Professional System Specification: Robust AI-Generated Content (AIGC) Detection Model

Target Application: Enterprise-grade AI-Generated Content Detection & Robustness Engine

Core Objective: High-precision, zero-leakage classification of authentic vs. AI-generated visual media under severe real-world redistribution degradations (e.g., social media re-compression, aggressive rescaling, and additive sensor noise).

Technical Performance & Empirical Benchmark

Primary Model Classification Metrics

Evaluated across internal validation (val_internal) and out-of-distribution transfer test splits (test_transfer). The deployed dual-tower ViT-L system delivers high operational accuracy and low False Positive Rates (FPR).

Split	AUC	TPR@1%FPR	Dataset size (Real / Fake)	Model Families & Counts
train	—	—	350,564 (175,670 / 174,894)	SD (5), Diffusion/Pixel (5), GAN (6), PixArt (3), SD-Adaptor (3), SD-Img2Img (3), Kandinsky (3), OmniGen (2), Infinity (2), FLUX (1), Z-Image (1), Open-Weight Singletons (5)
val	0.9654	0.7348	53,600 (26,800 / 26,800)	VQ/Autoencoder (4), SD (2), FLUX (2), Klein (2), Open-Weight Singletons (3), Commercial Closed-Weight (2)
test	0.9963	0.9825	37,295 (18,647 / 18,648)	Open-Weight Singletons (4), SD (1), FLUX (1), Z-Image (1), Commercial Closed-Weight (11)
demo	0.9989	0.9758	—	Coco val 2017 (1), Dall E Advanced

Robustness & Stress-Test Degradation Matrix

Evaluated against a baseline reference performance ($\text{AUC} = 0.9551$). Operational invariance is sustained across typical degradation pipelines, with catastrophic degradation observed only at extreme compression thresholds ($\text{JPEG } q=10$).

Applied Augmentation Pipeline	Metric (AUC)	Differential Score (Δ vs Clean Baseline)	Operational Impact
Clean Baseline Reference	0.9551	—	Baseline Nominal
JPEG Compression ($q = 90$)	0.9645	$+0.0094$	High Stability
JPEG Compression ($q = 70$)	0.9645	$+0.0094$	High Stability (Unseen Severity)
JPEG Compression ($q = 50$)	0.9550	$-0.0001$	Negligible Shift
JPEG Compression ($q = 10$)	0.8949	$-0.0601$	High Degradation
Gaussian Blur ($\sigma = 0.5$)	0.9673	$+0.0122$	Performance Gain
Gaussian Blur ($\sigma = 1.0$)	0.9585	$+0.0035$	High Stability (Unseen Severity)
Gaussian Blur ($\sigma = 2.0$)	0.9500	$-0.0051$	Nominal Shift
Spatial Downscale ($0.5\times$)	0.9600	$+0.0049$	High Stability
Spatial Downscale ($0.25\times$)	0.9353	$-0.0198$	Moderate Shift
Additive Noise ($\sigma = 0.02$)	0.9584	$+0.0033$	Nominal Shift
Additive Noise ($\sigma = 0.05$)	0.9565	$+0.0015$	Nominal Shift
Additive Noise ($\sigma = 0.10$)	0.9411	$-0.0140$	Moderate Shift
Color Jitter (Brightness $\pm 20\%$)	0.9488	$-0.0062$	Nominal Shift
Color Jitter (Contrast $\pm 20\%$)	0.9617	$+0.0066$	High Stability
Color Jitter (Saturation $\pm 20\%$)	0.9572	$+0.0021$	Nominal Shift
Spatial Center Crop ($80\%$)	0.9595	$+0.0045$	High Stability

Corpus Curation & Out-of-Distribution Splitting

The training corpus leverages external benchmark distributions combined with targeted internal synthetic image generations. It strictly isolates zero-day commercial architectures and unseen autoencoder lineages to prevent data leakage.

External Benchmarks: Baseline assets derived from NTIRE 2026 and WildFake datasets, capturing legacy GANs, foundational diffusion pipelines, and baseline organizer distributions.
In-House Targeted Generations: Synthetic visual suites anchored to real portrait photographs from Open Images V7. Designed specifically to capture modern mobile visual trends (e.g., social media portrait layouts), spanning both Text-to-Image (T2I) and Image-to-Image (I2I) generation paradigms.
Aspect Ratio Balancing: Open Images V7 portrait-oriented photographs ($18.0\%$ of sampled thumbnails $<0.8$ aspect ratio) are integrated to decouple aspect ratio structural shortcuts from the authentic class label.
OOD Generalization Strategy: Held-out validation and testing splits strictly isolate unseen open-weight architectures (e.g., FLUX variants, SANA) and closed-weight commercial APIs (e.g., GPT Image 2, Seedream 4.5, Ideogram 4.0). Splits are enforced at the decoder/autoencoder lineage level rather than model checkpoint names (e.g., grouping all flux2_vae lineages simultaneously).

