Janspiry/Image-Super-Resolution-via-Iterative-Refinement

Implements a diffusion-based architecture using DDPM-style ResNet blocks with attention mechanisms applied to low-resolution features (16×16), encoding timestep information via FiLM conditioning rather than affine transformation. Supports both conditional super-resolution (16×16→128×128, 64×64→512×512) and unconditional face generation, with multi-GPU training, Weights & Biases logging, and checkpoint resumption built in. Includes pretrained models for FFHQ/CelebA-HQ datasets and data preparation utilities for LMDB or PNG formats.

3,910 stars. No commits in the last 6 months.