# OracleZoom：基于在策略自蒸馏的参考约束递归图像超分辨率

- 来源：HuggingFace Daily Papers（社区热门论文）
- 发布时间：2026-09-06 08:00
- AIHOT 分数：38
- AIHOT 链接：https://aihot.news/items/cmtvf0he602x0ronbclxtst49
- 原文链接：https://arxiv.org/abs/2609.06490

## AI 摘要

OracleZoom 提出一种受在策略自蒸馏启发的参考约束递归图像超分辨率框架，在监督边界之外用最后一张可用真值图像作为参考，将深层预测拆分为可验证内容与未解析细节，分别用跨尺度一致性和无参考质量目标约束，并以 KL 约束的潜变量先验和 EMA 一致性稳定训练。

## 正文

Abstract

Recursive Image Super-Resolution (SR) extends fixed-scale SR to extreme magnification by feeding predictions back into the same model, analogous to zooming an image repeatedly. The source resolution required for ground truth grows geometrically, leaving deeper zoom predictions unsupervised. Inspired by on-policy self-distillation, Oracle Zoom trains on its own recursive predictions and uses the last available ground-truth image as a reference 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 latent prior limits quality drift, and EMA consistency stabilizes training. Across seven datasets, Oracle Zoom achieves state-of-the-art (SOTA) SR quality across scales (averaging 0.713 CLIPIQA), with larger gains on deeper scales while reducing hallucinations.

媒体内容 · 前往原文查看

Figure 1: Recursive SR exceeds practical ground-truth resolution: a 256× target stores 52 GB per image, supervision ends at 4× while the model continues to reuse its predictions. Oracle Zoom uses the last ground-truth image as a reference beyond that boundary, constrains content through cross-scale alignment, and guides remaining detail under a KL-constrained prior, reducing hallucination by 2–5×.

1 Introduction

Single-image Super-Resolution (SR) reconstructs a high-resolution image from a low-resolution observation. Recent generative SR methods [12, 7, 94, 74, 18] use diffusion priors to recover realistic high-frequency detail beyond conventional regression-based reconstruction [83, 87, 80, 82]. Recursive SR extends this setting to extreme magnification by repeatedly applying an SR model, where the prediction at one scale becomes the input to the next; equivalent to repeatedly zooming an image. Chain-of-Zoom (CoZ) [31] recursively applies a fixed-scale SR model under multi-scale VLM guidance to reach magnifications up to 256×.

High-fidelity recursive SR can support safety-critical applications across domains [59, 58, 50, 57, 27, 61, 60, 28, 63, 67, 62, 26, 66, 65, 64]. A fundamental difficulty in recursive SR is obtaining ground truth at every scale. Required source resolution grows rapidly with recursion: for a 512×512 input and successive 4× magnifications, targets at 4×, 16×, 64×, and 256× require source regions of 20482, 81922, 327682, and 1310722 pixels, respectively. A single uncompressed 131072×131072 RGB image requires about 52 GB of storage, making deep-scale supervision impractical. Ground truth is typically available only at earlier recursion stages. At deeper scales, recursive SR must synthesize increasingly fine detail without a corresponding visual target. VLM-generated captions provide semantic guidance [31, 53], but cannot directly verify whether synthesized textures and structures remain consistent with the observed image. This creates a supervision gap at deeper scales, where detailed synthesis lacks direct ground-truth verification.

Inspired by On-Policy Self-Distillation (OPSD) [95] and its recent visual and generative extensions [4, 88, 36, 43, 96], we train the SR model on its own recursive predictions. Each predicted image becomes the input to the next zoom during training, matching how the model operates at inference. We backpropagate through the recursive chain, so losses at deeper zooms can also update earlier predictions that become their inputs. However, following the inference trajectory does not solve the missing-ground-truth problem: beyond the last supervised scale, there is no target to constrain the newly generated detail. Our key observation is that the last available ground truth still contains verifiable information about the region being zoomed into. We therefore align this region with each deeper prediction and project the prediction back to the resolution where ground truth is observable. This preserves the ground-truth evidence that remains observable while constraining the unresolved details synthesized at deeper scales.

Building on the on-policy formulation, Oracle Zoom separates each target-unavailable prediction into what can still be verified and what cannot. (1) For the verifiable part, the aligned region of the last ground-truth target serves as a cross-scale reference: each deeper prediction is projected back and matched to this reference. (2) The remaining fine-scale detail cannot be determined by projection, since multiple high-resolution predictions can correspond to the same lower-resolution observation. We therefore use a frozen no-reference quality model to guide this detail. Because perceptual quality alone may favor sharp but unsupported patterns [3], a KL prior keeps the adapted latent distribution close to that of the pretrained SR model. Finally, EMA (exponential moving average) consistency stabilizes learning where direct supervision ends. Together, these objectives preserve observable evidence while guiding the detail that cannot be directly supervised.

Supervision Gap at Deep Recursive Scales. We identify and formulate the supervision gap in recursive SR: ground truth becomes prohibitively expensive at deeper magnifications, while the model increasingly relies on its own predictions where direct visual supervision is unavailable.

Reference-Constrained Recursion Beyond Ground Truth. We introduce Oracle Zoom, an on-policy self-distillation inspired, reference-constrained recursive SR framework that carries the last ground-truth beyond the supervision boundary without annotation at deeper scales.

Supervision for Verifiable and Unresolved Detail. We separate target-unavailable synthesis into verifiable and unresolved components: cross-scale consistency preserves observable ground-truth evidence, quality guidance supplies unresolved detail, and a KL-constrained pretrained prior with EMA consistency limits generation drift. We further establish a bound on quality-driven deviation under the KL constraint.

SOTA Quality and Fidelity with Lower Hallucination. On seven datasets, Oracle Zoom achieves (SOTA) 0.713 mean CLIPIQA, the best aggregate 4× fidelity with 0.199 LPIPS and 0.160 DISTS, 0.706 CLIPIQA at 256×. At 64× and 256×, an independent cross-family vision–language judge prefers Oracle Zoom in 68% and 78% of comparisons with a clear preference, respectively, while CoZ hallucinates 2–5× more often.

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.

2 Related Work

Super-Resolution and Extreme Magnification. Single-image SR has progressed from regression and adversarial reconstruction [41, 77, 40] to blind and real-world restoration that explicitly models unknown degradations [91, 78, 75]. More recently, diffusion and large generative priors have enabled stronger perceptual detail synthesis [17, 74, 90, 42, 83, 87, 80, 82, 86, 71]. Parallel work supports progressive or arbitrary-scale SR through pyramidal reconstruction and continuous image representations [33, 24, 11, 35, 6, 10, 79]. These approaches extend the attainable output scale, but not to the setting where model predictions are recursively reused as inputs once ground-truth supervision is no longer available. Among recent approaches, Chain-of-Zoom (CoZ) [31] addresses this extreme-magnification setting by recursively applying a fixed-scale SR model and using multi-scale VLM guidance to steer each zoom step. Our work instead focuses on preserving visual supervision along the same recursive zoom process once ground-truth targets are no longer available.

Learning Beyond Direct Supervision. When targets are unavailable, prior work has used perceptual objectives, quality estimators, teacher-student consistency, and generative priors for indirect supervision [3, 30, 85, 73, 9, 72, 8, 54]. EMA teachers provide slowly varying consistency targets [72, 8], while On-Policy Self-Distillation (OPSD) [95] trains a student on its own trajectories using privileged information available to a teacher, reducing mismatch between training and deployment states. We adopt this on-policy self-distillation perspective for recursive SR: the student is trained on images produced by preceding zooms, with ground truth providing privileged visual information during training. However, no-reference quality alone can reward plausible but unsupported detail [3, 46, 30, 85, 73, 9, 45, 68]. Motivated by distributional regularization in diffusion SR [74, 82, 70], we combine quality guidance with a KL-constrained pretrained latent prior.

3 Method

3.1 Preliminary and Problem Setup

Given a low-resolution image I0, we construct a recursive zoom sequence using cumulative magnification factors s0(=1)<s1<⋯<sK. The relative magnification (zoom) at step i is γi=si/si−1. Let 𝒵γi denote a zoom operator that selects the region to be magnified at step i. Importantly, 𝒵γi specifies the region of interest but does not perform super-resolution. The resulting input to the SR model is

Iiin=𝒵γi​(I^i−1),I^0=I0. (1)

At each step, a frozen multi-scale VLM G takes two images at different scales and generates a caption-based prompt that provides textual guidance for the selected region. At the first step, we obtain p1=G⁡(I0) while for the subsequent scales, the VLM input prompts are constructed using the preceding prediction and the current SR input

pi=G⁡(I^i−1,Iiin),i>1. (2)

Let Fθ denote the latent SR model, where the pretrained backbone remains frozen and θ represents the trainable adapter parameters. Given Iiin and pi, the model produces the latent prediction ziθ. A frozen VAE decoder D then maps latent ziθ to image space to perform super resolution

ziθ=Fθ​(Iiin,pi),I^i=D⁡(ziθ). (3)

Applying this recursively produces the SR sequence

I0→I^1→I^2→⋯→I^K. (4)

We assume that ground-truth targets are available only up to step m. Specifically, Iigt is available for i≤m, while no ground truth is available for i>m. Our goal is to learn an SR model that remains reliable to the available targets within the supervised range while maintaining reliable recursive behavior beyond target availability at deeper scales. Thus, the supervision gap emerges where recursion continues, but direct visual evidence no longer exists, motivating us to carry the last available target beyond sm.

3.2 Oracle Zoom

We propose Oracle Zoom to address this with five complementary objectives: Direct supervision anchors target-available scales, cross-scale consistency carries verifiable ground-truth information deeper into the recursion, quality guidance encourages unresolved fine-scale detail, KL prior regularization constrains this detail to the pretrained SR latent distribution, while EMA (Exponential Moving Average) consistency stabilizes learning at the supervision boundary as shown in Fig. 2. Together, these objectives separate the target-unavailable scales into what can still be verified from Imgt and what must be synthesized faithfully under constrained prior knowledge.