                       Total Evaluated Corpus (N = 455,300)
                                        │
         ┌──────────────────────────────┼──────────────────────────────┐
         ▼                              ▼                              ▼
  Train Set (77.0%)             Val Set (11.8%)                Test Set (8.2%)
 Real: 175,670 | Fake: 174,894   Real: 26,800 | Fake: 26,800     Real: 18,647 | Fake: 18,648
 Total: 350,564                 Total: 53,600                  Total: 37,295

Quantitative Split Breakdown

Dataset Partition	Authentic (Real)	Synthetic (Fake)	Aggregate Total	Generative Families & Architectural Diversity
Train	175,670	174,894	350,564	SD (5), Diffusion/Pixel (5), GAN (6), PixArt (3), SD-Adaptor (3), SD-Img2Img (3), Kandinsky (3), OmniGen (2), Infinity (2), FLUX (1), Z-Image (1), Singletons (5)
Validation	26,800	26,800	53,600	VQ/Autoencoder (4), SD (2), FLUX (2), Klein (2), Open-Weight Singletons (3), Commercial Closed-Weight (2)
Test	18,647	18,648	37,295	Commercial Closed-Weight (11), Open-Weight Singletons (4), SD (1), FLUX (1), Z-Image (1)
Total	226,115	229,185	455,300	Full Architectural Cross-Section

Architectural Representation by Lineage

Generator Lineage Family	Train Partition	Validation Partition (Internal)	Test Partition (Out-of-Distribution)
Stable Diffusion (SD)	SD 1.4, SD 1.5 T2I, SD 2.1, SDXL 1.0, SDXL T2I	SDXL Turbo, SD 3 Medium	SD 3.5 Large
SD-Adaptor	SD Adaptor ControlNet, SD Adaptor LoRA, SD Adaptor LyCORIS	—	—
SD-Img2Img	SD 1.5 Img2Img, SDXL Self-Cond, SDXL Img2Img	—	—
FLUX	FLUX.1 Kontext Dev	FLUX.1 Kontext Dev, FLUX.1 Dev	FLUX.1 Schnell
Generative Adversarial Nets	DF-GAN, GALIP, StarGAN, StyleGAN, GigaGAN, BigGAN	—	—
Diffusion / Pixel UNet	ADM, DDIM, DDPM, VQDM, Imagen	—	—
PixArt	PixArt-512, PixArt-$\alpha$, PixArt-$\Sigma$	—	—
Kandinsky	Kandinsky 2, Kandinsky 3, Kandinsky 2.2 T2I	—	—
OmniGen	OmniGen, OmniGen 2	—	—
Infinity	Infinity 2B, Infinity 8B	—	—
Klein	—	Klein 4B T2I, Klein 4B Ref Image	—
VQ / Autoencoder	—	VQ-VAE, VQ-GAN, MAGE, MAE	—
Z-Image	Z-Image T2I	—	Z-Image Turbo
Open-Weight Singletons	YOSO, Kolors, Janus Pro 7B, Ovis Image, DeepFloyd IF	Playground v2.5, Lumina Image 2.0, Qwen Image	Würstchen T2I, Sana 1600M T2I, HiDream, Ideogram 4.0 Turbo
Commercial Closed-Weight	—	Ideogram v3 Turbo†, ImageGen-4 Fast†	Nano Banana Pro†, FLUX-2 Max†, ImageGen-4 Ultra†, Seedream 5 Lite†, Groq Imagine Image†, Google Nano Banana 2 Lite, ByteDance Seedream 4.5, OpenAI GPT Image 2, Meta Muse Image

Note: Models marked with (†) indicate preliminary proprietary evaluation suites.

Core System Architecture & Operational Pipeline

Dataset Ingest
     │
     ▼
Canonicalisation (200px Crop, strip short-side shortcuts)
     │
     ▼
Augmentation (11 Stochastic View Pipelines, 1-3 operations chained)
     │
     ▼
Dual Parallel Towers (DINO ViT-L @ 224px, 608M combined parameters)
     │
     ▼
Concatenation Head (2048-d Feature Vector -> 512-d MLP)
     │
     ▼
Loss Regularization (Class BCE + Aux Degradation + TeleAI KL/MSE Consistency)
     │
     ▼
Calibrated Probability Output & EQI Dynamic Temperature Scaling

1. Canonicalisation & Resolution Shortcut Stripping

Short-side pixel dimensions leak class labels due to source dataset encoding biases (e.g., COCO real photographs exhibiting short-side dimensions of 200px versus high-resolution synthetic renders). Canonicalisation enforces strict input standardization prior to feature extraction.

