Downloads · 30 days
0
miraklodiwska1/siglip-ai-vs-human-onnx
siglip-ai-vs-human-onnx is a image classification model from miraklodiwska1. Use it when you need a label for an image. It is set up for onnx. The card lists the license as apache-2.0.
A straight ONNX export of Ateeqq/ai-vs-human-image-detector. No fine-tuning, no architecture change, no recalibration — the weights are the upstream ones, in a format browsers can run.
Downloads · 30 days
0
Access
Public
Updated Aug 21, 2026
Repo size
343 MB
Likes
0
Public
Click a slice to open those files.
.onnx343 MB · 100%
From the Hugging Face model README
A straight ONNX export of Ateeqq/ai-vs-human-image-detector.
No fine-tuning, no architecture change, no recalibration — the weights are the
upstream ones, in a format browsers can run.
Published because upstream ships safetensors only, and Forensic Lens needs to load this model with ONNX Runtime Web inside a Chrome extension.
Dynamic INT8 quantization of this model costs 2.7 points of balanced accuracy (78.6% → 75.9% on a 3,559-image corpus) and makes it brittle across runtimes: ONNX Runtime's Python CPU INT8 kernels and ORT-web's WASM INT8 kernels disagree by up to 16.6 in logit space on the same input. FP32 is 343 MB instead of 87 MB and worth it. Details: docs/EVALUATION.md.
pixel_values, float32 [batch, 3, 224, 224], RGB, normalized
with mean [0.5, 0.5, 0.5] and std [0.5, 0.5, 0.5].logits, float32 [batch, 2]. id2label = {0: "ai", 1: "hum"},
so the log-odds of "AI-generated" is logits[0] - logits[1].import numpy as np, onnxruntime as ort
from PIL import Image
sess = ort.InferenceSession("model.onnx")
im = Image.open("image.jpg").convert("RGB").resize((224, 224), Image.BILINEAR)
x = ((np.asarray(im, np.float32) / 255.0 - 0.5) / 0.5).transpose(2, 0, 1)[None]
logits = sess.run(None, {"pixel_values": x})[0][0]
p_ai = 1 / (1 + np.exp(-(logits[0] - logits[1])))
Scores are uncalibrated model outputs, not probabilities. Forensic Lens applies its own calibration on top.
pip install torch transformers onnx onnxruntime
python3 scripts/convert-siglip.py # from the Forensic Lens repo
The export is deterministic; re-running reproduces the identical digest.
sha256: b0b82e15a57d1cb72a2920098fd39634c7013ff05a66ce7673c38bea12d3363b
Apache-2.0, inherited from the upstream model. Original work by Ateeqq; this repository only changes the serialization format.