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

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

cs.CVSep 6, 2026

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

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

Organizations: University of Maryland, Baltimore County

Abstract

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/ .

Figures & tables

Figure 2: Overview of Oracle Zoom . The shared LoRA [ 14 ] adapted SR is applied recursively; available targets provide direct and cross-scale supervision, while a frozen quality model, KL prior, and EMA consistency constrain target-unavailable predictions. All auxiliary branches are training-only. By carrying the last verifiable ground-truth evidence beyond the supervision boundary, Oracle Zoom addresses the key gap in recursive SR where deeper predictions must otherwise be synthesized without direct visual supervision.

Method	In-domain						Out-of-domain
	4KLSDB						DIV2K				DIV8K
	No-reference quality				GT fidelity		No-reference quality				No-reference quality				GT fidelity
	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	CLIPIQA $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	CLIPIQA $\uparrow$	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	CLIPIQA $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
$4\times$ magnification
HiT-SR [ 93 ]	8.76	40.11	0.396	0.396	0.408	0.232	7.89	39.21	0.403	0.374	8.05	37.06	0.395	0.366	0.448	0.265
MambaIR [ 20 ]	8.68	41.24	0.405	0.412	0.412	0.233	7.81	40.57	0.419	0.386	8.00	38.53	0.412	0.377	0.449	0.268

Table 1: Per-dataset results across recursion depth on 4KLSDB ( in-domain ), DIV2K, and DIV8K ( out-of-domain ). All methods use the same CoZ recursion with matched zoom paths, crops, prompts, and evaluation. GT fidelity is reported only where genuine

4\times

targets are available; deeper scales have no ground truth. Bold : best; underline : second-best.

Figure 3: Qualitative comparison across $4$ – $256\times$ . OSEDiff smooths local structure, while CoZ develops repetitive textures at deeper zooms. Oracle Zoom preserves sharper, more detailed fur and skin structure through $256\times$ . Colored boxes mark the next zoom region.

Figure 4: Effect of quality guidance and latent prior. Without $\mathcal{L}_{\mathrm{qual}}$ , predictions become smooth, while removing $\mathcal{L}_{\mathrm{prior}}$ introduces repetitive patterns at deeper zooms. The full objective preserves coherent detail through $256\times$ . Colored boxes mark the next zoom region.

Method	No-reference quality				GT fidelity @ $4\times$
Method	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	CLIPIQA $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
HiT-SR [ 93 ]	14.22	27.06	0.379	0.414	0.341	0.222
MambaIR [ 20 ]	14.69	27.34	0.382	0.428	0.343	0.224
SwinIR [ 40 ]	7.88	34.02	0.477	0.456	0.228	0.175
SeeSR [ 83 ]	8.35	48.46	0.532	0.546	0.216	0.164
OSEDiff [ 82 ]	6.70	51.38	0.548	0.581	0.336	0.228

Table 2: Aggregate results over all seven test sets. No-reference metrics average all four recursion scales; GT fidelity averages the four datasets with

4\times

targets. Oracle Zoom achieves the best MUSIQ, MANIQA, CLIPIQA, LPIPS, and DISTS. Bold : best; underline : second-best. Per-dataset results are in Table 1 .

Figure 5: Results across recursion depth. (a) Mean CLIPIQA over the test sets. (b) Projected-reference fidelity at $16\times$ , measured by P-DISTS ( $\downarrow$ ). (c) Anchored hallucination rate for Oracle Zoom and CoZ, with $95\%$ Wilson intervals. Oracle Zoom maintains perceptual quality, retains more of the last available reference, and reduces hallucination as recursion deepens.

Figure 6: Projected-reference fidelity at $16\times$ . Native predictions appear plausible but cannot be compared with a $16\times$ target. Projecting them to $4\times$ enables direct comparison with the aligned ground-truth region. CoZ and OSEDiff alter the cable and panel boundaries, whereas Oracle Zoom remains closer to the reference, as reflected by the residual maps and projected scores.

