OracleZoom: On-Policy Self-Distillation Inspired Reference-Constrained Recursive Image Super Resolution

Shubhashis Roy Dipta; Sourajit Saha; Shaswati Saha; Nobin Sarwar

Source-linked AI summary

OracleZoom: On-Policy Self-Distillation Inspired Reference-Constrained Recursive Image Super Resolution

Shubhashis Roy Dipta, Sourajit Saha, Shaswati Saha, Nobin Sarwar

arXiv:2609.06490v1Sep 6, 2026cs.CVcs.AIcs.CLcs.LG

TL;DR

Recursive SR lacks direct ground-truth verification at deeper magnification scales because required source resolution grows geometrically. OracleZoom trains on its own recursive trajectory while carrying the last verifiable evidence forward and constraining unresolved detail with quality guidance, a KL prior, and EMA consistency. Across seven datasets, it reports state-of-the-art quality with 0.713 mean CLIPIQA and reduced hallucinations at deeper scales.

Problem
Recursive SR lacks ground-truth targets at deeper scales because the required source resolution grows geometrically, leaving fine-scale predictions without direct visual verification.
Method
OracleZoom trains on recursive predictions, carries aligned last-target evidence into deeper scales, and combines quality guidance with KL prior regularization and EMA consistency.
Results
0.713 mean CLIPIQA across seven datasets is reported with larger gains at deeper scales and significantly reduced hallucinations.
Takeaways & Limitations
OracleZoom preserves observable ground-truth evidence while improving perceptual recursive SR quality beyond the last target-available scale.
Takeaways & Limitations
Evaluation beyond 4× cannot measure exact recovery, follows synthetic center-crop recursion, and relies on fixed quality and pretrained SR priors that may bias unsupported detail.

Abstract

from arXiv · show

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, an on-policy distillation-inspired, reference-constrained framework that trains on its trajectory while carrying the last ground-truth evidence beyond the supervision boundary. Direct and cross-scale supervision constrain verifiable content, while a no-reference quality objective guides unresolved fine-scale detail. A KL-constrained pretrained latent prior limits quality-driven drift, while EMA consistency stabilizes 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/ .

1. Introduction

Recursive SR enables extreme magnification by repeatedly applying an SR model, but geometrically growing source-resolution requirements leave deeper scales without direct ground-truth verification. OracleZoom addresses this gap by carrying verifiable evidence forward while constraining unresolved detail.

Motivation: Recursive SR repeatedly feeds each prediction into the same SR model to achieve extreme magnification.This process is analogous to repeatedly zooming an image.
Supervision Gap: 512×512 inputs with successive 4× magnifications require source regions growing from 2048^2 to 131072^2 pixels, making deep-scale supervision impractical.A single uncompressed 131072×131072 RGB image requires about 52 GB of storage.
Approach: At deeper scales, OracleZoom carries the last available ground-truth evidence beyond the supervision boundary to constrain recursive predictions.The framework trains on its own recursive trajectory, matching inference behavior.
Approach: OracleZoom separates deeper predictions into verifiable content constrained by cross-scale references and unresolved detail guided by a frozen no-reference quality model.A KL prior limits drift from the pretrained SR model, while EMA consistency stabilizes the supervision boundary.
Results: 0.713 mean CLIPIQA is achieved across seven datasets, with 0.706 CLIPIQA at 256× and lower hallucination rates at deeper scales.At 64× and 256×, an independent judge clearly prefers OracleZoom in 68% and 78% of comparisons, respectively.

2. Related Work

Prior work advances perceptual and arbitrary-scale super-resolution, but recursive SR still needs indirect supervision when deeper targets are unavailable. OracleZoom combines on-policy training, reference evidence, quality guidance, and distributional regularization to address that setting.

Super-Resolution: Modern SR methods use diffusion and large generative priors to synthesize perceptually stronger high-frequency detail.Related approaches also model unknown degradations and support progressive or arbitrary-scale reconstruction.
Indirect Supervision: When targets are unavailable, prior methods use perceptual objectives, quality estimators, teacher-student consistency, and generative priors for indirect supervision.EMA teachers provide slowly varying consistency targets in related work.
On-Policy Learning: OPSD trains students on their own trajectories using privileged information available to a teacher, reducing mismatch between training and deployment states.OracleZoom adapts this idea to recursive SR, where ground truth provides privileged visual information during training.
Distributional Regularization: No-reference quality objectives can reward plausible but unsupported detail, motivating quality guidance combined with a KL-constrained pretrained latent prior.The prior regularizes the adapted latent distribution toward the pretrained SR model.

