Pick an image, watch it sharpen
Click any image to run it through the live pipeline. It's queued as a job, super-resolved by the model, and shown below with a comparison slider.
Or bring your own
PNG or JPEG up to 25 MB. Large images are downscaled to ~480p first, since the model's job is upscaling low-res inputs.
How it works
CUDA PyTorch + PixelShuffle / AMP
A residual CNN with PixelShuffle upsampling and mixed-precision (AMP) training/inference, cutting VRAM by 40% on NVIDIA T4 GPUs while keeping tiled, seam-free upscaling.
ONNX / TensorRT + Rust API
The model is exported to ONNX/TensorRT and served through a Rust API backed by PostgreSQL, with jobs queued, claimed, and tracked end to end.
PSNR / SSIM vs bicubic
Benchmarked against a bicubic baseline on held-out images, delivering +3.1 dB PSNR and +0.05 SSIM over the classic upsampler.