媒体内容 · 前往原文查看

Algorithm 1 Oracle Zoom Training and Inference

Input: Input I0, targets {Iigt}i=1m, scales {si}i=0K

Output: Trained adapter θ; recursive predictions {I^i}i=1K

1 Init: θ¯=θ

Training

2 for each training iteration do

3 Recursively obtain {I^i,ziθ,pi}i=1K ⊳ Eqs. (1)–(3)

4 Compute ℒsup for i≤m ⊳ available targets

5 Compute ℒxscale,ℒqual,ℒprior for j>m ⊳ deeper scales

6 Compute ℒema at sm ⊳ boundary

7 Update θ using Eq. (12); update θ¯ by EMA

Inference

8 for i=1 to K do

9 Iiin=𝒵γi​(I^i−1); obtain pi ⊳ zoom and prompt

10 I^i=D⁡(Fθ​(Iiin,pi)) ⊳ recursive SR

3.2.1 Learning Objectives

Supervision at Target-Available Scales. For i≤m, we supervise predictions using the available ground truth:

ℒsup=1m​∑i=1mLPIPS⁡(I^i,Iigt). (5)

This anchors the adapted SR model to observed image detail before direct supervision disappears beyond sm.

Cross-Scale Ground-Truth Consistency. For j>m, Ijgt is unavailable, but the recursive zoom path identifies the corresponding region within Imgt. We define

Im→jref=𝒜m→j​(Imgt),I^j→m=𝒫j→m​(I^j), (6)

𝒜m→j extracts aligned ground-truth region, 𝒫j→m projects deeper predictions to observable resolutions to impose

ℒxscale=1K−m​∑j=m+1KLPIPS⁡(I^j→m,Im→jref). (7)

Thus, predictions beyond sm remain constrained by verifiable ground-truth information without requiring Ijgt. This keeps the last target as a visual reference beyond sm.

Quality-Guided Detail Synthesis. Cross-scale projection cannot fully constrain high-frequency detail, since different fine-scale predictions may produce similar lower-resolution observations. We therefore use a frozen no-reference quality model 𝒬:

ℒqual=−1K−m∑j=m+1K𝒬(I^j). (8)

This encourages perceptually detailed predictions where direct high-resolution supervision is unavailable.

KL-Constrained Latent Prior. Optimizing image quality may favor unsupported high-frequency patterns [15]. We therefore constrain the adapted latent distribution toward that of the frozen SR model F0. Let zj0=sg⁡[F0​(Ijin,pj)] and model the adapted and base latents as ρjθ=𝒩⁡(zjθ,σ2​I) and ρj0=𝒩⁡(zj0,σ2​I). For latent dimensionality d,

ℒprior =1K−m∑j=m+1KDKL(ρjθ∥ρj0) (9)

=d2​σ2​(K−m)​∑j=m+1KMSE⁡(zjθ,zj0).

Thus, the KL prior reduces to latent MSE under shared isotropic covariance, with the constant incorporated into λprior. While ℒqual encourages unresolved detail, ℒprior limits deviation from the pretrained latent distribution. Together, they synthesize unverifiable detail while preventing unconstrained drift at deep scales.

EMA Consistency. At the supervision boundary sm, we maintain EMA adapter θ¯ to obtain a training-only latent:

zmema=sg⁡[Fθ¯​(Imgt,pm)],ℒema=MSE⁡(zmθ,zmema). (10)

After each optimization step, θ¯←μ​θ¯+(1−μ)​θ.

媒体内容 · 前往原文查看

Method In-domain Out-of-domain

4KLSDB DIV2K DIV8K

No-reference quality GT fidelity No-reference quality No-reference quality GT fidelity

NIQE↓ MUSIQ↑ MANIQA↑ CLIPIQA↑ LPIPS↓ DISTS↓ NIQE↓ MUSIQ↑ MANIQA↑ CLIPIQA↑ NIQE↓ MUSIQ↑ MANIQA↑ CLIPIQA↑ LPIPS↓ DISTS↓

4× 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

SwinIR [40] 6.02 51.31 0.500 0.465 0.280 0.184 5.66 51.00 0.510 0.458 5.85 49.42 0.512 0.447 0.320 0.210

SeeSR [83] 5.44 63.47 0.569 0.615 0.260 0.174 4.70 64.26 0.608 0.619 4.66 63.94 0.612 0.610 0.263 0.174

OSEDiff [82] 5.00 60.31 0.559 0.608 0.335 0.215 4.93 59.54 0.585 0.638 4.88 60.59 0.596 0.639 0.380 0.243

CoZ [31] 5.74 64.38 0.582 0.623 0.248 0.186 4.65 66.94 0.633 0.699 4.67 67.04 0.638 0.697 0.238 0.167

Oracle Zoom 6.00 59.80 0.556 0.638 0.169 0.142 4.46 66.47 0.638 0.735 4.48 67.31 0.648 0.736 0.223 0.161

16× magnification

HiT-SR [93] 14.13 19.70 0.333 0.370 – – 13.11 18.39 0.310 0.312 13.27 18.99 0.314 0.318 – –

MambaIR [20] 14.42 20.40 0.333 0.390 – – 13.28 19.23 0.312 0.329 13.42 19.78 0.315 0.334 – –

SwinIR [40] 7.34 28.38 0.427 0.454 – – 6.78 28.96 0.391 0.401 7.21 29.18 0.388 0.391 – –

SeeSR [83] 7.81 47.83 0.507 0.552 – – 6.50 51.78 0.523 0.540 6.49 52.96 0.528 0.535 – –

OSEDiff [82] 6.10 51.92 0.531 0.586 – – 5.96 53.01 0.545 0.605 5.79 53.48 0.550 0.611 – –

CoZ [31] 8.08 52.26 0.551 0.578 – – 6.29 58.49 0.595 0.652 6.32 58.89 0.602 0.663 – –

Oracle Zoom 7.16 54.80 0.556 0.658 – – 5.82 61.64 0.599 0.737 5.85 61.56 0.605 0.735 – –

64× magnification

HiT-SR [93] 16.69 23.76 0.388 0.455 – – 16.52 21.82 0.367 0.408 16.55 22.03 0.370 0.408 – –

MambaIR [20] 17.82 23.33 0.388 0.469 – – 17.10 21.64 0.366 0.419 17.09 21.77 0.369 0.418 – –

SwinIR [40] 9.02 26.52 0.489 0.477 – – 8.09 23.20 0.478 0.465 8.62 22.93 0.474 0.460 – –

SeeSR [83] 9.83 37.31 0.495 0.497 – – 8.98 42.00 0.502 0.517 8.88 43.76 0.509 0.524 – –

OSEDiff [82] 7.20 46.22 0.531 0.533 – – 7.30 47.71 0.536 0.576 7.04 47.66 0.537 0.580 – –

CoZ [31] 9.54 45.70 0.555 0.544 – – 7.86 52.06 0.581 0.623 7.78 51.99 0.584 0.630 – –

Oracle Zoom 8.47 50.80 0.579 0.651 – – 7.12 56.02 0.597 0.727 7.23 55.61 0.597 0.726 – –

256× magnification

HiT-SR [93] 17.81 26.23 0.428 0.494 – – 17.64 26.32 0.421 0.480 17.52 26.34 0.419 0.487 – –

MambaIR [20] 18.97 26.04 0.426 0.509 – – 18.47 26.23 0.418 0.493 18.43 26.39 0.416 0.497 – –

SwinIR [40] 10.77 28.54 0.495 0.467 – – 9.96 27.35 0.504 0.466 10.01 27.34 0.504 0.473 – –

SeeSR [83] 11.56 33.66 0.503 0.496 – – 11.16 36.26 0.502 0.497 10.79 37.64 0.505 0.510 – –

OSEDiff [82] 8.18 42.57 0.525 0.502 – – 8.53 43.71 0.528 0.543 8.25 43.98 0.528 0.551 – –

CoZ [31] 10.41 44.13 0.555 0.538 – – 9.24 48.26 0.576 0.599 8.83 48.57 0.577 0.608 – –

Oracle Zoom 9.31 49.31 0.589 0.664 – – 8.30 51.83 0.595 0.703 8.07 51.85 0.593 0.710 – –

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

3.2.2 Constrained Learning and Overall Objective

Beyond target-available scales, Oracle Zoom improves perceptual quality while preserving observable ground-truth evidence and proximity to pretrained SR latent distribution:

maxθ 1K−m​∑j=m+1K𝒬⁡(I^j) (11)

s.t. ℒsup≤ϵsup,ℒxscale≤ϵxscale,ℒprior≤ϵprior.

The constraints preserve target fidelity, cross-scale agreement with available visual evidence, and proximity to the pretrained latent distribution. In practice, we optimize the corresponding penalized objective:

ℒ= ℒsup+λxscale​ℒxscale+λqual​ℒqual (12)

+λprior​ℒprior+λema​ℒema.

Taken together, the objective converts target-unavailable recursive SR from unconstrained synthesis into optimization around observed evidence and the pretrained SR prior [47].

Bounded Quality Deviation. The cross-scale and KL constraints control different aspects beyond target availability. Cross-scale consistency keeps predictions aligned with the available ground-truth evidence, while the KL constraint keeps quality optimization close to the pretrained SR model. Thus, quality optimization can add unresolved detail without drifting arbitrarily far at deeper scales.

Proposition 1. Let q⁡(z)=𝒬⁡(D⁡(z)) be locally Lq-Lipschitz around zj0 under dz​(z,z′)=MSE⁡(z,z′). For any j>m, if

DKL(ρjθ∥ρj0)≤ϵprior⟹ |𝒬⁡(I^j)−𝒬⁡(D⁡(zj0))| (13)

≤Lq​σ​2​ϵpriord.

Proof. For shared isotropic covariance in the latent distributions, DKL(ρjθ∥ρj0)=d2​σ2MSE(zjθ,zj0). Equation (13) therefore gives dz​(zjθ,zj0)≤σ​2​ϵprior/d. Applying the local Lipschitz condition to q yields Eq. (13). Together with ℒxscale≤ϵxscale, this controls observable disagreement and quality-driven deviation (detailed proof is provided in Appendix). The proof shows that the quality objective can improve unseen detail without moving the prediction arbitrarily far from the pretrained SR model.

3.3 Training and Inference

Algorithm 1 summarizes training and inference for Oracle Zoom. During training, the adapter parameters θ are shared across scales, while the SR backbone, VAE decoder D, VLM prompter G, quality model 𝒬, and base model F0 remain frozen. Only θ is optimized by gradients, while θ¯ is updated by EMA. We backpropagate through successive predictions, allowing deeper-scale objectives to also update earlier steps. Ground-truth targets, cross-scale references, 𝒬, F0, and the EMA branch are used only for training.

At inference, the learned adapter is used with the frozen SR backbone, decoder, and VLM prompter over scales s1,…,sK. Thus, without requiring ground truth or training-only branch, Oracle Zoom improves the shared SR transition beyond the supervision boundary.

