🖼️ Image Restoration¶

🧪 ICML2026 · 2 paper notes

Hierarchical Image Tokenization for Multi-Scale Image Super Resolution: H-VAR re-slices the "residual quantization for multi-scale generation" VAR paradigm into hierarchical image tokenization (HIT), enabling a 310M small model to output three meaningful intermediate resolutions (128 / 256 / 512) in a single forward pass. A DPO regularization term, which does not require an external reward model, is added to bias outputs toward HR. On standard ISR datasets, it competes with the 1B-parameter VARSR.
Image Restoration via Diffusion Models with Dynamic Resolution: SubDAPS / SubDAPS++ adapts pixel-space diffusion restoration methods like DPS and DAPS into a "dynamic resolution diffusion model" framework—sampling in \(64^2 / 128^2\) subspaces in early stages and returning to \(256^2\) full resolution later. It replaces Langevin with conjugate gradient, switches between stochastic/deterministic sampling via thresholding, and adds a corrector step that requires no extra network evaluation. On four linear and two nonlinear restoration tasks, it outperforms both pixel and latent diffusion methods on most metrics while being faster at inference.