Speaker
Description
Recent JPEG restoration systems achieve strong visual quality with large models, but their computational cost often limits efficient on-device deployment. We present a 65M-parameter generative restorer that improves LPIPS over three regression baselines in seven of nine held-out conditions on LIVE-1, Urban100, and DIV2K validation, while sustaining 8.05 images/s at 720p on a single RTX 3090. Trained from scratch, the model replaces learned VAE encoding and decoding with an exactly invertible two-level Haar transform. A rank-enhanced linear-attention DiT predicts clean wavelet-domain residuals and internally estimates compression severity. A MeanFlow objective enables inference with one or two network evaluations without distillation. Although large pretrained priors remain stronger on several quality metrics, our approach offers a compact, high-throughput alternative for resource-constrained JPEG restoration.