# HEART

    Lightweight by transformer standards, and built for real-world use. Drop an
    image, pick a model, press **Upscale**, and download the full-quality PNG.

    **What is HEART?** A ~16.7M-parameter attention network that matches HAT-M
    on quality while using **~half the training VRAM** and running **~1.8x
    faster** (3.6x less VRAM) at inference. One file, standard ONNX, no
    fused/unfused pairs.

    - **release** models → **clean** images.
    - the **OTF GAN** model → **degraded** images (blurry, noisy, JPEG-compressed).
    Large images are processed in 32px-aligned tiles.