NVIDIA's PiD upscaler pushes from 4K to 64MP using novel tiled diffusion
64-megapixel outputs from 1K inputs? NVIDIA’s PiD tiled upscaler delivers.
Deep Dive
From 1K–4K to 64MP output samples: a model pushes beyond previous resolution limits. The full story explains how it was built, what was discovered, and what the model can and cannot do.
Key Points
- PiD (Pixel-wise Diffusion) is a tiled upscaler that achieves 64MP output from 1K–4K inputs — a 16x resolution boost over state-of-the-art methods.
- Uses overlapping tile denoising with a diffusion backbone to bypass GPU memory limits, enabling parallel inference.
- Outperforms Real-ESRGAN and SwinIR on detail recovery but shows minor structural artifacts at tile boundaries.
Why It Matters
64MP AI upscaling from low-res inputs – a game-changer for film, medical, and satellite imaging industries.