4 Experiments

We evaluate Oracle Zoom on seven benchmarks across four magnifications, studying no-reference quality, reference-based fidelity, and hallucination beyond target.

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

Figure 4: Effect of quality guidance and latent prior. Without ℒqual, predictions become smooth, while removing ℒprior introduces repetitive patterns at deeper zooms. The full objective preserves coherent detail through 256×. Colored boxes mark the next zoom region.

4.1 Experiment Setup

Datasets. For training Oracle Zoom, we sample a 1,000-image training set from 129,484 candidates in 4KLSDB [98] training split. We retain images with a short side of at least 2048 pixels, discard the lowest-quality decile, balance caption-derived content groups, filter with DFN5B/SigLIP2 agreement on photographic content, and remove near-duplicates from all evaluation sets using image similarity and CLIP verification to improve diversity on training data [16]. We evaluate on 4KLSDB [98], DIV2K [1], DIV8K [19], DRealSR [81], RealSR [5], FFHQ [29], and Flickr2K [41]. Following CoZ [31], every method processes the same 512×512 center crop through four 4× SR steps, producing 4×, 16×, 64×, and 256× outputs.

Evaluation Metrics.

For no-reference quality, when ground truth is unavailable for comparison (at 16×, 64×, 256×); we evaluate on NIQE [46] (↓), MUSIQ [30] (↑), MANIQA [85] (↑), and CLIPIQA [73] (↑). At 4×, where high-resolution ground truth is available, we measure fidelity using LPIPS [92], DISTS [13] (↓), and DINOv2 cosine similarity [49] (↑) on 4KLSDB, DIV8K, DRealSR, and RealSR. At 16×, we project each prediction back to the last target-available resolution and compare it with the aligned 4× ground-truth region using projected DISTS (P-DISTS) and projected DINOv2. Beyond 4×, we additionally use InternVL3.5-38B [76] as an anchored pairwise judge from a different model family than the Qwen prompter borrowed from [31]. The judge evaluates 120 region-aligned examples per scale; ties and abstentions are excluded from the win rate. Moreover, TOPIQ-NR [9], which supplies ℒqual, is excluded from the primary evaluation.

Implementation Details.

We build Oracle Zoom on CoZ’s one-step OSEDiff [82] using SD3-medium as the frozen SR backbone and use its GRPO-tuned Qwen2.5-VL-3B-Instruct as the frozen prompter; the VAE decoder is kept fixed. We perform parameter-efficient adaptation with a rank-16 LoRA [23, 69] on the SD3 transformer, with 7.1M trainable parameters. We train the shared adapter across the 4×→16× recursive chain and backpropagate through the 4× prediction. We set λxscale=1.0, λqual=0.4, λprior=8.0, and λema=0.1, with EMA decay 0.95. We optimize in fp32 using AdamW with a learning rate of 5×10−5, weight decay of 10−2, a 500-step warmup, and an effective batch size of 4. Final checkpoint is selected by early stopping after approximately 9.3k optimization steps. Complete implementation and hyperparameter details are provided in Appendix.

Baseline Models.

We compare with three regression SR models, SwinIR [40], HiT-SR [93], MambaIR [20], two diffusion SR models, SeeSR [83] and OSEDiff [82], and CoZ [31]. For comparability, in every baseline, we use identical same CoZ recursion [31], with matched inputs, zoom paths, crop geometry, prompts, and metrics.

媒体内容 · 前往原文查看

Method No-reference quality GT fidelity @4×

NIQE↓ MUSIQ↑ MANIQA↑ CLIPIQA↑ LPIPS↓ DISTS↓

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

CoZ [31] 7.42 55.23 0.587 0.621 0.215 0.170

Oracle Zoom 6.90 57.80 0.598 0.713 0.199 0.160

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

4.2 Results

Table 1 jointly reports no-reference quality at all four recursion scales and GT fidelity at 4×, separating the in-domain set from the two out-of-domain benchmarks. At 4×, the two metric families show whether improved perceptual quality is accompanied by closer agreement with ground truth; beyond this scale (4×), the GT-fidelity entries are omitted because no target exists at 16×→256×. Table 2 aggregates every metric; averaging performance over all seven test sets and all zooming scales.

Results Across Recursion Depth.

In Tab. 1, Oracle Zoom achieves the highest CLIPIQA on all three datasets, and the best LPIPS and DISTS at 4× on the two sets that provides ground-truth target. Thus, its perceptual-quality gain does not come at the expense of fidelity at the target-available scale. From 16× onward, Oracle Zoom ranks first in CLIPIQA, MUSIQ, and MANIQA on all three datasets. The margin is wider away from the training domain: at 4× Oracle Zoom leads CoZ by 0.036 and 0.040 CLIPIQA on DIV2K and DIV8K, against 0.015 on the in-domain set. Therefore, the performance gains are generalizable. The same trend holds across the test sets in Figure 5(a), where Oracle Zoom remains above 0.70 CLIPIQA throughout the recursion. At 256×, it obtains 0.706, compared with 0.579 for CoZ, 0.532 for OSEDiff, and 0.463 for SwinIR. Averaged over all seven test sets (Table 2), Oracle Zoom leads MUSIQ, MANIQA, CLIPIQA, LPIPS, and DISTS, and is second on NIQE.

Figure 5: Results across recursion depth.(a) Mean CLIPIQA over the test sets. (b) Projected-reference fidelity at 16×, measured by P-DISTS (↓). (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.

Results on Projected-Reference Fidelity.

At 16×, when target is unavailable, but the corresponding region remains observable in the 4× ground truth. We therefore project each prediction back to the target-available resolution and compare it with the aligned reference. As shown in Figure 5(b), Oracle Zoom achieves the lowest P-DISTS (0.215) and the highest projected DINOv2 similarity (0.691), compared with 0.239 and 0.633 for CoZ. These results showcase the superiority of Oracle Zoom in preserving prior visual evidence beyond supervision boundary.

Figure 6: Projected-reference fidelity at 16×. Native predictions appear plausible but cannot be compared with a 16× target. Projecting them to 4× 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.

Results Beyond Available Ground Truth.

At 4×, 16×, Oracle Zoom and CoZ have similar hallucination rates. Their behavior diverges at deeper scales: at 64× and 256×, the hallucination rate of Oracle Zoom decreases to 0.21 and 0.14, while that of CoZ rises to 0.55 and 0.70, respectively [Figure 5(c)]. Thus, improved no-reference quality is not accompanied by greater contradiction with the observable anchor. Since ground truth is unavailable at these scales, the judge measures consistency with the preceding zooms, not the explicit recovery of unseen fine detail.

Qualitative Results.

Figure 3 depicts two examples through the 4×→16×→64×→256× trajectory. OSEDiff progressively removes local structure, while CoZ develops repetitive textures as recursion deepens. In contrast, Oracle Zoom maintains the orientation and continuity of the visible fur and skin patterns much better while producing a coherent fine-scale structure through 256×.

媒体内容 · 前往原文查看

Variant LPIPS4×↓ P-DISTS16×↓ CLIPIQA16×↑ Halluc.16×↓

Full Oracle Zoom 0.199 0.215 0.714 0.303

−ℒsup 0.215 0.222 0.719 0.353

−ℒxscale 0.196 0.232 0.727 0.339

−ℒqual 0.202 0.208 0.594 0.370

−ℒprior 0.197 0.330 0.794 0.907

−ℒema 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.

4.3 Analysis

Ablation Study.

Table 3 analyzes the contribution of each objective in Oracle Zoom. Removing ℒsup increases 4× LPIPS from 0.199 to 0.215, while removing ℒxscale increases P-DISTS from 0.215 to 0.232. Without ℒqual, CLIPIQA decreases from 0.714 to 0.594. Removing ℒprior instead increases CLIPIQA to 0.794, but P-DISTS degrades to 0.330 and hallucination rises from 0.303 to 0.907. The smaller changes after removing ℒema indicate that it acts as a lightweight training stabilizer, while the full objective provides the optimal fidelity and quality.

Quality Guidance and Latent Prior.

Figure 4 visualizes the complementary roles of ℒqual and ℒprior. Removing quality guidance produces smooth predictions with limited fine-scale structure. Removing the latent prior instead introduces repetitive patterns that score highly on the quality objective but contradict the preceding zoom. Consistent with Table 3, Figure 4 shows that variants using no-prior obtain the highest quality at the cost of fidelity and hallucination.

Qualitative Projected-Reference Comparison.

Figure 6 illustrates how deeper predictions are comparable to the last available ground truth. At 16×, no target exists, so each prediction is projected back to the observable 4× resolution and compared with the aligned ground-truth region. OSEDiff and CoZ visibly alter the cable and panel boundaries, producing larger residuals. In contrast, Oracle Zoom preserves these structures more closely and yields the smallest projected error (absolute difference), showing that its deeper predictions remain better aligned with the visual evidence available before the supervision boundary.

5 Conclusion

We presented Oracle Zoom, an OPSD-inspired, reference-constrained framework for recursive SR beyond the last target-available scale. Oracle Zoom trains on its own recursive trajectory: direct supervision anchors target-available predictions, aligned projection carries the last observable ground-truth evidence into deeper scales, and a no-reference quality objective guides unresolved detail, while a KL-constrained pretrained prior and EMA consistency limit drift. Across seven datasets, Oracle Zoom achieves SOTA fidelity, no-reference super resolution quality while significantly reducing hallucinations. These gains require only a 7.1M-parameter LoRA trained on 1,000 images in addition to the base SR model. Overall, Oracle Zoom turns recursive SR beyond available ground truth from unconstrained synthesis into generation that remains anchored to the last verifiable visual evidence. This provides a practical path toward reliable extreme magnification even when direct supervision can no longer follow the recursion.

6 Limitations

Our evaluation beyond 4× cannot measure exact recovery because the ground truth is unavailable at deeper scales; projected-reference metrics and the anchored VLM judge instead assess consistency with observable evidence. The current evaluation also follows synthetic center-crop recursion, so it does not directly establish performance for physical camera zoom or arbitrary user-selected regions. Training uses a curated 4K photographic set, and the smaller fidelity gains on DRealSR and RealSR indicate that domain-specific degradations remain challenging. Finally, Oracle Zoom relies on a fixed no-reference quality model and a pretrained SR prior; both can bias the type of detail encouraged at unsupported scales [38]. These limitations matter because outputs may appear plausible even when fine-scale recovery cannot be verified [22]. Future work should extend supervision to degradation models, adaptive zoom paths, and stronger reference-aware evaluation for deeper recursive magnification.

