---
title: UNIStainNet - Virtual IHC Staining
emoji: 🔬
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: "6.10.0"
app_file: app.py
pinned: false
license: mit
hardware: cpu-basic
---

# UNIStainNet: Foundation-Model-Guided Virtual Staining

Virtual staining of H&E histopathology images to IHC (HER2, Ki67, ER, PR) using a single unified 42M-parameter SPADE-UNet conditioned on dense spatial tokens from a frozen UNI pathology foundation model.

## Features

- **Gallery** — Browse 16 pre-computed examples from BCI and MIST datasets (no GPU needed)
- **Virtual Staining** — Upload an H&E image and generate any IHC stain (GPU required)
- **Cross-Stain Comparison** — Generate all 4 stains from a single H&E input (GPU required)

## Architecture

| Component | Details |
|-----------|---------|
| Generator | SPADE-UNet with UNI spatial conditioning + FiLM stain embeddings |
| UNI Features | 4x4 sub-crop tiling → UNI ViT-L/16 → 32x32 spatial tokens (1024-dim) |
| Parameters | 42M (generator), UNI frozen (303M) |

## GPU Support

- **Gallery tab** works on CPU (default hardware)
- **Live inference** requires GPU — set Space hardware to ZeroGPU (HF Pro) or run locally with `python app.py`