Super-telephoto imaging on mobile phones suffers from complex, device-dependent degradations that are poorly captured by conventional handcrafted pipelines. Collecting large-scale aligned training pairs across different phones and diverse scenes, however, is both costly and difficult to scale. We therefore study whether device-specific degradations can be learned from only a small number of physical captures and transferred for training a unified super-resolution (SR) model. Our key idea is to factor degradation learning into two learnable components: one captures the device-specific imaging characteristics under well-resolved conditions, while the other captures the additional information loss and artifacts that arise in super-telephoto imaging. We build a paired benchmark across three phone brands (Huawei, Apple, and vivo) by photographing the same content at two display sizes and zoom settings. The larger displayed image is captured at the phone's maximum optical zoom setting, while the smaller image is photographed directly at the phone's highest native zoom setting through the phone's native imaging pipeline. The benchmark contains 519 training and 60 held-out pairs. We use the benchmark to learn device-specific degradation, then apply it to external images to synthesize thousands of additional pairs to train our UniTele SR model. Relative to the pretrained Qwen and FLUX backbones before adaptation, the full UniTele training recipe substantially improves paired restoration fidelity and perceptual similarity on the held-out benchmark and yields markedly more stable and visually faithful restorations.
Figures & tables
Figure 1: Overview of UniTele. Two learned mappings synthesize low-resolution (LR) images for SR training, using their corresponding synthetic high-resolution (HR) observations as targets.
Figure 2: Paired HR (left) and LR (right) observations of the same content across Huawei, iPhone, and vivo phones.
Backbone
Training setting
PSNR ↑
SSIM ↑
MAE ↓
LPIPS ↓
HF ratio →1
HP ↑
Input
No restoration
16.99
0.5234
27.10
0.5704
0.680
–
AdcSR ( Chen et al., 2025 )
Official pretrained model
16.90
0.5673
27.03
0.4383
0.698
2.623
OSEDiff ( Wu et al., 2024a )
Official pretrained model
16.92
0.5810
26.90
0.4617
0.641
–
Qwen SR
Base (untrained)
11.66
0.4931
54.16
0.6618
0.471
1.459
Real-ESRGAN degradations
16.90
0.5297
27.39
0.5288
0.740
2.262
Real-world only
18.84
0.6098
20.43
0.3187
0.740
5.180
Table 1: SR performance on 60 paired test images. Ratios denote real-to-synthetic sampling; dashes indicate unavailable HP scores.
Figure 3: Qualitative SR comparison on the three-phone paired test set. GT denotes the aligned real-world HR image.
Figure 4: Effect of the real-to-synthetic sampling ratio on UniTele (Qwen) and UniTele (FLUX).
Figure 5: Qualitative SR comparison on in-the-wild images. Red boxes mark the regions shown in the adjacent close-ups.
Method
MUSIQ ↑
CLIP-IQA ↑
NIQE ↓
Qwen SR base
44.33
0.2101
6.89
FLUX.2 SR base
65.60
0.5569
5.35
UniTele (Qwen), 1:2
54.49
0.2335
7.14
UniTele (FLUX), 1:2
58.27
0.3721
5.85
Qwen SR, Real-ESRGAN degradations
45.88
0.1767
8.03
GPT Image 2
67.83
0.5686
3.73
Table 2: No-reference quality on 16 in-the-wild iPhone images.
Figure 6: Additional in-the-wild SR comparisons. Rows show the input, Real-ESRGAN-trained Qwen, real-only Qwen, UniTele (Qwen), and GPT Image 2; red boxes mark the adjacent close-ups.
Device pool
Training data
PSNR ↑
SSIM ↑
MAE ↓
LPIPS ↓
HF ratio →1
All three devices
Real-world only
18.84
0.6098
20.43
0.3187
0.740
Vivo only
Real-world only
14.47
0.5451
38.99
0.4192
0.832
All three devices
UniTele (Qwen), 1:2
18.81
0.6031
19.87
0.3199
0.766
Vivo only
UniTele (Qwen), 1:2
13.51
0.5363
45.06
0.4617
0.815
Table 3: Effect of single- versus multi-device training on Qwen SR, evaluated across all three phones.
