Add ONNX models for visual similarity and perceptual comparison

README.md

@@ -0,0 +1,56 @@
+---
+license: mit
+tags:
+  - onnx
+  - clip
+  - lpips
+  - image-similarity
+  - computer-vision
+---
+
+# ONNX Models for vidupe.net
+
+This repository contains ONNX-exported models used by [vidupe.net](https://vidupe.net) for visual similarity and perceptual comparison tasks.
+
+## Models
+
+### `vidupe.net/models/clip_visual_vit_b32.onnx`
+
+CLIP visual encoder (ViT-B/32) exported to ONNX. This model encodes images into a 512-dimensional embedding space, enabling semantic image similarity comparisons.
+
+- **Source:** [openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32)
+- **Input:** RGB image tensor `[batch, 3, 224, 224]`, normalized
+- **Output:** Image embeddings `[batch, 512]`
+
+### `vidupe.net/models/lpips_alexnet.onnx`
+
+LPIPS (Learned Perceptual Image Patch Similarity) model with an AlexNet backbone exported to ONNX. Computes perceptual distance between two image patches.
+
+- **Source:** [richzhang/PerceptualSimilarity](https://github.com/richzhang/PerceptualSimilarity)
+- **Input:** Two normalized RGB image tensors `[batch, 3, H, W]`
+- **Output:** Perceptual distance score `[batch, 1, 1, 1]`
+
+## Usage
+
+```python
+import onnxruntime as ort
+import numpy as np
+
+# CLIP visual encoder
+session = ort.InferenceSession("vidupe.net/models/clip_visual_vit_b32.onnx")
+image = np.random.randn(1, 3, 224, 224).astype(np.float32)
+embeddings = session.run(None, {"input": image})[0]
+
+# LPIPS perceptual similarity
+session = ort.InferenceSession("vidupe.net/models/lpips_alexnet.onnx")
+img0 = np.random.randn(1, 3, 64, 64).astype(np.float32)
+img1 = np.random.randn(1, 3, 64, 64).astype(np.float32)
+distance = session.run(None, {"input0": img0, "input1": img1})[0]
+```
+
+## Requirements
+
+```
+onnxruntime>=1.16.0
+numpy
+```