OracleZoom: On-Policy Self-Distillation Inspired Reference-Constrained Recursive Image Super Resolution
OracleZoom trains recursive super-resolution past the point where ground truth is available by carrying forward the last verified reference.
The paper targets extreme zooming, where each generated image is fed back into the same model and deeper scales become hard to supervise. Its framework combines direct and cross-scale supervision for content that can be checked, then uses no-reference quality scoring for finer detail. A pretrained latent prior and EMA consistency are used to limit drift and stabilize training. The authors report state-of-the-art results across seven datasets, averaging 0.713 CLIPIQA, with larger gains at deeper zoom levels and fewer hallucinations. HF Daily Papers' note
The paper targets extreme zooming, where each generated image is fed back into the same model and deeper scales become hard to supervise. Its framework combines direct and cross-scale supervision for content that can be checked, then uses no-reference quality scoring for finer detail. A pretrained latent prior and EMA consistency are used to limit drift and stabilize training. The authors report state-of-the-art results across seven datasets, averaging 0.713 CLIPIQA, with larger gains at deeper zoom levels and fewer hallucinations. HF Daily Papers' note
score 4