OracleZoom Combines On-Policy Self-Distillation and Reference Constraints to Cut Hallucinations in Extreme Super-Resolution

Shubhashis Roy Dipta · hf · 2026-09-10

OracleZoom, released on Hugging Face, tackles recursive image super-resolution at extreme magnifications. The method combines trajectory-based training with cross-scale supervision and a latent prior, constrained by reference signals, to reduce hallucinations that typically appear at high zoom factors. Inspired by on-policy self-distillation, it aims for more faithful high-resolution outputs.

Original post →

More from Multimodal

Multimodal channel →