Forward-model variant
PSNR ↑
SSIM ↑
MAE ↓
LPIPS ↓
HF ratio →1
Without Stage I: direct digital input
12.05
0.4657
54.39
0.6337
1.992
Without Stage III: no reference-based refinement
15.20
0.5473
36.05
0.4001
0.950
Full three-stage model
15.52
0.5439
35.01
0.3942
1.018
Table 4: Stage ablations for LR synthesis on 50 paired test images from three phones.
Figure 7: Stage I ablation for Huawei-targeted degradation synthesis. Columns show the digital input, iPhone LR, synthesis without Stage I, the full model, and Huawei LR; the bottom row shows yellow-box close-ups.
Digital zoom on smartphones relies on learning-based super-resolution (SR) models that operate on RAW sensor images, but obtaining sensor-specific training data is challenging due to the lack of ground-truth images. Synthetic data generation via ``unprocessing'' pipelines offers a potential solution by simulating the degradations that transform high-resolution (HR) images into their low-resolution (LR) counterparts. However, these pipelines can introduce domain gaps due to incomplete or unrealistic degradation modeling. In this paper, we demonstrate that principled and carefully designed degradation modeling can enhance SR performance in real-world conditions. Instead of relying on generic priors for camera blur and noise, we model device-specific degradations through calibration and unprocess publicly available rendered images into the RAW domain of different smartphones. Using these image pairs, we train a single-image RAW-to-RGB SR model and evaluate it on real data from a held-out device. Our experiments show that accurate degradation modeling leads to noticeable improvements, with our SR model outperforming baselines trained on large pools of arbitrarily chosen degradations.
Universal image restoration aims to recover clean images from arbitrary real-world degradations using a single inference model. Despite significant progress, existing all-in-one restoration networks do not scale to multiple degradations. As the number of degradations increases, training becomes unstable, models grow excessively large, and performance drops across both seen and unseen domains. In this work, we show that scaling universal restoration is fundamentally limited by interference across degradations during joint learning, leading to catastrophic task forgetting. To address this challenge, we introduce a unified inference pipeline with a multi-branch mixture-of-experts architecture that decomposes restoration knowledge across specialized task-adaptable experts. Our approach enables scalable learning (over sixteen degradations), adapts and generalizes robustly to unseen domains, and supports user-controllable restoration across degradations. Beyond achieving superior performance across benchmarks, this work establishes a new design paradigm for scalable and controllable universal image restoration.
Debabrata Mandal, Soumitri Chattopadhyay, Yujie Wang +2
University of North Carolina at Chapel Hill · University of California, San Diego
Recent real-world image super-resolution (SR) methods often adapt text-to-image (T2I) backbones with ControlNet-style branches or spatial conditioning tokens, which increases memory and computes with resolution and often constrains training to a fixed scale. We propose Fill2SR, which repurposes a masked-inpainting Diffusion Transformer for SR without extra spatial branches. Our Inpainting-Interface Evidence Adapter (IIEA) writes the low-quality (LQ) observation into the native masked-image slot under a full-image mask, turning inpainting into a reverse-degradation conditional rectified flow trained with LoRA-only tuning. We further introduce RCDT, an offline pipeline that distills degradation descriptors from unpaired real images and transfers them onto clean targets using frozen open-source models. Fill2SR supports mixed-resolution training up to QHD and yields stable performance across 512/1024/2048 outputs. On synthetic benchmarks, our base model with IIEA achieves the best LPIPS on DIV2K and LSDIR; adding RCDT trades a small LPIPS drop for consistently stronger no-reference quality on RealLQ250 and RealPhoto60. Fill2SR remains memory-predictable, running 15362 inference on a single 32GB GPU and extending to multi-megapixel outputs via tiled restoration.
Xingfu Yi, Xiaoxue Yu
Independent Researcher, Hangzhou, China · Zhejiang University, Hangzhou, China