Variant	LPIPS ${}_{4\times}\downarrow$	P-DISTS ${}_{16\times}\downarrow$	CLIPIQA ${}_{16\times}\uparrow$	Halluc. ${}_{16\times}\downarrow$
Full Oracle Zoom	$\mathbf{0.199}$	$0.215$	$0.714$	$\mathbf{0.303}$
$-\mathcal{L}_{\mathrm{sup}}$	$0.215$	$0.222$	$0.719$	$0.353$
$-\mathcal{L}_{\mathrm{xscale}}$	$0.196$	$0.232$	$0.727$	$0.339$
$-\mathcal{L}_{\mathrm{qual}}$	$0.202$	$\mathbf{0.208}$	$0.594$	$0.370$
$-\mathcal{L}_{\mathrm{prior}}$	$0.197$	$0.330$	$0.794$	$0.907$
$-\mathcal{L}_{\mathrm{ema}}$	$0.203$	$0.216$	$0.702$	$0.305$

Table 3: Ablation of the training objectives. Removing each term degrades a different aspect of performance, while the full objective gives the optimal fidelity, quality, and hallucination.

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Value
Backbone & prompter (both frozen)
SR backbone	OSEDiff on SD3-medium
Prompter (VLM)	Qwen2.5-VL-3B-Instruct + CoZ adapter
Prompt length	$32$ max new tokens
LoRA adapter (only trainable weights)
Placement	SD3 transformer (MMDiT)

Appendix

Table 4: Training configuration for Oracle Zoom . Only the rank-16 LoRA adapter is optimized; the SR backbone, VAE decoder, VLM prompter, quality model, and base SR model remain frozen.

per recursion step	SR	PE	total
CoZ (the blind baseline)	$0.152$	$1.334$	$1.486$
Ours (as evaluated)	$0.171$	$1.334$	$1.505$
Ours, adapter merged	$0.153$	$1.347$	$1.500$

Appendix

Table 5: Inference latency per recursive step on 500 DIV2K images using one H100. Oracle Zoom adds only

0.019

s to the SR stage over CoZ, while merging the LoRA adapter reduces the SR-stage gap to

0.001

Scale	ours	CoZ	tie	abst.	win-rate [95% CI]
$4\times$	29	29	54	8	$0.50$ $[0.38,0.63]$
$16\times$	60	53	6	1	$0.53$ $[0.44,0.62]$
$64\times$	77	37	6	0	$\mathbf{0.68}$ $[0.59,0.75]$
$256\times$	91	25	4	0	$\mathbf{0.78}$ $[0.70,0.85]$

Appendix

Table 6: Pairwise faithfulness evaluation across recursion depth. Oracle Zoom and CoZ are comparable at

4\times

and

16\times

, while Oracle Zoom is preferred at

64\times

and

256\times

. Win rates use decided comparisons only; brackets show Wilson

95\%

confidence intervals.

Supervision	LPIPS $\downarrow$	DISTS $\downarrow$	DINOv2 $\uparrow$
CoZ (baseline)	0.2092	0.1585	0.9173
Latent MSE	0.2425	0.1936	0.9049
Decode-space LPIPS	0.1707	0.1407	0.9439

Appendix

Table 7: Where to apply target-available supervision at

4\times

. Decode-space LPIPS improves LPIPS, DISTS, and DINOv2 similarity over CoZ, whereas direct latent-space MSE degrades fidelity.

Reference	LPIPS $\downarrow$	DISTS $\downarrow$	DINOv2 $\uparrow$	FID $\downarrow$	TOPIQ-FR $\uparrow$
Exact GT	0.178	0.145	0.945	41.1	0.641
Mildly degraded	0.183	0.146	0.942	42.3	0.633
Combined	0.188	0.147	0.940	44.8	0.621
CoZ (baseline)	0.234	0.164	0.908	60.7	0.548

Appendix

Table 8: Effect of reference quality at

4\times

. Exact ground truth gives the strongest fidelity, while a mildly degraded reference retains most of the improvement over CoZ.

Panel of Figure 7 — Figure 7: The rubric used for pairwise judging at $4\times$ , with the target high-resolution image as the anchor.

Panel of Figure 8 — Figure 8: The rubric used for pairwise judging at $16\times$ and deeper, with the previous zoom level as the anchor.

Figure 9: Effect of the no-reference quality objective across $4$ – $256\times$ . Two 4KLSDB examples compare TOPIQ-NR, HyperIQA, and Laplacian sharpness with all other settings fixed. TOPIQ-NR better retains fine-scale structure at deeper zooms, consistent with Table 9 ; colored boxes mark the region enlarged at the next step.

