OracleZoom: On-Policy Self-Distillation Inspired Reference-Constrained Recursive Image Super Resolution
Summary
OracleZoom improves recursive image super-resolution by combining on-policy self-distillation with reference constraints and cross-scale supervision to reduce hallucinations, achieving state-of-the-art results across multiple datasets.
View Cached Full Text
Cached at: 09/10/26, 02:17 PM
Paper page - OracleZoom: On-Policy Self-Distillation Inspired Reference-Constrained Recursive Image Super Resolution
Source: https://huggingface.co/papers/2609.06490
Abstract
OracleZoom improves recursive super-resolution by combining trajectory-based training with cross-scale supervision and a latent prior to reduce hallucinations at extreme magnifications.
Recursive Super-Resolution(SR) extends fixed-scale SR to extreme magnification by repeatedly feeding predictions back into the same model, analogous to zooming an image repeatedly. However, ground truth availability at every scale, especially at depth, remains challenging as the required source resolution grows geometrically, leaving deeper predictions unsupervised. We present OracleZoom, anon-policy distillation-inspired, reference-constrained framework that trains on its trajectory while carrying the last ground-truth evidence beyond the supervision boundary. Direct andcross-scale supervisionconstrain verifiable content, while ano-reference quality objectiveguides unresolved fine-scale detail. AKL-constrained pretrained latent priorlimits quality-driven drift, whileEMA consistencystabilizes the supervision boundary. Across seven datasets, OracleZoom achieves the state-of-the-art SR quality across zooming scales, averaging 0.713 CLIPIQA, with larger gains on deeper scales, while significantly reducing hallucinations. Code, data, and models are available at https://dipta007.github.io/OracleZoom/ .
View arXiv pageView PDFProject pageGitHub0Add to collection
Get this paper in your agent:
hf papers read 2609\.06490
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper1
#### dipta007/OracleZoom Image-to-Image• Updatedabout 3 hours ago
Datasets citing this paper1
#### dipta007/OracleZoom-4KLSDB-train Viewer• Updatedabout 4 hours ago • 52.8k • 399
Spaces citing this paper2
Collections including this paper2
Similar Articles
Super-Resolution in The Right Latent Space: A Frozen Vision-Foundation Substrate
提出RAESR，利用冻结DINOv3-L的23层融合层作为超分辨率的潜空间基底，训练415M解码器将退化嵌入映射回流形，在RealSR、DRealSR、LSDIR和DIV2K-Val上实现最优的保真度-感知权衡，且单张H20 GPU上每512×512图像仅需37ms。
Recursive Self-Improvement via On-Policy Distillation for Reasoning
This paper proposes a recursive framework using Dynamic Co-Evolution and Self-Refined Concise Learning to improve on-policy self-distillation for reasoning, demonstrating significant gains on the Qwen3-8B model across mathematics benchmarks.
RefGC-SR^2: Reference-guided Generated Content Super-Resolution and Refinement
This paper introduces a new task, reference-guided generated content super-resolution-refinement (RefGC-SR²), which simultaneously recovers high-resolution details and refines generative artifacts using a frequency-aware diffusion transformer model. The method leverages a high-resolution reference image to improve the quality of AI-generated images during post-processing.
Improving OCR Faithfulness via Gated and Attenuated On-Policy Distillation
The paper introduces GAD-RL, a method that adaptively gates and attenuates on-policy distillation supervision during joint post-training of vision-language models to improve OCR transcription faithfulness, achieving 59.92% Micro Recall on CHAOS-Bench on Qwen3.5-2B, surpassing GRPO by 8.45 points.
Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes
Where-OPD proposes spatially guided on-policy self-distillation for MLLMs, where a teacher with privileged spatial text guidance is distilled into a student using procedurally generated synthetic scenes, yielding a 3.23-point average gain on real-world perception benchmarks via synthetic-to-real transfer.