joynix28/clip-vit-large-patch14-pooler

CLIP ViT-L/14 vision tower — pooled 1024-d output (ONNX)

A browser-ready ONNX export of the CLIP ViT-L/14 vision tower that exposes the 1024-d pooled CLS output as pooler_output. Every other browser-friendly export of this encoder stops at image_embeds, the 768-d projected vector, which is not the feature an AI-image probe is trained on.

This repository exists so that the linear probe published by Rinne414/SynthID-detector (MIT) can run client-side. That probe — ai_detector/model.npz, a logistic regression over these very features — is validated held out by its author (AUC 0.998; 0.0 % missed AI; 1.8 % false positives; 0.97–1.00 across generators it never saw).

What it is

• input pixel_values — float32[batch, 3, 224, 224], CLIP normalisation (mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711]), bicubic resize to 224×224.
• outputs
  • pooler_output — float32[batch, 1024], the post-layernorm pooled CLS feature
  • image_embeds — float32[batch, 768], the original projected embedding

How it was made

Derived from Xenova/clip-vit-large-patch14 (onnx/vision_model_quantized.onnx) by adding a single Identity node that exposes the tensor feeding the visual projection as a graph output. Graph surgery only — no retraining, no weight modification. Weights are OpenAI's original CLIP ViT-L/14.

Licence

MIT, following openai/clip-vit-large-patch14 and the MIT-licensed tooling it was derived from.