	No-reference quality		Fidelity @ $4\times$ (GT)			Coherence	Quality score	Hallucination
Quality model	CLIPIQA $\uparrow$	CQ ${}_{256}\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	DINO ${}_{4\times}\uparrow$	Coher. $\uparrow$	TOPIQ ${}_{256}\uparrow$	Halluc. $\downarrow$
Ours (TOPIQ-NR)	0.711	0.707	0.199	0.160	0.916	0.792	0.585	0.221
HyperIQA	0.680 $-0.031$	0.680 $-0.027$	0.197	0.158	0.917	0.816 $+0.024$	0.553	0.203
Laplacian	0.610 $-0.100$	0.579 $-0.128$	0.201	0.158	0.914	0.808 $+0.016$	0.467	0.247

Appendix

Table 9: Choice of no-reference quality model. TOPIQ-NR gives the strongest overall and

256\times

no-reference quality, while all three choices give similar

4\times

fidelity. HyperIQA slightly improves coherence, whereas Laplacian sharpness substantially reduces deep-zoom quality. Hallucination differences between the three variants are not significant.

	4KLSDB			DIV8K			DRealSR			RealSR
Method	LPIPS $\downarrow$	DISTS $\downarrow$	DINO $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	DINO $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	DINO $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	DINO $\uparrow$
HiT-SR [ 93 ]	0.408	0.231	0.907	0.448	0.265	0.894	0.322	0.223	0.920	0.186	0.169	0.927
MambaIR [ 20 ]	0.412	0.233	0.903	0.449	0.268	0.897	0.326	0.225	0.916	0.185	0.169	0.928
SwinIR [ 40 ]	0.280	0.184	0.927	0.320	0.210	0.919	0.197	0.174	0.930	0.115	0.131	0.886
SeeSR [ 83 ]	0.260	0.174	0.917	0.263	0.174	0.917	0.177	0.156	0.928	0.164	0.151	0.885
OSEDiff [ 82 ]	0.335	0.215	0.793	0.380	0.243	0.788	0.318	0.231	0.805	0.310	0.223	0.729

Appendix

Table 10: Full-reference fidelity at

4\times

on the four datasets with genuine high-resolution targets. Oracle Zoom performs best on 4KLSDB and DIV8K, while results are more mixed on the real-camera DRealSR and RealSR sets. Best is bold ; second-best is underlined .

Scale	Method	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	CLIPIQA $\uparrow$
$4\times$	HiT-SR [ 93 ]	8.76	40.1	0.396	0.396
	MambaIR [ 20 ]	8.68	41.2	0.405	0.412
	SwinIR [ 40 ]	6.02	51.3	0.500	0.465
	SeeSR [ 83 ]	5.44	63.5	0.569	0.615
	OSEDiff [ 82 ]	5.00	60.3	0.559	0.608
	CoZ [ 31 ]	5.74	64.4	0.582	0.623

Appendix

Table 11: Per-scale no-reference quality on 4KLSDB. Oracle Zoom achieves the highest CLIPIQA at every scale and the best MUSIQ and MANIQA from

16\times

onward. Best is bold ; second-best is underlined .

Figure 10: Qualitative results on 4KLSDB across $4$ – $256\times$ . The zoom follows a waffle-knit garment. OSEDiff progressively flattens the fabric into a nearly uniform field with regular grid-like texture, while CoZ stretches the folds into vertical bands. Oracle Zoom better preserves the fold relief and interlocking knit structure as recursion deepens; colored boxes mark the region enlarged at the next step.

Scale	Method	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	CLIPIQA $\uparrow$
$4\times$	HiT-SR [ 93 ]	7.89	39.2	0.403	0.374
	MambaIR [ 20 ]	7.81	40.6	0.419	0.386
	SwinIR [ 40 ]	5.66	51.0	0.510	0.458
	SeeSR [ 83 ]	4.70	64.3	0.608	0.619
	OSEDiff [ 82 ]	4.93	59.5	0.585	0.638
	CoZ [ 31 ]	4.65	66.9	0.633	0.699

Appendix

Table 12: Per-scale no-reference quality on DIV2K. Oracle Zoom achieves the highest CLIPIQA at every scale and leads MUSIQ and MANIQA from

16\times

onward. Best is bold ; second-best is underlined .

Figure 11: Qualitative results on DIV2K across $4$ – $256\times$ . The zoom follows hanging cloth on a market stall. OSEDiff gradually loses the garment structure and produces a flat texture, while CoZ smooths the region into a largely uniform gradient. Oracle Zoom better preserves the visible weave, fiber structure, and seam-like boundaries through the deeper zooms; colored boxes mark the next enlarged region.

