecture Variants

Compare predictions across all 15 trained variants on this exact photo. The paper's Table 4 lists 10 of them; the full set is in RESULTS.md. Click or hover any variant to view its benchmark AUCs on WildRF, SynthBuster, and OpenFake2026.

Variants Leaderboard & Ablation Chart Threshold marker on each bar
Section 06

Runs On

The live hardware and software behind every score on this page.