3. Method

OracleZoom recursively applies a shared latent SR model while separating supervised scales from deeper target-unavailable scales. Direct supervision, projected ground-truth references, quality guidance, KL regularization, and EMA consistency jointly constrain the recursive trajectory.

Recursive Setup: A zoom operator selects the region for magnification at each step, while the latent SR model performs super-resolution on the resulting input.The recursive sequence uses cumulative magnification factors and relative zooms between successive scales.
Problem Setup: OracleZoom assumes ground-truth targets exist only through step m and seeks reliable behavior both within and beyond that supervised range.The supervision gap begins when recursion continues without direct visual evidence.
Reference-Constrained Supervision: Direct supervision anchors target-available scales, while cross-scale consistency carries verifiable ground-truth information into deeper recursion.Aligned regions are extracted from the last available target and deeper predictions are projected back to observable resolution.
Unresolved Detail: A frozen no-reference quality model guides unresolved fine-scale detail that projection cannot determine.Different high-resolution predictions can produce similar lower-resolution observations.
Constrained Learning: The KL prior limits deviation from the pretrained latent distribution, while EMA consistency anchors the adapted model at the supervision boundary.Together, these constraints add detail without allowing arbitrary drift at deeper scales.
Evaluation Scope: Ground-truth fidelity is reported only at 4× in Table 1 because deeper recursion lacks genuine target images.The table compares in-domain 4KLSDB with out-of-domain DIV2K and DIV8K under matched recursion and evaluation settings.
Training and Inference: Only adapter parameters are optimized through the recursive chain; the backbone, decoder, VLM prompter, quality model, and base model remain frozen.Training-only branches are removed at inference, when the learned adapter operates with the frozen components.

4. Experiments

ORACLE ZOOM is evaluated across seven benchmarks and four magnification scales using perceptual quality, fidelity, projected-reference consistency, and hallucination measures. It generally preserves quality and observable evidence deeper into recursion while reducing unsupported artifacts.

Experiment setup: ORACLE ZOOM is evaluated on seven benchmarks across four magnification scales using no-reference quality, reference-based fidelity, and hallucination measures.The evaluation includes aggregate metrics across recursion scales and ground-truth fidelity where 4× targets are available.
Results across recursion depth: ORACLE ZOOM leads MUSIQ, MANIQA, CLIPIQA, LPIPS, and DISTS when results are averaged over all seven test sets, ranking second on NIQE.Table 2 aggregates no-reference metrics across all four recursion scales and ground-truth fidelity across datasets with 4× targets.
Results across recursion depth: 0.706 CLIPIQA at 256× compares with 0.579 for CoZ, 0.532 for OSEDiff, and 0.463 for SwinIR, while ORACLE ZOOM remains above 0.70 throughout recursion.The method also leads CoZ by 0.036 and 0.040 CLIPIQA on DIV2K and DIV8K at 4×, versus 0.015 in-domain.
Projected-reference fidelity: 0.215 P-DISTS and 0.691 projected DINOv2 similarity at 16× outperform CoZ's 0.239 and 0.633, respectively, preserving more observable reference evidence.At 16×, predictions are projected back to the last target-available resolution for comparison with the aligned reference.
Results beyond available ground truth: At 64× and 256×, ORACLE ZOOM's hallucination rate decreases to 0.21 and 0.14, while CoZ's rises to 0.55 and 0.70.At 4× and 16×, the two methods have similar hallucination rates; deeper-scale evaluation measures consistency with preceding zooms rather than exact unseen-detail recovery.
Ablation analysis: Removing Lqual lowers CLIPIQA from 0.714 to 0.594, while removing Lprior raises CLIPIQA to 0.794 but worsens P-DISTS to 0.330 and hallucination to 0.907.The ablation indicates that quality guidance supplies fine-scale detail, whereas the latent prior constrains quality-driven drift and unsupported patterns.

5. Conclusion

ORACLE ZOOM extends recursive super-resolution beyond the last target-available scale by training on its own trajectory while retaining verifiable visual evidence. Its combined objectives support high quality and fidelity with fewer hallucinations across seven datasets.

Conclusion: ORACLE ZOOM carries observable ground-truth evidence into deeper recursive scales through aligned projection, while a no-reference objective guides unresolved detail.Direct supervision anchors target-available predictions, and KL prior regularization plus EMA consistency limit drift.
Conclusion: Across seven datasets, ORACLE ZOOM achieves state-of-the-art fidelity and no-reference super-resolution quality while significantly reducing hallucinations.The approach uses a 7.1M-parameter LoRA trained on 1,000 images in addition to the base SR model.