Scale	Method	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	CLIPIQA $\uparrow$
$4\times$	HiT-SR [ 93 ]	8.05	37.1	0.395	0.366
	MambaIR [ 20 ]	8.00	38.5	0.412	0.377
	SwinIR [ 40 ]	5.85	49.4	0.512	0.447
	SeeSR [ 83 ]	4.66	63.9	0.612	0.610
	OSEDiff [ 82 ]	4.88	60.6	0.596	0.639
	CoZ [ 31 ]	4.67	67.0	0.638	0.697

Appendix

Table 13: Per-scale no-reference quality on DIV8K. Oracle Zoom achieves the highest CLIPIQA at every scale and leads MUSIQ and MANIQA from

16\times

onward. Best is bold ; second-best is underlined .

Figure 12: Qualitative results on DIV8K across $4$ – $256\times$ . The zoom follows the carved folds of a sandstone statue. OSEDiff introduces regular mesh-like patterns, while CoZ smooths the surface into broad wavy structures. Oracle Zoom better retains the carved relief and irregular stone texture across recursion; colored boxes mark the next enlarged region.

Scale	Method	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	CLIPIQA $\uparrow$
$4\times$	HiT-SR [ 93 ]	8.72	41.0	0.429	0.377
	MambaIR [ 20 ]	8.63	42.8	0.446	0.382
	SwinIR [ 40 ]	6.51	54.6	0.542	0.482
	SeeSR [ 83 ]	5.85	66.9	0.640	0.639
	OSEDiff [ 82 ]	5.59	59.3	0.587	0.636
	CoZ [ 31 ]	6.14	67.2	0.642	0.670

Appendix

Table 14: Per-scale no-reference quality on DRealSR. Oracle Zoom achieves the highest CLIPIQA at every scale and leads MUSIQ and MANIQA throughout the target-unavailable zooms. Best is bold ; second-best is underlined .

Figure 13: Qualitative results on DRealSR across $4$ – $256\times$ . The zoom follows the surface of an unglazed stoneware planter. Both baselines progressively lose the original surface variation and approach a nearly flat field at deeper scales. Oracle Zoom retains a granular, directionally varying surface structure through $256\times$ ; colored boxes mark the region enlarged at the next step.

Scale	Method	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	CLIPIQA $\uparrow$
$4\times$	HiT-SR [ 93 ]	9.55	42.4	0.407	0.542
	MambaIR [ 20 ]	9.44	42.8	0.407	0.561
	SwinIR [ 40 ]	5.56	52.1	0.503	0.570
	SeeSR [ 83 ]	4.79	65.6	0.597	0.709
	OSEDiff [ 82 ]	4.76	64.2	0.594	0.726
	CoZ [ 31 ]	5.22	65.6	0.616	0.741

Appendix

Table 15: Per-scale no-reference quality on FFHQ. Oracle Zoom achieves the highest CLIPIQA at every scale and leads MUSIQ and MANIQA from

16\times

through

256\times

. Best is bold ; second-best is underlined .

Figure 14: Qualitative results on FFHQ across $4$ – $256\times$ . The zoom moves from the nose bridge toward the cheek. OSEDiff develops a regular canvas-like pattern and color smearing, while CoZ produces increasingly smooth skin with isolated artificial-looking structures. Oracle Zoom better preserves connected skin texture and local pore-like variation as recursion deepens; colored boxes mark the region enlarged at the next step.

Scale	Method	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	CLIPIQA $\uparrow$
$4\times$	HiT-SR [ 93 ]	7.95	36.4	0.402	0.362
	MambaIR [ 20 ]	7.88	37.6	0.418	0.367
	SwinIR [ 40 ]	5.74	49.0	0.513	0.427
	SeeSR [ 83 ]	4.62	63.2	0.611	0.590
	OSEDiff [ 82 ]	4.81	59.5	0.587	0.620
	CoZ [ 31 ]	4.64	66.8	0.642	0.685

Appendix

Table 16: Per-scale no-reference quality on Flickr2K. Oracle Zoom achieves the highest CLIPIQA at every scale and leads MUSIQ and MANIQA from

16\times

onward, except MANIQA at

16\times

0.001

. Best is bold ; second-best is underlined .

