Beta
AI-Generated Image Detection
Cross-generator deepfake detection, no retraining when new models drop
Detect AI-generated images across modern generation families — diffusion, transformer, and emerging architectures. The same frozen Bhala encoder that powers our text moderation, extended to images.
Features
What you get
Trained across 8 generator families (Stable Diffusion, DALL-E, Midjourney, Flux, SDXL, SD2.1, DeepFloyd, mixed deepfake pools)
9 real-image cohorts for balanced calibration (photos, art, faces, scenes, food, flora, captioned web)
Generalizes to unseen generators without retraining — same encoder, swappable head
8 independent forensic decision methods per image (geometric, learned linear, k-NN constructor, set operations)
Returns calibrated probability + per-method scores + verdict (fake / real / uncertain)
Same auth + billing + endpoint as Bhala text moderation
Pricing: $0.02 per image (Beta access)
Request
curl -X POST https://api.bhala.ai/v1/score-image \
-H "Authorization: Bearer bh_sk_..." \
-H "Content-Type: application/json" \
-d '{
"images_b64": ["iVBORw0KGgoAAAANS..."]
}'Response
{
"items": [
{
"is_fake": true,
"fake_probability": 0.847,
"verdict": "fake",
"probes": {
"ce_classifier": 0.823,
"linear": 0.851,
"whitened_linear": 0.842,
"centroid": 0.811,
"vector_algebra": 0.806,
"invert_a": 0.819,
"constructor_5nn": 0.872,
"zfc_intersection": 0.808
}
}
],
"latency_ms": 142,
"model_version": "bhala-image-mod-v1"
}Use Cases
What you can build with AI-Generated Image Detection
Trust & safety: flag AI-generated political imagery before it spreads
News verification: triage submissions in seconds, not hours
Identity / KYC: catch GenAI face-swap attempts in onboarding
Content platforms: comply with EU DSA / AI Act synthetic-content labelling
Brand protection: detect AI counterfeit product imagery