6. Limitations

The evaluation cannot verify exact recovery beyond 4× because deeper-scale ground truth is unavailable, and its settings do not cover physical camera zoom or arbitrary user-selected regions. Additional limitations concern domain-specific degradations and dependence on fixed pretrained quality and SR priors.

Evaluation boundary: Beyond 4×, projected-reference metrics and an anchored VLM judge assess consistency with observable evidence rather than exact recovery.Deeper-scale ground truth is unavailable, so exact fine-detail reconstruction cannot be measured directly.
Scope boundary: The synthetic center-crop recursion setting does not directly establish performance for physical camera zoom or arbitrary user-selected regions.The training data are also curated 4K photographic images, with domain-specific degradations remaining challenging on DRealSR and RealSR.
Model dependence: A fixed no-reference quality model and pretrained SR prior can bias the type of detail encouraged at unsupported scales.Outputs may appear plausible even when fine-scale recovery cannot be verified.

A. Additional Method Details and Analysis

Beyond the last supervised scale, ORACLE ZOOM preserves verifiable evidence by projecting deeper predictions back to the final ground-truth resolution, while separately addressing unobservable fine detail. Its complementary objectives combine cross-scale consistency, quality guidance, and a pretrained latent prior.

Extending target evidence: The last ground-truth target remains useful at deeper scales because deterministic zoom paths identify corresponding regions for cross-scale comparison.Projecting a deeper prediction back to the last supervised resolution preserves the observable field of view, but not newly generated fine detail.
Limits of cross-scale supervision: Cross-scale supervision constrains information that remains visible after projection, whereas multiple high-resolution images can share the same lower-resolution observation.Projection removes fine-scale information, creating ambiguity among deeper predictions.
Complementary objectives: The beyond-boundary objectives preserve verifiable evidence, guide unresolved fine-scale detail, and keep predictions close to the pretrained SR model.Lxscale, Lqual, and Lprior provide complementary supervision when direct ground truth is unavailable.
Latent prior: A bounded KL constraint limits how far quality-guided predictions can move from the pretrained SR model in quality-score space.Under the shared isotropic Gaussian assumption, latent MSE is equivalent to the closed-form KL divergence up to a constant absorbed into λprior.

A.4. Proof of Bounded Quality Deviation

The method bounds quality-driven deviation from the pretrained model while propagating deeper-scale objectives through the recursive chain. EMA consistency stabilizes the transition where direct ground-truth supervision ends.

Bounded quality deviation: A bounded KL constraint limits the adapted prediction’s movement from the pretrained SR model in quality-score space, while cross-scale consistency preserves verifiable evidence.The two constraints address complementary risks: quality guidance can alter detail, while cross-scale consistency retains observable content.
Objective formulation: The constrained optimization formulation preserves target fidelity, cross-scale evidence, and pretrained-model proximity while optimizing quality.The auxiliary EMA term is added to the objective after introducing non-negative multipliers for the principal constraints.
Recursive gradient propagation: Deeper-scale losses backpropagate through preceding predictions, allowing objectives beyond the supervision boundary to shape the complete recursive chain.The computation graph is retained rather than detached, so earlier model-generated inputs receive gradients from later losses.
Supervision boundary: EMA consistency stabilizes the boundary where direct supervision ends by keeping the current latent representation near a slower-moving parameter average.Lsup matches available targets, whereas Lema stabilizes the representation before recursion enters target-free scales.

A.8. Training-Only Supervision and Inference

ORACLE ZOOM uses additional references and objectives only during training, while inference retains the base recursive SR transition. Training beyond the ground-truth boundary combines carried-forward evidence, quality guidance, prior regularization, and EMA stabilization.

Inference pipeline: At inference, only the trained recursive SR transition is used; ground-truth references, quality models, priors, and EMA branches are training-only signals.The additional supervision mechanisms therefore do not introduce new inference-time supervision.
Objective roles: For target-available scales, direct supervision matches observed targets; beyond them, cross-scale consistency carries verifiable evidence while quality guidance handles unresolved detail.Prior regularization constrains the generated detail, and EMA consistency stabilizes the transition out of direct supervision.
Relation to self-distillation: ORACLE ZOOM shares OPSD’s use of model-generated intermediate predictions but differs by addressing the absence of high-resolution targets beyond the ground-truth boundary.Its supervision combines carried-forward reference evidence with guidance for target-unavailable fine detail.