References

[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In IEEE conference on computer vision and pattern recognition workshops, 2017.

[2] Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. In Advances in Neural Information Processing Systems, 2015.

[3] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018.

[4] Walid Bousselham, Hilde Kuehne, and Cordelia Schmid. Vold: Reasoning transfer from llms to vision-language models via on-policy distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026.

[5] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In IEEE/CVF International Conference on Computer Vision, 2019.

[6] Jiezhang Cao, Qin Wang, Yongqin Xian, Yawei Li, Bingbing Ni, Zhiming Pi, Kai Zhang, Yulun Zhang, Radu Timofte, and Luc Van Gool. Ciaosr: Continuous implicit attention-in-attention network for arbitrary-scale image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.

[7] Pu Cao, Feng Zhou, Qing Song, and Lu Yang. Controllable generation with text-to-image diffusion models: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.

[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021.

[9] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024.

[10] Hao-Wei Chen, Yu-Syuan Xu, Min-Fong Hong, Yi-Min Tsai, Hsien-Kai Kuo, and Chun-Yi Lee. Cascaded local implicit transformer for arbitrary-scale super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.

[11] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021.

[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 2021.

[13] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022.

[14] Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature machine intelligence, 5(3):220–235, 2023.

[15] Shubhashis Roy Dipta, Daniel Bis, Kun Zhou, Lichao Wang, Benjamin Z Yao, Chenlei Guo, and Ruhi Sarikaya. Pa3: Policy-aware agent alignment through chain-of-thought. arXiv preprint arXiv:2603.14602, 2026a.

[16] Shubhashis Roy Dipta, Ankur Padia, and Francis Ferraro. Decomposerl: Learning to ask useful, informative, and diverse questions for semi-supervised, traceable claim verification. arXiv preprint arXiv:2605.27858, 2026b.

[17] Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.

[18] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.

[19] Shuhang Gu, Andreas Lugmayr, Martin Danelljan, Manuel Fritsche, Julien Lamour, and Radu Timofte. Div8k: Diverse 8k resolution image dataset. In IEEE/CVF International Conference on Computer Vision Workshop, 2019.

[20] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European conference on computer vision, 2024.

[21] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015.

[22] Elias Hossain, Shubhashis Roy Dipta, Subash Neupane, Rajib Rana, Ravid Shwartz-Ziv, Ivan Garibay, and Niloofar Yousefi. Uat-lite: Inference-time uncertainty-aware attention for pretrained transformers. arXiv preprint arXiv:2602.02952, 2026.

[23] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022.

[24] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019.

[25] Yuxuan Jiang, Runchao Li, Shubhashis Roy Dipta, Dawei Li, and Zhao Yang. Cornerstones or stumbling blocks? deciphering the rock tokens in on-policy distillation. arXiv preprint arXiv:2605.09253, 2026.

[26] Abhinav Joshi, Shaswati Saha, Divyaksh Shukla, Sriram Vema, Harsh Jhamtani, Manas Gaur, and Ashutosh Modi. Towards robust evaluation of unlearning in llms via data transformations. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 12100–12119, 2024.

[27] Sharif Amit Kamran, Sourajit Saha, Ali Shihab Sabbir, and Alireza Tavakkoli. Optic-net: A novel convolutional neural network for diagnosis of retinal diseases from optical tomography images. In 2019 18th IEEE international conference on machine learning and applications (ICMLA), pages 964–971. IEEE, 2019.

[28] Sharif Amit Kamran, Sourajit Saha, Ali Shihab Sabbir, and Alireza Tavakkoli. A comprehensive set of novel residual blocks for deep learning architectures for diagnosis of retinal diseases from optical coherence tomography images. In Deep Learning Applications, Volume 2, pages 25–48. Springer, 2020.

[29] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE/CVF conference on computer vision and pattern recognition, 2019.

[30] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In IEEE/CVF International Conference on Computer Vision, 2021.

[31] Bryan Sangwoo Kim, Jeongsol Kim, and Jong Chul Ye. Chain-of-Zoom: Extreme super-resolution via scale autoregression and preference alignment. In Advances in Neural Information Processing Systems, 2025.

[32] Yoon Kim and Alexander M. Rush. Sequence-level knowledge distillation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1317–1327, 2016.

[33] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017.

[34] Alex M. Lamb, Anirudh Goyal, Ying Zhang, Saizheng Zhang, Aaron C. Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In Advances in Neural Information Processing Systems, 2016.

[35] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.

[36] Pengyu Li, Zhitao Gao, Lingling Zhang, Muye Huang, Yuanming Li, Fangzhi Xu, and Jun Liu. Visual-opsd: Cross-modal on-policy self-distillation for efficient unified multimodal reasoning. arXiv preprint arXiv:2606.18974, 2026a.

[37] Pengyu Li, Zhitao Gao, Lingling Zhang, Muye Huang, Yuanming Li, Fangzhi Xu, and Jun Liu. Visual-opsd: Cross-modal on-policy self-distillation for efficient unified multimodal reasoning. arXiv preprint arXiv:2606.18974, 2026b.

[38] Nusrat Jahan Lia and Shubhashis Roy Dipta. Cross-lingual sentiment misalignment: Auditing multilingual language models for inversion risk, dialectal representation, and affective stability. In Proceedings of the 1st Workshop on Multilinguality in the Era of Large Language Models (MeLLM 2026), San Diego, United States, 2026. Association for Computational Linguistics.

[39] Nusrat Jahan Lia, Shubhashis Roy Dipta, Abdullah Khan Zehady, Naymul Islam, Madhusodan Chakraborty, and Abdullah Al Wasif. Read between the lines: A benchmark for uncovering political bias in Bangla news articles. In Proceedings of the Second Workshop on Bangla Language Processing (BLP-2025), Mumbai, India, 2025. Association for Computational Linguistics.

[40] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In IEEE/CVF international conference on computer vision workshops, 2021.

[41] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2017.

[42] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diffbir: Toward blind image restoration with generative diffusion prior. In Proceedings of the European Conference on Computer Vision, 2024.

[43] Hongyu Liu, Chun Wang, Feng Gao, Xuanhua He, Yue Ma, Ziyu Wan, Yong Zhang, Xiaoming Wei, and Qifeng Chen. Opsd-v: On-policy self-distillation for post-training few-step autoregressive video generators. arXiv preprint arXiv:2607.08766, 2026.

[44] David Lopez-Paz, Léon Bottou, Bernhard Schölkopf, and Vladimir Vapnik. Unifying distillation and privileged information. In International Conference on Learning Representations, 2016.

[45] Aritra Mazumder, Shubhashis Roy Dipta, Nusrat Jahan Lia, Tanzila Khan, Kainat Raisa Hossain, Nehaa Shri, Shubhrangshu Debsarkar, Humayra Tasnim, Gour Gupal Talukder Shawon, Debjoty Mitra, et al. Agentcollabbench: Diagnosing when good agents make bad collaborators. arXiv preprint arXiv:2605.08647, 2026.

[46] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a “completely blind” image quality analyzer. IEEE Signal Processing Letters, 2013.

[47] Zabir Al Nazi, Shubhashis Roy Dipta, and Sudipta Kar. †DAGGER: Distractor-aware graph generation for executable reasoning in math problems. arXiv preprint arXiv:2601.06853, 2026a.

[48] Zabir Al Nazi, Shubhashis Roy Dipta, and Md Rizwan Parvez. Omni-modal dissonance benchmark: Systematically breaking modality consensus to probe robustness and calibrated abstention. arXiv preprint arXiv:2603.27187, 2026b.

[49] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024.

[50] Nishanjan Ravin, Sourajit Saha, Alan Schweitzer, Ameena Elahi, Farouk Dako, Daniel Mollura, and David Chapman. Mitigating domain shift in ai-based tb screening with unsupervised domain adaptation. IEEE Access, 10:45997–46013, 2022.

[51] Stephane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, 2011.

[52] Shubhashis Roy Dipta and Francis Ferraro. If we may de-presuppose: Robustly verifying claims through presupposition-free question decomposition. In Proceedings of the 14th Joint Conference on Lexical and Computational Semantics (*SEM 2025), Suzhou, China, 2025a. Association for Computational Linguistics.

[53] Shubhashis Roy Dipta and Francis Ferraro. Q2E: Query-to-event decomposition for zero-shot multilingual text-to-video retrieval. In Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics, Mumbai, India, 2025b. The Asian Federation of Natural Language Processing and The Association for Computational Linguistics.

[54] Shubhashis Roy Dipta, Mehdi Rezaee, and Francis Ferraro. Semantically-informed hierarchical event modeling. In Proceedings of the 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023), Toronto, Canada, 2023. Association for Computational Linguistics.

[55] Shubhashis Roy Dipta, Tz-Ying Wu, and Subarna Tripathi. VC-inspector: Advancing reference-free evaluation of video captions with factual analysis. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), San Diego, California, United States, 2026. Association for Computational Linguistics.

[56] Andrei A. Rusu, Sergio Gomez Colmenarejo, Caglar Gulcehre, Guillaume Desjardins, James Kirkpatrick, Razvan Pascanu, Volodymyr Mnih, Koray Kavukcuoglu, and Raia Hadsell. Policy distillation. In International Conference on Learning Representations, 2016.

[57] Sourajit Saha and Shubhashis Roy Dipta. Seebel: Seeing is believing. arXiv preprint arXiv:2312.10933, 2023.

[58] Sourajit Saha and Tejas Gokhale. Improving shift invariance in convolutional neural networks with translation invariant polyphase sampling. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 620–629. IEEE, 2025.

[59] Sourajit Saha and Tejas Gokhale. Zero-shot multimodal retrieval with multi-scale contextual representations. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 20304–20324, 2026.

[60] Sourajit Saha and Nisha Saha. A lightning fast approach to classify bangla handwritten characters and numerals using newly structured deep neural network. Procedia computer science, 132:1760–1770, 2018.

[61] Sourajit Saha and Yaacov Yesha. Pairwise meta learning pipeline: classifying covid-19 abnormalities on chest radio-graphs. SPIE Medical Imaging 2022: Computer-Aided Diagnosis; PC1203302 (2022) Proceedings Volume PC12033, Medical Imaging 2022: Computer-Aided Diagnosis; PC1203302 (2022), 2022.