Figure 15: Qualitative results on Flickr2K across $4$ – $256\times$ . The zoom follows dried seed pods surrounded by grass. OSEDiff introduces regular repeated texture before becoming increasingly smooth, while CoZ fragments and smears the pod structure at deeper scales. Oracle Zoom better preserves the pod boundaries, curled husk structure, and nearby leaf blades through the recursive zoom; colored boxes mark the next enlarged region.

Scale	Method	NIQE $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$	CLIPIQA $\uparrow$
$4\times$	HiT-SR [ 93 ]	9.22	34.7	0.372	0.346
	MambaIR [ 20 ]	9.29	35.0	0.381	0.364
	SwinIR [ 40 ]	6.78	50.4	0.510	0.447
	SeeSR [ 83 ]	6.41	61.1	0.588	0.557
	OSEDiff [ 82 ]	5.83	57.7	0.578	0.612
	CoZ [ 31 ]	6.18	64.3	0.619	0.643

Appendix

Table 17: Per-scale no-reference quality on RealSR. Oracle Zoom achieves the highest CLIPIQA at every scale and leads MUSIQ and MANIQA throughout the target-unavailable zooms. Best is bold ; second-best is underlined .

Figure 16: Qualitative results on RealSR across $4$ – $256\times$ . The zoom follows a garment hanging on a clothing rack. At deeper scales, OSEDiff and CoZ increasingly reduce the garment to smooth or regularly patterned regions. Oracle Zoom retains more of the garment structure and fine fabric variation, with the clearest difference appearing at $256\times$ ; colored boxes mark the region enlarged at the next step.

Image Super-Resolution

Image super-resolution reconstructs high-resolution images from lower-resolution observations, balancing detail, fidelity, and computational cost.

See all 250 papers

OARS: Process-Aware Online Alignment for Generative Real-World Image Super-Resolution
Mar 13, 2026Shijie Zhao, Xuanyu Zhang, Bin Chen +8Image Super-Resolution Diffusion Model Alignment
ByteDance Inc. · Peking University · The Chinese University of Hong Kong
Single Image Super-Resolution via Bivariate `A Trous Wavelet Diffusion
Mar 7, 2026Maryam Heidari, Nantheera Anantrasirichai, Alin AchimImage Super-Resolution Diffusion Models for Image Restoration
Visual Information Laboratory, University of Bristol, Bristol, UK
TOC-SR: Task-Optimal Compact diffusion for Image Super Resolution
May 4, 2026Sowmya Vajrala, Akshay Bankar, Manjunath Arveti +5Diffusion Distillation Few-Step Diffusion Sampling
Samsung Research Institute Bangalore, India

Mar 13, 2026J/K move · Enter open · S save

OARS: Process-Aware Online Alignment for Generative Real-World Image Super-Resolution

Shijie Zhao, Xuanyu Zhang, Bin Chen, Weiqi Li+7

ByteDance Inc. · Peking University · The Chinese University of Hong Kong

Image Super-Resolution Diffusion Model Alignment RL for Image Generation

Open paper AskPDF arXiv

Aligning generative real-world image super-resolution models with human visual preference is challenging due to the perception--fidelity trade-off and diverse, unknown degradations. Prior approaches rely on offline preference optimization and static metric aggregation, which are often non-interpretable and prone to pseudo-diversity under strong conditioning. We propose OARS, a process-aware online alignment framework built on COMPASS, a MLLM-based reward that evaluates the LR to SR transition by jointly modeling fidelity preservation and perceptual gain with an input-quality-adaptive trade-off. To train COMPASS, we curate COMPASS-20K spanning synthetic and real degradations, and introduce a three-stage perceptual annotation pipeline that yields calibrated, fine-grained training labels. Guided by COMPASS, OARS performs progressive online alignment from cold-start flow matching to full-reference and finally reference-free RL via shallow LoRA optimization for on-policy exploration. Extensive experiments and user studies demonstrate consistent perceptual improvements while maintaining fidelity, achieving state-of-the-art performance on Real-ISR benchmarks.

Abstract

Image Super-Resolution

Explore similar work

OARS: Process-Aware Online Alignment for Generative Real-World Image Super-Resolution

Single Image Super-Resolution via Bivariate `A Trous Wavelet Diffusion

TOC-SR: Task-Optimal Compact diffusion for Image Super Resolution