Window Dimensions: CANON_CROP_SIDE = 200 native pixel window, matching the bandwidth floor of authentic benchmark classes. Detail above 200px is omitted to prevent the network from learning high-frequency resolution artifacts as binary class indicators.
Crop vs. Band Resampling: Native cropping preserves high-frequency generator artifacts and pixel-grid structures. It provides measured gains for DINO-lineage ViTs (+0.079 on dinov2l) without introducing resampling grid interference.

2. Stochastic Multi-View Augmentation Pipeline

Training views ($N_{\text{views}} = 11$; 1 clean, 10 degraded) apply continuous, non-linear augmentation chains.

Compound Recipes: Each degraded view applies 1 to 3 chained operations drawn from 6 distinct transform families (JPEG compression, Gaussian blur, spatial resizing, additive noise, color jitter, and center cropping).
Withheld Severity Bands: To evaluate generalization against unseen degradation severities, specific operational parameters (e.g., JPEG quality $q \in [65, 75]$ and Gaussian blur $\sigma \in [0.85, 1.15]$) are strictly withheld during training.

3. Dual-Tower Vision Backbone

The primary extraction pipeline comprises two parallel 300M-parameter DINO-lineage Vision Transformer (ViT-L/14) towers operating at 224px native input resolution.

Parameter Topology: Complete unfreezing across all 24 Transformer blocks (--depth 24). Master weights are maintained in float32 precision with bfloat16 mixed-precision execution.
Optimization Differential: Tower learning rates are decoupled from downstream head parameters ($\text{LR}{\text{tower}} = 1\times 10^{-5}$ vs. $\text{LR}{\text{head}} = 1\times 10^{-3}$) to preserve foundational spatial representations.

4. Classification Head & Multi-Task Consistency Loss

Embeddings from both backbones are concatenated into a 2048-dimensional joint vector and fed into a 1M-parameter MLP head (Linear → GELU → LayerNorm(512) → Linear → GELU → Linear(1)). Training is regularized via a tri-part multi-task loss structure:

$$\mathcal{L}{\text{total}} = \mathcal{L}{\text{BCE}} + 0.3 \cdot \mathcal{L}{\text{degradation}} + 1.0 \cdot \text{KL}\left(P{\text{clean}} \parallel P_{\text{degraded}}\right) + 1.0 \cdot \text{MSE}\left(h_{\text{clean}}, h_{\text{degraded}}\right)$$

Asymmetric Reference Target: The clean feature branch is detached during backpropagation, establishing a fixed reference state to pull degraded feature representations toward without risking representation collapse.
Auxiliary Degradation Head: Predicts degradation presence and severity parameters to output a continuous degradation embedding vector ($d$).

5. Calibration & Abstention Policy

Inference delivers calibrated probability scores $P(\text{AI-generated}) \in [0, 1]$ mapped to a strict target operational threshold ($\text{FPR} = 0.01$).

Dynamic Temperature Scaling: Incorporates predicted degradation vectors ($d$) and handcrafted proxy metrics ($h$) to dynamically adjust logit scaling:

$$T(d, h) = \text{softplus}\left(\text{Linear}([d; h])\right)$$

Under severe input corruption, confidence scores automatically soften toward $0.5$.

Evidence Quality Index (EQI): Estimates residual forensic evidence within input images to support a three-way decision output: Clear (Authentic), Flag (Synthetic), or Review (High Uncertainty / Excessively Degraded).

Key Systemic Takeaways & Architectural Decisions

System Metric / Component	Selection Rationale	Empirical Justification & Evidence
Lineage-Based Data Splitting	Grouped by generator family and VAE decoder lineage rather than model name.	Prevents model memorization of shared latent spaces; holding out `klein4b_t2i` alongside `klein4b_ref_image` measures actual structural generalizability (`splits.py`).
200px Native Spatial Crop	Standardizes input resolution prior to transform execution.	Resolves resolution shortcuts where short-side dimensions yielded standard synthetic classification AUC of 1.0000 on baseline distributions (`docs/resolution_shortcut.md`).
Multi-Task Loss Regularization	Integrates KL prediction divergence and MSE hidden-state consistency.	Reduces calibration drift from 1.24 pts down to 0.37 pts (a 3.4× reduction), maintaining threshold stability across degraded deployment environments (`docs/robustness_table.md`).
Dynamic Temperature Scaling ($T$)	Rescales raw classification logits using auxiliary degradation vectors ($d$).	Mitigates overconfident false alarms on corrupted real images by softening probability outputs toward maximum uncertainty under high noise levels.
Dual ViT-L Parallel Architecture	Concatenates two 300M DINO ViT-L towers at 224px input resolution.	Provides high out-of-distribution transfer metrics (Transfer Test $\text{AUC} = 0.9961$; $\text{TPR @ 1\% FPR} = 0.9823$). Note: Evaluation warrants fine-tuned comparison against single-tower multi-policy baselines.