[62] Sourajit Saha, Md Saiful Islam, Md Asif Bin Khaled, and Suraiya Tairin. An efficient traffic sign recognition approach using a novel deep neural network selection architecture. In Emerging Technologies in Data Mining and Information Security: Proceedings of IEMIS 2018, Volume 3, pages 849–862. Springer, 2018a.

[63] Sourajit Saha, Sharif Amit Kamran, and Ali Shihab Sabbir. Total recall: understanding traffic signs using deep convolutional neural network. In 2018 21st international conference of computer and information technology (ICCIT), pages 1–6. IEEE, 2018b.

[64] Shaswati Saha, Lauren Selingo, Emily Olejniczak, Hanna Noyce, Vaskar Raychoudhury, Roger O Smith, and Md Osman Gani. Mypath: Accessible routing for wheelchair users. Rehabilitation Engineering and Assistive Technology Society of North America (RESNA), 2022.

[65] Shaswati Saha, Sourajit Saha, Manas Gaur, and Tejas Gokhale. Side effects of erasing concepts from diffusion models. arXiv preprint arXiv:2508.15124, 2025.

[66] Shaswati Saha, Rajasekhar Anguluri, and Manas Gaur. To erase, or not to erase: Robust training-free concept erasure with preservation aware adaptive ranked subspace expansion. arXiv preprint arXiv:2607.23492, 2026.

[67] Nobin Sarwar. Fedmentalcare: towards privacy-preserving fine-tuned llms to analyze mental health status using federated learning framework. arXiv preprint arXiv:2503.05786, 2025a.

[68] Nobin Sarwar. Filterrag: zero-shot informed retrieval-augmented generation to mitigate hallucinations in vqa. arXiv preprint arXiv:2502.18536, 2025b.

[69] Nobin Sarwar and Shubhashis Roy Dipta. Fedmentor: Domain-aware differential privacy for heterogeneous federated llms in mental health. arXiv preprint arXiv:2509.14275, 2025.

[70] Nobin Sarwar, Shubhashis Roy Dipta, Zheyuan Liu, and Vaidehi Patil. Multimodal unlearning across vision, language, video, and audio: Survey of methods, datasets, and benchmarks. In Findings of the Association for Computational Linguistics: ACL 2026, pages 27702–27730, 2026.

[71] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024.

[72] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in Neural Information Processing Systems, 2017.

[73] Jianyi Wang, Kelvin C. K. Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023a.

[74] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C. K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 2024a.

[75] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021a.

[76] Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al. Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265, 2025.

[77] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European Conference on Computer Vision Workshops, 2018.

[78] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2021b.

[79] Xiaohang Wang, Xuanhong Chen, Bingbing Ni, Hang Wang, Zhengyan Tong, and Yutian Liu. Deep arbitrary-scale image super-resolution via scale-equivariance pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023b.

[80] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: Diffusion-based image super-resolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024b.

[81] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In European conference on computer vision, 2020.

[82] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. In Advances in Neural Information Processing Systems, 2024a.

[83] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. SeeSR: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024b.

[84] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V. Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020.

[85] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniqa: Multi-dimension attention network for no-reference image quality assessment. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2022.

[86] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In Proceedings of the European Conference on Computer Vision, 2024.

[87] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024.

[88] Qianhao Yuan, Jie Lou, Xing Yu, Hongyu Lin, Le Sun, Xianpei Han, and Yaojie Lu. Vision-opd: Learning to see fine details for multimodal llms via on-policy self-distillation. arXiv preprint arXiv:2605.18740, 2026a.

[89] Qianhao Yuan, Jie Lou, Xing Yu, Hongyu Lin, Le Sun, Xianpei Han, and Yaojie Lu. Vision-opd: Learning to see fine details for multimodal llms via on-policy self-distillation. arXiv preprint arXiv:2605.18740, 2026b.

[90] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift: Efficient diffusion model for image super-resolution by residual shifting. In Advances in Neural Information Processing Systems, 2023.

[91] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021.

[92] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018.

[93] Xiang Zhang, Yulun Zhang, and Fisher Yu. Hit-sr: Hierarchical transformer for efficient image super-resolution. In European conference on computer vision, 2024.

[94] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in neural information processing systems, 2023.

[95] Siyan Zhao, Zhihui Xie, Mengchen Liu, Jing Huang, Guan Pang, Feiyu Chen, and Aditya Grover. Self-distilled reasoner: On-policy self-distillation for large language models. In Forty-third International Conference on Machine Learning, 2026.

[96] Wei Zhou, Xiongwei Zhu, Lingdong Kong, Bo Chen, Lei Zhang, Yongyuan Liang, Xiaoxia Hou, Ye Tian, Xian Sun, Yingshuo Wang, Linfeng Li, Shengqiong Wu, Leigang Qu, Feng Li, Wei Liu, Julian McAuley, and Tat-Seng Chua. On-policy self-distillation in diffusion models. arXiv preprint arXiv:2608.24646, 2026a.

[97] Wei Zhou, Xiongwei Zhu, Lingdong Kong, Bo Chen, Lei Zhang, Yongyuan Liang, Xiaoxia Hou, Ye Tian, Xian Sun, Yingshuo Wang, Linfeng Li, Shengqiong Wu, Leigang Qu, Feng Li, Wei Liu, Julian McAuley, and Tat-Seng Chua. On-policy self-distillation in diffusion models. arXiv preprint arXiv:2608.24646, 2026b.

[98] Zihao Zhu, Kuan-Ru Huang, Zhaoming Xu, Renjie Li, Bo Wu, Ruizheng Bai, Mingyang Wu, Sayak Paul, and Zhengzhong Tu. 4klsdb: A large-scale dataset for 4k image restoration and generation. arXiv preprint arXiv:2605.24762, 2026.

Appendix

Supplementary overview. The supplementary material provides additional method analysis, implementation and evaluation details, and extended quantitative and qualitative results and analysis.

Appendix A Additional Method Details and Analysis

This section provides additional details and analysis for Oracle Zoom. We first show how the last available ground truth remains useful at deeper recursive scales and clarify what information can still be verified. We then derive the KL-constrained latent prior, provide the full proof of Proposition 1, and explain how supervision propagates through the recursive chain. Finally, we relate Oracle Zoom to prior self-distillation approaches and clarify the distinction in supervision.

A.1 Extending Target Evidence to Deeper Zooms

Recall that Imgt is the last available ground-truth target, while no target Ijgt exists for j>m. Although Imgt is limited to scale sm, it still contains the image region followed by every subsequent deterministic zoom. The available target can therefore remain useful after direct supervision ends. For a deeper scale sj, define the cumulative relative magnification from sm as

Γm→j=sjsm=∏r=m+1jγr. (14)

The recursive zoom path from sm to sj is

𝒜m→j=𝒵γj∘⋯∘𝒵γm+1, (15)

where each operator follows the same deterministic spatial selection used during recursive SR. Applying this path to the last available target gives

Im→jref=𝒜m→j​(Imgt). (16)

The prediction I^j represents the same field of view at scale sj. We project it back to the resolution of sm:

𝒫j→m=Downsamplem→j,I^j→m=𝒫j→m​(I^j). (17)

Consequently, Im→jref and I^j→m describe the same spatial region at the same observable resolution.

Observation 1 (Persistent visual reference). For every j>m, the deterministic recursive zoom path identifies a region Im→jref within Imgt that can be compared with 𝒫j→m​(I^j) without requiring Ijgt.

Justification. Each 𝒵γr selects the region enlarged at the next recursive step. Composing these operators therefore identifies the nested region represented at scale sj. Projecting I^j back to the resolution of sm allows this region to be compared with the corresponding region in Imgt. Thus, the last available ground truth continues to provide visual evidence beyond sm, although it cannot verify newly generated fine-scale detail.

A.2 Limits of cross-scale supervision

Cross-scale supervision preserves the information that can still be verified from Imgt, but it cannot determine all fine-scale detail at sj. Multiple high-resolution images can produce the same image after projection. Consider

𝒫j→m:𝒳j→𝒳m, (18)

where 𝒳j denotes images at scale sj. Since sj>sm, the projection removes fine-scale information. Therefore, two different images X1≠X2 can satisfy

𝒫j→m​(X1)=𝒫j→m​(X2). (19)

For a linear projection, this ambiguity can be written directly. Any fine-scale change δ removed by the projection satisfies

𝒫j→m​(δ)=0, (20)

and therefore

𝒫j→m​(X+δ)=𝒫j→m​(X). (21)

Observation 2 (Unobservable fine-scale detail). Minimizing ℒxscale constrains the information that remains visible after projection, but not the fine-scale detail removed by it.

This motivates the three objectives used beyond the supervision boundary. ℒxscale preserves the evidence that can still be verified from Imgt. ℒqual guides fine-scale detail that this evidence cannot supervise. ℒprior keeps the resulting prediction close to the pretrained SR model. Together, these objectives provide complementary supervision at scales without direct ground truth.

A.3 Derivation of the KL-Constrained Latent Prior

We derive the form of ℒprior. For a target-unavailable scale j, let

ρjθ=𝒩⁡(zjθ,σ2​I),ρj0=𝒩⁡(zj0,σ2​I), (22)

where zjθ and zj0 are the adapted and base-model latent representations, respectively.

For two d-dimensional Gaussians 𝒩⁡(μ1,Σ1) and 𝒩⁡(μ0,Σ0),

DKL(𝒩(μ1,Σ1)∥𝒩(μ0,Σ0)) (23)

=12[tr(Σ0−1Σ1)+(μ0−μ1)⊤Σ0−1(μ0−μ1)

−d+logdetΣ0detΣ1].

Under the shared isotropic covariance

Σ0=Σ1=σ2​I, (24)

the trace and dimensionality terms cancel, while the log-determinant ratio is zero. Therefore,

DKL(ρjθ∥ρj0)=12​σ2‖zjθ−zj0‖22. (25)

Using

MSE⁡(zjθ,zj0)=1d​‖zjθ−zj0‖22, (26)

we obtain

DKL(ρjθ∥ρj0)=d2​σ2MSE(zjθ,zj0)

(27)

Thus, under the shared isotropic Gaussian assumption, latent MSE is equivalent to the closed-form KL divergence up to the constant d/(2​σ2). Since this constant is independent of θ, it can be incorporated into λprior. The latent prior therefore keeps the adapted model close to the pretrained SR model while the quality objective guides detail at scales without ground truth.

