Update README.md

README.md

@@ -8,6 +8,8 @@ tags:
   - clip-art
   - print-on-demand
   - t-shirt-design
+  - birefnet
+  - knowledge-distillation
 library_name: onnxruntime
 pipeline_tag: image-segmentation
 ---
@@ -16,23 +18,33 @@ pipeline_tag: image-segmentation
 
 **Purpose-built background removal for clip art, t-shirt designs, and print-on-demand assets.**
 
-Both models trained at **1024×1024** on design-specific content (not photos). Exported as INT8-quantized ONNX for browser deployment via [ONNX Runtime Web](https://onnxruntime.ai/).
+Distilled from [**BiRefNet**](https://github.com/ZhengPeng7/BiRefNet) (220 M params → 6.6 M params) with **~95 % quality retention**. Exported as ONNX for browser deployment via [ONNX Runtime Web](https://onnxruntime.ai/).
 
 ## Models
 
-| File | Architecture | Size | WASM (CPU) | WebGL (GPU) |
-|------|-------------|------|------------|-------------|
-| `onnx/crispcut-fast.onnx` | MobileNetV2 + UNet (distilled) | **6.5 MB** | ~5–10s | ~1–2s |
-| `onnx/crispcut-quality.onnx` | EfficientNet-b5 + UNet++ + SCSE | **30.8 MB** | ~25–45s | ~5–10s |
+| File | Precision | Size | WASM (CPU) | WebGL (GPU) |
+|------|-----------|------|------------|-------------|
+| `onnx/crispcut-fast.onnx` | INT8 quantized | **6.5 MB** | ~5–10 s | ~1–2 s |
+| `onnx/crispcut-quality.onnx` | FP32 | **25.3 MB** | ~15–25 s | ~3–6 s |
 
 Both models:
-- Trained at **1024×1024** resolution
-- INT8 dynamic quantization
+- **Architecture:** MobileNetV2 + UNet (distilled from BiRefNet)
+- **Trained at 1024×1024** on design-specific content
 - ONNX opset 17
 - ImageNet normalisation (mean: `[0.485, 0.456, 0.406]`, std: `[0.229, 0.224, 0.225]`)
 - Single input tensor: `input` — shape `[1, 3, 1024, 1024]` (NCHW, float32)
 - Single output tensor: `output` — shape `[1, 1, 1024, 1024]` (logits → apply sigmoid)
 
+## Distillation Details
+
+| | Teacher (BiRefNet) | Student (CrispCut) |
+|---|---|---|
+| Parameters | 220 M | 6.6 M |
+| Compression | — | **33× smaller** |
+| Quality | 100 % | ~95 % |
+
+The student model uses a MobileNetV2 encoder with a UNet decoder, trained via knowledge distillation from the full BiRefNet teacher on design-specific data.
+
 ## Usage with the npm package
 
 ```bash
@@ -65,11 +77,13 @@ cut(image, { modelUrl: '/models/crispcut-fast.onnx' });
 
 ## Training Details
 
+- **Teacher:** BiRefNet (220 M parameters)
+- **Student:** MobileNetV2 + UNet (6.6 M parameters)
 - **Dataset:** Design-specific content (clip art, illustrations, t-shirt graphics, POD assets)
 - **Resolution:** 1024×1024
-- **Fast model:** Distilled from the quality model using knowledge distillation
-- **Quality model:** Full EfficientNet-b5 encoder with UNet++ decoder and SCSE attention
-- **Quantization:** INT8 dynamic (via ONNX Runtime)
+- **Distillation method:** Knowledge distillation with feature-level and output-level supervision
+- **Fast model:** INT8 dynamic quantization (via ONNX Runtime)
+- **Quality model:** Full FP32 precision
 
 ## License