B. Implementation Details

The implementation freezes the main components while optimizing a rank-16 LoRA adapter, adds little inference latency, and evaluates faithfulness across recursion depth using anchored pairwise judgments. Ablations examine supervision placement and reference quality.

Compute and latency: 1.505 s per recursive step is reported for ORACLE ZOOM versus 1.486 s for CoZ, while adapter merging reduces the SR-stage gap to 0.001 s.The latency evaluation uses 500 DIV2K images on one H100 with batch size 1.
Training configuration: Only the rank-16 LoRA adapter is optimized; the SR backbone, VAE decoder, VLM prompter, quality model, and base SR model remain frozen.The training configuration uses LoRA on selected SD3 transformer projection layers.
Faithfulness evaluation: ORACLE ZOOM and CoZ are comparable at 4× and 16×, while ORACLE ZOOM is preferred at 64× and 256× in pairwise faithfulness evaluation.Win rates use decided comparisons only, with Wilson 95% confidence intervals.
Supervision ablation: Decode-space LPIPS improves LPIPS, DISTS, and DINOv2 similarity over CoZ, whereas direct latent-space MSE degrades fidelity.This ablation evaluates where target-available supervision should be applied at 4×.
Reference quality: Exact ground truth gives the strongest 4× fidelity, while a mildly degraded reference retains most of the improvement over CoZ.The comparison shows that reference quality affects fidelity but need not be exact to retain most of the gain.

C.1. Analysis

The analysis examines supervision choices, judging protocols, reference quality, and no-reference quality objectives for recursive super-resolution. Decode-space supervision and TOPIQ-NR provide the strongest reported quality, while reference quality affects supervision outcomes.

Target-available supervision: Decode-space LPIPS substantially improves LPIPS, DISTS, and DINOv2 similarity over latent MSE at 4×.Latent MSE performs worse than CoZ on all three fidelity metrics.
Evaluation protocol: Pairwise fidelity judging uses the target high-resolution image as the anchor at 4×, then the previous zoom level at 16× and deeper.Candidates are evaluated by structural and textural agreement, with unsupported or contradictory detail treated as hallucination.
Reference quality: The exact ground-truth reference performs best, while a mildly degraded reference remains close and still improves substantially over CoZ.Combining exact and degraded references is slightly weaker than using the exact target alone.
Quality model: TOPIQ-NR gives the strongest overall and 256× no-reference quality among the evaluated quality models.All three choices give similar 4× fidelity.
Quality model: HyperIQA reduces mean CLIPIQA by 0.031 and 256× quality by 0.027, while Laplacian sharpness reduces them by 0.100 and 0.128.Hallucination differences between the variants are not significant.

C.2. Results

Across seven datasets, OracleZoom consistently improves recursive SR quality, with the clearest advantages after the target-available scale. Qualitative examples likewise show better preservation of material structure and less flattening or synthetic texture at deeper zooms, although deep-scale ground truth is unavailable.

Quantitative results: ORACLE ZOOM achieves the highest CLIPIQA at every scale on every dataset, with separation clearer beyond the target-available scale.MUSIQ and MANIQA differences are smaller at 4× but favor ORACLE ZOOM more consistently at deeper scales.
Quantitative results: 0.667 versus 0.540 CLIPIQA is the 256× comparison on RealSR, where the gap increases with recursion despite smaller earlier-scale differences.This supports persistence under real-camera degradations.
Quantitative results: 0.771 versus 0.579 CLIPIQA is ORACLE ZOOM’s 256× result against CoZ on FFHQ, its largest reported margin.ORACLE ZOOM also preserves the strongest MUSIQ and MANIQA at deeper scales.
Quantitative results: 0.102–0.192 is the 256× CLIPIQA margin over CoZ across datasets, exceeding 0.10 on every dataset.The differences are most pronounced at 64× and 256× after several recursive SR steps.
Qualitative results: ORACLE ZOOM better preserves fold relief, knit structure, weave, fiber structure, and seam boundaries than the baselines in recursive qualitative examples.The examples span 4KLSDB and DIV2K, where OSEDiff flattens textures and CoZ loses material structure.
Qualitative results: ORACLE ZOOM retains irregular stone texture, surface grain, garment structure, skin variation, pod boundaries, and leaf structure through deeper zooms.The pattern appears across carved stone, real-camera images, faces, and vegetation.
Scope of qualitative evidence: The qualitative examples do not establish exact recovery of unseen high-frequency detail because ground truth is unavailable at deeper scales.They instead assess compatibility with visual structures present in preceding zooms.