A.4 Proof of Bounded Quality Deviation

Proposition 1. Let

q⁡(z)=𝒬⁡(D⁡(z)) (28)

be locally Lq-Lipschitz around zj0 with respect to

dz​(z,z′)=MSE⁡(z,z′). (29)

If

DKL(ρjθ∥ρj0)≤ϵprior, (30)

then

|𝒬⁡(I^j)−𝒬⁡(D⁡(zj0))|≤Lq​σ​2​ϵpriord. (31)

Proof. From Eq. (27),

DKL(ρjθ∥ρj0)=d2​σ2MSE(zjθ,zj0). (32)

Combining Eqs. (30) and (32) gives

MSE⁡(zjθ,zj0)≤2​σ2​ϵpriord. (33)

Taking the square root yields

dz​(zjθ,zj0)≤σ​2​ϵpriord. (34)

By the local Lq-Lipschitz assumption,

|q⁡(zjθ)−q⁡(zj0)|≤Lq​dz​(zjθ,zj0). (35)

Substituting Eq. (34) gives

|q⁡(zjθ)−q⁡(zj0)|≤Lq​σ​2​ϵpriord

(36)

Finally, since q⁡(zjθ)=𝒬⁡(I^j) and q⁡(zj0)=𝒬⁡(D⁡(zj0)), Eq. (36) is exactly Eq. (31). □

The proposition shows that a bounded KL constraint limits how far the adapted prediction can move from the pretrained SR model in quality-score space. Cross-scale consistency complements this constraint by preserving the visual evidence that can still be verified.

A.5 From Constraints to Training Objective

The overall objective can also be written as a constrained optimization problem. Omitting the auxiliary EMA term, we seek

maxθ 1K−m​∑j=m+1K𝒬⁡(I^j) (37)

s.t. ℒsup≤ϵsup,

ℒxscale≤ϵxscale,

ℒprior≤ϵprior.

Introducing non-negative multipliers αsup, αxscale, and αprior gives

𝒥⁡(θ)= −1K−m∑j=m+1K𝒬(I^j) (38)

+αsup​(ℒsup−ϵsup)

+αxscale​(ℒxscale−ϵxscale)

+αprior​(ℒprior−ϵprior).

The terms involving ϵsup, ϵxscale, and ϵprior are constant with respect to θ and do not affect gradient-based optimization. Rescaling the remaining multipliers and adding EMA consistency gives the training objective used in the main paper:

ℒ= ℒsup+λxscale​ℒxscale+λqual​ℒqual (39)

+λprior​ℒprior+λema​ℒema.

This formulation makes the role of each term explicit: the quality objective is optimized while target fidelity, cross-scale evidence, and proximity to the pretrained SR model are preserved.

A.6 Gradient Flow Through Recursive SR

Recursive SR uses each prediction as the input to the next scale. We therefore retain the computation graph through successive predictions rather than detaching them. Let the differentiable image path at step i be written as

I^i=Tθ(i)​(I^i−1), (40)

where Tθ(i) includes the zoom operation, adapted SR model, and frozen decoder for the conditioning used at that step. For a loss ℒj applied at a deeper scale j, differentiation through the recurrence gives

d​I^jd​θ=∂Tθ(j)∂θ+∂Tθ(j)∂I^j−1​d​I^j−1d​θ. (41)

Expanding recursively shows that

d​ℒjd​θ=∂ℒj∂I^j​d​I^jd​θ (42)

contains gradients through the preceding predictions that contribute to I^j.

Consequently, objectives evaluated beyond sm do not only update the final deep-scale prediction. They also update earlier predictions that later become recursive inputs. This exposes the shared SR transition to the same model-generated inputs it receives during inference and allows deeper objectives to shape the complete recursive chain.

A.7 EMA Consistency at the Supervision Boundary

The transition from sm to sm+1 is where direct supervision ends: sm still has exact ground truth, while sm+1 does not. We therefore apply EMA consistency at this boundary.

The EMA parameters are updated as

θ¯t=μ​θ¯t−1+(1−μ)​θt, (43)

and are not updated by gradients. Since θ¯ averages the model parameters over training, it changes more slowly than the current parameters θ. We use it to form a stable latent reference

zmema=sg⁡[Fθ¯​(Imgt,pm)], (44)

and encourage the current representation to remain close to this reference using

ℒema=MSE⁡(zmθ,zmema). (45)

The EMA term therefore stabilizes training where direct supervision ends. It complements ℒsup: while ℒsup matches the prediction to the available ground truth, ℒema keeps the latent representation stable before recursion proceeds to scales without targets.

A.8 Training-Only Supervision and Inference

The additional supervision mechanisms are used only during training. Specifically, Iigt, the aligned cross-scale reference, 𝒬, F0, and the EMA branch are required only to compute the objectives in Equation 39. At inference, only the trained recursive SR transition is used.

Each recursive step is

Iiin=𝒵γi​(I^i−1), (46)

pi=G⁡(I^i−1,Iiin), (47)

and

I^i=D⁡(Fθ​(Iiin,pi)). (48)

Thus, Oracle Zoom uses the same recursive inference pipeline as the base SR system. The additional signals are used only during training and introduce no new inference-time supervision.

A.9 Putting the Objectives Together

The five objectives play different roles depending on whether ground truth is available. For i≤m, ℒsup directly matches the prediction to the observed target. For j>m, ℒxscale carries forward the information from Imgt that can still be verified along the recursive zoom path. Since this evidence does not determine all fine-scale detail, ℒqual guides the remaining detail while ℒprior keeps the prediction close to the pretrained SR model. Finally, ℒema stabilizes the transition where direct supervision ends.

Together, these objectives allow training to continue beyond the last available ground-truth scale without discarding the evidence that remains available. Oracle Zoom preserves what can still be verified, guides the remaining fine-scale detail, and constrains that detail with the pretrained SR prior.

A.10 Relation to On-Policy Self-Distillation

Oracle Zoom is inspired by On-Policy Self-Distillation (OPSD) [95], but differs in how supervision is provided beyond the ground-truth boundary. Classical knowledge distillation transfers outputs from a reference model to another model [21, 56, 32], while privileged-information methods use additional training-only information that is unavailable at inference [44]. Related work also reduces the mismatch between training and inference by learning from model-generated intermediate predictions or targets [51, 2, 34, 84]. OPSD trains on model-generated sequences while using additional training-only information to supervise those sequences [95]. Recent extensions apply related self-distillation strategies to vision-language reasoning, multimodal reasoning, fine-detail perception, autoregressive video generation, and diffusion models [4, 37, 89, 43, 97, 25].

The connection to Oracle Zoom is limited to training on model-generated intermediate predictions. In recursive SR, each prediction becomes the input to the next scale, but beyond sm no high-resolution target exists for the deeper prediction. Oracle Zoom therefore uses the last available ground truth as a cross-scale reference for what remains verifiable, while ℒqual guides unresolved fine-scale detail and ℒprior limits deviation from the pretrained SR model. Unlike standard distillation or self-training [21, 32, 84], F0 is not used to construct a replacement supervision target, and 𝒬 is used only to evaluate the current prediction during training. While OPSD trains on model-generated intermediate predictions, Oracle Zoom introduces the reference-constrained supervision when recursive SR continues beyond the last available ground-truth scale.

Appendix B Implementation Details

媒体内容 · 前往原文查看

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)

Rank r 16

α (scaling) 32 (=2​r)

Dropout 0.0

Target modules to_q,to_k,to_v,add_q_proj, add_k_proj,add_v_proj

Trainable parameters 7.1M

Optimization

Optimizer AdamW (β1=0.9, β2=0.999, ϵ=10−8)

Learning rate 5×10−5

Weight decay 10−2

LR schedule linear warmup (500 steps) then constant

Effective batch size 4 (accumulated one image at a time)

Max epochs 200 (cap)

Early stopping patience 8 evaluations, min-delta 10−3, eval every 100 steps

Seed 0

Training objectives

4× supervision λsup 1.0

Cross-scale consistency λxscale 1.0

Quality objective λqual 0.4

Latent prior λprior 8.0

EMA consistency λema 0.1

EMA decay 0.95

Quality model TOPIQ-NR

Latent reference base SR model (adapter off)

Data & protocol

Training set 1,000-image curated tier

Validation set held-out set, 2,000 images

Supervision scale 4× only (GT)

Training recursion 4×→16×

Crop size 512×512

Recursions at test 4 (4×/16×/64×/256×)

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.

媒体内容 · 前往原文查看

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

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 s.

媒体内容 · 前往原文查看

Scale ours CoZ tie abst. win-rate [95% CI]

4× 29 29 54 8 0.50 [0.38,0.63]

16× 60 53 6 1 0.53 [0.44,0.62]

64× 77 37 6 0 0.68 [0.59,0.75]

256× 91 25 4 0 0.78 [0.70,0.85]

Table 6: Pairwise faithfulness evaluation across recursion depth. Oracle Zoom and CoZ are comparable at 4× and 16×, while Oracle Zoom is preferred at 64× and 256×. Win rates use decided comparisons only; brackets show Wilson 95% confidence intervals.

Training configuration.

The LoRA adapter is applied to the to_q, to_k, to_v, add_q_proj, add_k_proj, and add_v_proj layers of the SD3 transformer. We set the LoRA scaling to α=32 with zero dropout and limit the VLM prompter to 32 generated tokens. AdamW uses (β1,β2)=(0.9,0.999) and ϵ=10−8, with a constant learning rate after warmup. We evaluate every 100 optimization steps and use early stopping with patience 8 and a minimum improvement of 10−3. Training is capped at 200 epochs, uses random seed 0, and accumulates the effective batch one image at a time. Table 4 summarizes the complete training configuration.

Compute and latency.

Training is performed on a single 8-GPU H100-class node, with one GPU used for optimization and the remaining GPUs parallelizing validation. The reported model finishes training in under one day. For latency, we evaluate 500 DIV2K images on one H100 with batch size 1, CUDA synchronization around each stage, and three discarded warm-up samples. As shown in Table 5, Oracle Zoom requires 1.505 s per recursive step compared with 1.486 s for CoZ. The additional cost comes entirely from the LoRA-adapted SR pass, while prompt extraction remains unchanged. Since each recursive input is resampled to 512×512, the per-step cost remains approximately constant across 4×–256×. Merging the linear LoRA weights into the backbone reduces the SR latency to within 0.001 s of CoZ.

媒体内容 · 前往原文查看

Supervision LPIPS ↓ DISTS ↓ DINOv2 ↑

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

Table 7: Where to apply target-available supervision at 4×. Decode-space LPIPS improves LPIPS, DISTS, and DINOv2 similarity over CoZ, whereas direct latent-space MSE degrades fidelity.

