PatrickKillinni/clearsnap-detector

ClearSnap offline AI-image detector (int8 ONNX)

On-device build of haywoodsloan/ai-image-detector-deploy (SwinV2, Apache-2.0) used by the ClearSnap app. Exported to ONNX (opset 17) and dynamically quantised to int8 (790 MB → 190 MB); output parity with the original PyTorch model was verified.

• Input pixel_values: float32 [1, 3, 256, 256], RGB, PIL bicubic resize to 256×256, scaled to 0–1, normalised with ImageNet mean/std (see preprocessing.json).
• Output logits: [1, 2] → index 0 = artificial, 1 = real.

Benchmark (ClearSnap test set, 393 images)

216 real photos (27 phone/camera models) + 177 AI images (20 generators, 2024–2026), Wikimedia Commons originals.

	Accuracy	AI caught	Real wrongly flagged
Original files (model + metadata provenance)	92%	88%	5%
Re-shared copies (1600 px, JPEG q85, metadata stripped)	88.5%	77%	2.3%

License

Apache-2.0, inherited from the base model by haywoodsloan. This is a converted, quantised derivative; no retraining was performed.

---

v2 (ClearSnap 7.4): Fast + Thorough detectors

File	Base model	License	Size	Output
clearsnap_detector_v2_int8.onnx	haywoodsloan/ai-image-detector-deploy (SwinV2)	Apache-2.0	190 MB	logits [1,2], features [1,1536]
clearsnap_siglip_w8.onnx	Ateeqq/ai-vs-human-image-detector (SigLIP)	Apache-2.0	85 MB	logits [1,2], features [1,768] (mean-pooled)
clearsnap_clip_int8.onnx	openai/clip-vit-base-patch16 (vision tower)	MIT	83 MB	features [1,512] (L2-normalised image embedding)

• SwinV2 / CLIP: dynamic int8. SigLIP: weight-only int8 (activation outliers break dynamic int8).
• head_hay.json (Fast) and head_hay_ate_clip.json (Thorough): logistic-regression decision layers trained by ClearSnap on frozen features of 2,056 images (Nano Banana, GPT-Image, DALL-E 3, FLUX, Midjourney/SD mix, Wikimedia Commons generators; real photos from 22 phone/camera models). CLIP features are scaled ×10.
• Preprocessing: SwinV2 — PIL bicubic 256×256, ImageNet mean/std. SigLIP — PIL bilinear 224×224, mean/std 0.5. CLIP — PIL bicubic shortest-edge 224 + centre crop 224, CLIP mean/std.

Held-out benchmark (393 Wikimedia Commons images, never used in training), compressed models, full app logic, accuracy when a verdict is given: Thorough 98.4% (originals) / 96.9% (re-shared); Fast 96.2% / 95.7%.