媒体内容 · 前往原文查看

Reference LPIPS ↓ DISTS ↓ DINOv2 ↑ FID ↓ TOPIQ-FR ↑

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

Table 8: Effect of reference quality at 4×. Exact ground truth gives the strongest fidelity, while a mildly degraded reference retains most of the improvement over CoZ.

Judge protocol.

For the pairwise faithfulness evaluation [55], each item contains an anchor and two same-region candidate zooms, with candidate order randomized independently for every item [39]. Each image carries an explicit ANCHOR, A, or B label, and the judge first reports these labels before returning its comparison. At 4×, the anchor is the ground-truth patch, whereas from 16× onward the preceding zoom serves as the observable reference. The output contains a comparison in {A,B,tie} together with separate hallucination indicators for the two candidates. Unparseable outputs are recorded as abstentions, and win rates are computed only over decided comparisons [48]. Table 6 reports the raw counts and Wilson 95% confidence intervals. Figures 7 and 8 reproduce the exact rubrics used at 4× and at deeper scales, respectively [52].

媒体内容 · 前往原文查看

Figure 7: The rubric used for pairwise judging at 4×, with the target high-resolution image as the anchor.

媒体内容 · 前往原文查看

Figure 8: The rubric used for pairwise judging at 16× and deeper, with the previous zoom level as the anchor.

Figure 9: Effect of the no-reference quality objective across 4–256×. 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.

媒体内容 · 前往原文查看

No-reference quality Fidelity @4× (GT) Coherence Quality score Hallucination

Quality model CLIPIQA↑ CQ256↑ LPIPS↓ DISTS↓ DINO4×↑ Coher.↑ TOPIQ256↑ Halluc.↓

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

Table 9: Choice of no-reference quality model. TOPIQ-NR gives the strongest overall and 256× no-reference quality, while all three choices give similar 4× fidelity. HyperIQA slightly improves coherence, whereas Laplacian sharpness substantially reduces deep-zoom quality. Hallucination differences between the three variants are not significant.

媒体内容 · 前往原文查看

4KLSDB DIV8K DRealSR RealSR

Method LPIPS↓ DISTS↓ DINO↑ LPIPS↓ DISTS↓ DINO↑ LPIPS↓ DISTS↓ DINO↑ LPIPS↓ DISTS↓ DINO↑

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

CoZ [31] 0.248 0.186 0.887 0.238 0.166 0.910 0.183 0.158 0.900 0.193 0.170 0.860

Ours 0.169 0.142 0.925 0.223 0.161 0.930 0.182 0.161 0.917 0.220 0.176 0.890

Table 10: Full-reference fidelity at 4× 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.

Appendix C Additional Results and Analysis

C.1 Analysis

We further examine the target-available supervision, reference quality, and choice of no-reference quality model.

Target-available supervision.

Table 7 compares supervision in latent and image space at 4×. Latent MSE performs worse than CoZ on all three fidelity metrics, whereas decode-space LPIPS substantially improves LPIPS, DISTS, and DINOv2 similarity. This supports applying ℒsup after decoding rather than directly matching latent features.

Reference quality.

Table 8 examines how the quality of the available reference affects supervision. The exact ground-truth reference performs best. A mildly degraded reference remains close and still improves substantially over CoZ. Combining the exact and degraded references is slightly weaker than using the exact target alone.

Quality model.

Table 9 compares TOPIQ-NR with HyperIQA and a Laplacian sharpness measure. TOPIQ-NR gives the highest mean CLIPIQA and CLIPIQA at 256×, while all three choices give similar 4× fidelity. HyperIQA slightly improves coherence but reduces mean CLIPIQA by 0.031 and 256× quality by 0.027. The Laplacian variant reduces these quantities by 0.100 and 0.128, respectively. Hallucination differences between the three variants are not significant. Figure 9 shows the corresponding qualitative behavior, with the Laplacian variant losing fine structure and HyperIQA producing less realistic textures in these examples. Despite its slightly better coherence in Table 9, HyperIQA adds repetitive textures across the zoom scales shown in Figure 9, whereas Oracle Zoom does not.

Per-dataset fidelity.

Table 10 expands the 4× fidelity evaluation to all four datasets with high-resolution targets. Oracle Zoom gives the strongest LPIPS and DISTS on 4KLSDB and the strongest LPIPS, DISTS, and DINOv2 similarity on DIV8K. Results are more mixed on DRealSR and RealSR. The largest fidelity gains occur on the high-resolution photographic datasets.

C.2 Results

We complement the aggregate results in the main paper with a per-dataset quantitative breakdown and qualitative recursive zoom examples across all seven test sets.

Quantitative Results across datasets.

Tables 11, 12, 13, 14, 15, 16 and 17 report the complete per-scale no-reference results on all seven test sets. Oracle Zoom achieves the highest CLIPIQA at every scale on every dataset. At 4×, differences in MUSIQ and MANIQA are smaller, and these metrics sometimes favor another method. This is consistent with the target-available objective, which emphasizes agreement with ground truth in stead of no-reference quality alone. The separation becomes clearer once recursion proceeds beyond the target-available scale.

On 4KLSDB (Table 11), Oracle Zoom improves CLIPIQA from 0.638 at 4× to 0.658 at 16× and remains above CoZ through 256×, while also leading MUSIQ and MANIQA from 16× onward. On DIV2K (Table 12), the advantage generalizes outside the training domain: Oracle Zoom reaches 0.737 CLIPIQA at 16× and remains strongest through 256×. A similar pattern appears on DIV8K (Table 13), where Oracle Zoom leads CLIPIQA at all four scales and MUSIQ and MANIQA at every target-unavailable scale. These results show that the improvement is not limited to the 4KLSDB training domain.

The real-camera datasets show the same deeper-scale trend. On DRealSR (Table 14), Oracle Zoom rises from 0.704 CLIPIQA at 4× to 0.739 at 16× and remains ahead of CoZ at deeper scales. On RealSR (Table 17), where the earlier-scale differences are smaller, the gap increases with recursion and reaches 0.667 versus 0.540 CLIPIQA at 256×. This shows that the benefit of Oracle Zoom persists under real-camera degradations.

The same behavior extends to different image content. On FFHQ (Table 15), Oracle Zoom gives the largest 256× CLIPIQA margin over CoZ, reaching 0.771 versus 0.579, while preserving the strongest MUSIQ and MANIQA at deeper scales. On Flickr2K (Table 16), Oracle Zoom again leads CLIPIQA at every scale and leads the other learned no-reference metrics from 16× onward, except MANIQA at 16× by only 0.001. Thus, the gains hold across faces, natural scenes, high-resolution photographs, and real-camera images.

The differences are most pronounced at 64× and 256×, after several recursive SR steps. At 256×, the CLIPIQA margin over CoZ ranges from 0.102 on DIV8K to 0.192 on FFHQ and exceeds 0.10 on every dataset. This is substantially larger than the separation near the supervision boundary, showing that the benefit becomes more visible as predictions are repeatedly reused as inputs. NIQE is less consistent across methods, while CLIPIQA, MUSIQ, and MANIQA consistently favor Oracle Zoom at deeper scales. Overall, the seven per-dataset tables support the same conclusion: reference-constrained training becomes increasingly useful for recursive SR beyond supervision.

Qualitative Results across datasets.

Figures 10, 11, 12, 13, 14, 15 and 16 show complete 4×→16×→64×→256× recursive zooms across all seven datasets. A consistent failure pattern emerges as recursion deepens. OSEDiff often replaces the original material with regular synthetic textures before becoming increasingly flat, while CoZ more often preserves the coarse region but gradually loses its material structure.

On 4KLSDB (Figure 10), the zoom follows a waffle-knit garment. OSEDiff flattens the fabric and introduces a regular grid, while CoZ stretches the folds into striations that no longer resemble the original knit. Oracle Zoom better preserves the fold relief and interlocking knit structure through the deeper zooms. On DIV2K (Figure 11), OSEDiff turns the hanging market cloth into smooth regions and regular dots, while CoZ approaches a nearly uniform gradient. Oracle Zoom retains more of the visible weave, fiber structure, and seam boundaries through 256×. On DIV8K (Figure 12), the carved folds of a sandstone statue expose a different failure mode: OSEDiff develops mesh-like and brick-like patterns, while CoZ turns the surface increasingly smooth and waxy. Oracle Zoom maintains and sharpens both the carved relief and irregular stone grain.

The real-camera examples show a similar trend. On DRealSR (Figure 13), both baselines progressively lose the granular surface of an unglazed stoneware planter and approach a nearly uniform field. Oracle Zoom retains substantially more irregular surface variation and directional grain at the deepest scale. On RealSR (Figure 16), the methods remain closer at the earlier zooms, making this a more difficult example. At 256×, however, OSEDiff and CoZ become largely flat, while Oracle Zoom still preserves visible garment structure and fine fabric variation.

The pattern also extends beyond rigid materials. On FFHQ (Figure 14), OSEDiff introduces a regular canvas-like texture and color smearing, while CoZ produces increasingly smooth skin with isolated artificial structures. Oracle Zoom preserves more connected local skin variation and pore-like detail as magnification increases. On Flickr2K (Figure 15), OSEDiff first introduces repeated texture and later becomes smooth, while CoZ progressively fragments the dried seed pods. Oracle Zoom better preserves the pod boundaries, curled husk structure, and nearby leaf blades throughout the recursive chain.

Across these examples, the differences are most visible at 64× and 256×, where earlier prediction errors have passed through several recursive steps. The qualitative results also show that the benefit is not limited to one material type: it appears on fabric, carved stone, pottery, skin, vegetation, and real-camera clothing images. These examples do not establish exact recovery of unseen high-frequency detail, since ground truth is unavailable at the deeper scales. Instead, they show whether newly generated detail remains compatible with visual structures already present in the preceding zooms. Overall, the qualitative evidence matches the quantitative trend: the methods are relatively close near the supervision boundary, while the advantage of Oracle Zoom becomes increasingly visible as recursion proceeds deeper with higher magnifications or bigger zooms.

媒体内容 · 前往原文查看

Scale Method NIQE↓ MUSIQ↑ MANIQA↑ CLIPIQA↑

4× 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

Ours 6.00 59.8 0.556 0.638

16× HiT-SR [93] 14.13 19.7 0.333 0.370

MambaIR [20] 14.42 20.4 0.333 0.389

SwinIR [40] 7.34 28.4 0.427 0.454

SeeSR [83] 7.81 47.8 0.507 0.552

OSEDiff [82] 6.10 51.9 0.531 0.586

CoZ [31] 8.08 52.3 0.551 0.578

Ours 7.16 54.8 0.556 0.658

64× HiT-SR [93] 16.69 23.8 0.388 0.455

MambaIR [20] 17.82 23.3 0.388 0.469

SwinIR [40] 9.02 26.5 0.489 0.477

SeeSR [83] 9.83 37.3 0.495 0.497

OSEDiff [82] 7.20 46.2 0.531 0.533

CoZ [31] 9.54 45.7 0.555 0.544

Ours 8.47 50.8 0.579 0.651

256× HiT-SR [93] 17.81 26.2 0.428 0.494

MambaIR [20] 18.97 26.0 0.426 0.509

SwinIR [40] 10.77 28.5 0.495 0.467

SeeSR [83] 11.56 33.7 0.503 0.496

OSEDiff [82] 8.18 42.6 0.525 0.502

CoZ [31] 10.41 44.1 0.555 0.538

Ours 9.31 49.3 0.589 0.664

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

Figure 10: Qualitative results on 4KLSDB across 4–256×. 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.

媒体内容 · 前往原文查看

Scale Method NIQE↓ MUSIQ↑ MANIQA↑ CLIPIQA↑

4× 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

Ours 4.46 66.5 0.638 0.735

16× HiT-SR [93] 13.11 18.4 0.310 0.312

MambaIR [20] 13.28 19.2 0.312 0.329

SwinIR [40] 6.78 29.0 0.391 0.401

SeeSR [83] 6.50 51.8 0.523 0.540

OSEDiff [82] 5.96 53.0 0.545 0.605

CoZ [31] 6.29 58.5 0.595 0.652

Ours 5.82 61.6 0.599 0.737

64× HiT-SR [93] 16.52 21.8 0.367 0.408

MambaIR [20] 17.10 21.6 0.366 0.419

SwinIR [40] 8.09 23.2 0.478 0.465

SeeSR [83] 8.98 42.0 0.502 0.517

OSEDiff [82] 7.30 47.7 0.536 0.576

CoZ [31] 7.86 52.1 0.581 0.623

Ours 7.12 56.0 0.597 0.727

256× HiT-SR [93] 17.64 26.3 0.421 0.480

MambaIR [20] 18.47 26.2 0.418 0.493

SwinIR [40] 9.96 27.4 0.504 0.466

SeeSR [83] 11.16 36.3 0.502 0.497

OSEDiff [82] 8.53 43.7 0.528 0.543

CoZ [31] 9.24 48.3 0.576 0.599

Ours 8.30 51.8 0.595 0.703

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

Figure 11: Qualitative results on DIV2K across 4–256×. 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.

媒体内容 · 前往原文查看

Scale Method NIQE↓ MUSIQ↑ MANIQA↑ CLIPIQA↑

4× 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

Ours 4.48 67.3 0.648 0.736

16× HiT-SR [93] 13.27 19.0 0.314 0.318

MambaIR [20] 13.42 19.8 0.315 0.334

SwinIR [40] 7.21 29.2 0.388 0.391

SeeSR [83] 6.49 53.0 0.528 0.535

OSEDiff [82] 5.79 53.5 0.550 0.611

CoZ [31] 6.32 58.9 0.602 0.663

Ours 5.85 61.6 0.605 0.735

64× HiT-SR [93] 16.55 22.0 0.370 0.408

MambaIR [20] 17.09 21.8 0.369 0.418

SwinIR [40] 8.62 22.9 0.474 0.460

SeeSR [83] 8.89 43.8 0.509 0.524

OSEDiff [82] 7.04 47.7 0.537 0.580

CoZ [31] 7.78 52.0 0.584 0.630

Ours 7.23 55.6 0.597 0.726

256× HiT-SR [93] 17.52 26.3 0.419 0.487

MambaIR [20] 18.43 26.4 0.417 0.497

SwinIR [40] 10.01 27.3 0.504 0.473

SeeSR [83] 10.79 37.6 0.505 0.510

OSEDiff [82] 8.25 44.0 0.528 0.551

CoZ [31] 8.83 48.6 0.577 0.608

Ours 8.07 51.8 0.593 0.710

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

Figure 12: Qualitative results on DIV8K across 4–256×. 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.

媒体内容 · 前往原文查看

Scale Method NIQE↓ MUSIQ↑ MANIQA↑ CLIPIQA↑

4× 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

Ours 5.71 67.1 0.642 0.704

16× HiT-SR [93] 13.37 19.5 0.321 0.354

MambaIR [20] 13.73 20.6 0.322 0.375

SwinIR [40] 7.37 35.9 0.398 0.435

SeeSR [83] 7.34 57.2 0.540 0.583

OSEDiff [82] 6.57 52.3 0.523 0.608

CoZ [31] 7.28 59.2 0.574 0.649

Ours 6.63 61.9 0.580 0.739

64× HiT-SR [93] 16.60 22.4 0.359 0.420

MambaIR [20] 17.20 22.0 0.359 0.432

SwinIR [40] 8.43 26.4 0.466 0.462

SeeSR [83] 9.31 48.0 0.514 0.553

OSEDiff [82] 7.99 46.0 0.525 0.582

CoZ [31] 8.22 53.1 0.563 0.634

Ours 7.29 55.7 0.580 0.735

256× HiT-SR [93] 17.95 25.1 0.411 0.477

MambaIR [20] 18.78 24.8 0.409 0.492

SwinIR [40] 10.51 27.1 0.487 0.459

SeeSR [83] 11.94 36.6 0.490 0.505

OSEDiff [82] 9.03 41.8 0.522 0.542

CoZ [31] 9.27 48.1 0.572 0.590

Ours 8.04 52.8 0.595 0.709

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×. 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×; colored boxes mark the region enlarged at the next step.

媒体内容 · 前往原文查看

Scale Method NIQE↓ MUSIQ↑ MANIQA↑ CLIPIQA↑

4× 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

Ours 5.48 66.8 0.611 0.777

16× HiT-SR [93] 15.95 22.2 0.327 0.358

MambaIR [20] 16.11 21.7 0.325 0.362

SwinIR [40] 6.93 27.8 0.482 0.478

SeeSR [83] 7.76 43.8 0.509 0.523

OSEDiff [82] 6.40 50.4 0.556 0.536

CoZ [31] 7.46 47.3 0.561 0.541

Ours 7.01 51.2 0.584 0.692

64× HiT-SR [93] 17.93 29.2 0.393 0.508

MambaIR [20] 18.60 28.3 0.390 0.518

SwinIR [40] 8.16 35.1 0.514 0.470

SeeSR [83] 10.59 39.5 0.506 0.514

OSEDiff [82] 6.92 53.0 0.554 0.516

CoZ [31] 7.78 50.2 0.572 0.554

Ours 7.74 54.7 0.606 0.736

256× HiT-SR [93] 17.85 29.9 0.438 0.523

MambaIR [20] 19.13 29.8 0.433 0.530

SwinIR [40] 9.20 34.5 0.513 0.467

SeeSR [83] 12.66 38.6 0.506 0.541

OSEDiff [82] 7.35 52.4 0.552 0.514

CoZ [31] 8.10 51.4 0.585 0.579

Ours 8.78 56.0 0.617 0.771

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× through 256×. Best is bold; second-best is underlined.

Figure 14: Qualitative results on FFHQ across 4–256×. 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.

媒体内容 · 前往原文查看

Scale Method NIQE↓ MUSIQ↑ MANIQA↑ CLIPIQA↑

4× 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

Ours 4.45 66.4 0.645 0.722

16× HiT-SR [93] 13.12 17.8 0.306 0.301

MambaIR [20] 13.29 18.5 0.309 0.319

SwinIR [40] 6.78 27.1 0.378 0.375

SeeSR [83] 6.46 50.4 0.512 0.509

OSEDiff [82] 5.57 53.5 0.543 0.597

CoZ [31] 6.08 58.8 0.602 0.656

Ours 5.63 61.9 0.601 0.735

64× HiT-SR [93] 16.44 21.4 0.362 0.394

MambaIR [20] 17.01 21.2 0.362 0.403

SwinIR [40] 8.06 22.5 0.476 0.459

SeeSR [83] 8.40 42.5 0.500 0.511

OSEDiff [82] 6.72 48.4 0.542 0.571

CoZ [31] 7.51 52.8 0.583 0.629

Ours 6.85 57.0 0.598 0.730

256× HiT-SR [93] 17.50 26.1 0.417 0.479

MambaIR [20] 18.31 26.1 0.415 0.489

SwinIR [40] 9.05 28.4 0.510 0.454

SeeSR [83] 10.45 37.7 0.503 0.500

OSEDiff [82] 8.02 45.7 0.540 0.556

CoZ [31] 8.40 50.5 0.582 0.602

Ours 7.73 54.1 0.599 0.719

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× onward, except MANIQA at 16× by 0.001. Best is bold; second-best is underlined.

Figure 15: Qualitative results on Flickr2K across 4–256×. 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.

媒体内容 · 前往原文查看

Scale Method NIQE↓ MUSIQ↑ MANIQA↑ CLIPIQA↑

4× 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

Ours 5.95 63.5 0.602 0.687

16× HiT-SR [93] 13.95 19.5 0.320 0.360

MambaIR [20] 14.45 19.6 0.322 0.377

SwinIR [40] 7.20 33.3 0.422 0.439

SeeSR [83] 7.70 52.5 0.527 0.533

OSEDiff [82] 6.75 53.7 0.543 0.603

CoZ [31] 7.60 57.3 0.578 0.626

Ours 6.92 59.4 0.580 0.702

64× HiT-SR [93] 16.65 23.8 0.387 0.457

MambaIR [20] 17.30 23.5 0.387 0.470

SwinIR [40] 8.31 25.6 0.484 0.470

SeeSR [83] 10.34 44.2 0.499 0.513

OSEDiff [82] 7.64 47.0 0.527 0.557

CoZ [31] 8.66 50.0 0.560 0.585

Ours 7.86 54.0 0.575 0.708

256× HiT-SR [93] 17.58 26.2 0.429 0.479

MambaIR [20] 18.67 25.9 0.429 0.495

SwinIR [40] 10.78 28.4 0.497 0.454

SeeSR [83] 12.27 34.2 0.495 0.490

OSEDiff [82] 8.39 43.1 0.521 0.517

CoZ [31] 9.83 45.2 0.556 0.540

Ours 8.72 49.1 0.585 0.667

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×. 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×; colored boxes mark the region enlarged at the next step.