acceptodds
Withdrawn from ICLR 2027

UniTele: Learning Device-Conditioned Degradations for Unified Smartphone Telephoto Super-Resolution

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

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.

Primary area:
datasets and benchmarks
Keywords:
Real-World Super-Resolution, Degradation Modeling, Super-Telephoto Imaging

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Ri9COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces UniTele, a smartphone telephoto super-resolution (SR) framework, which decomposes device-specific degradation into accurately characterized camera rendering effects and supplementary telephoto degradation components. Evaluation uses 60 controlled test pairs and 16 unpaired in-the-wild iPhone images.

Critical strengths
  1. Practical decomposition strategy and data acquisition. Disentangling camera rendering from telephoto signal loss follows an intuitive design. This approach enables a small number of physically captured samples to be applied to external image content.
  2. When trained with mixed real and synthetic data, Qwen achieves a 0.49 dB gain compared to training solely on real data. FLUX also receives performance benefits, which demonstrates that the synthesized image pairs provide meaningful supervisory signals.
Critical weaknesses
  1. The main test set has just 60 monitor-captured pairs following the same controlled protocol as the training data. The in-the-wild experiment uses only 16 unpaired iPhone images. Additionally, UniTele-FLUX is inferior to its base model on all three no-reference metrics. Hence, generalization to practical telephoto photography cannot be reliably confirmed.

  2. Estimating degradations from real pairs to generate SR training data is a known approach. The paper lacks baselines with learned degradation and dedicated per-device models. Tab. 3 evaluates a Vivo-trained model on all three devices, which inherently handicaps the single-device model. Tab. 4 measures synthesis quality but does not prove that factorization improves downstream SR performance compared to a capacity-matched one-stage generator.

Questions for authors
  1. Can the authors evaluate on paired outdoor telephoto scenes from all three devices and report results across multiple seeds or confidence intervals??

  2. How does UniTele compare with learned-degradation baselines under the same Qwen/FLUX backbone, data volume, and training budget??

  3. Can the authors compare the unified model against three per-device specialists on their corresponding devices and provide a downstream SR ablation for a capacity-matched one-stage degradation model?

Reviewer_CDToOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UniTele, a framework for learning device-specific smartphone telephoto degradations. To address the scarce paired data, it employs generative models to learn the mapping between device-domain HR observation and corresponding degraded LR images. Combining small size of real data with such synthetic data, the SR model (finetuned Qwen and FLUX) achieves favorable performance. Evaluation includes a three-phone paired benchmark, natural-scene images, and component ablations, providing some evidence for the effectiveness of UniTele.

Critical strengths
  1. A practically motivated data synthesis strategy. Using captured HR observations as an intermediate domain gives the factorization a supervision signal and makes the synthetic SR targets consistent with the captured targets.
  2. Relative to real-only training, UniTele improves both Qwen and FLUX on PSNR and LPIPS. These results exhibit potential value in the complete training pipeline.
Critical weaknesses

The contributions of UniTele is insufficiently established.

The related-work discussion omits recent generative degradation approaches such as RealDGen and GenDeg, leaving the methodological distinction insufficiently developed. UniTele has not been shown to provide a sufficiently distinct and well-supported contribution.

For data capture, there is a gap between the “device-specific” degradation learning and the practical data capture protocol. Although the authors have acknowledged that the learned mappings may capture display–camera interactions as well as device-specific characteristics, I am still concerned that this could weaken the main contribution of UniTele.

During the evaluation on in-the-wild images, UniTele (FLUX) undergoes a decline in performance compared with the FLUX.2 SR base model. Removing Stage I appears to feed digital images directly into a generator trained on device-domain HR observations, inducing a domain gap and does not establish that factorization outperforms a properly trained single mapping. Likewise, removing Stage III removes both reference conditioning and additional optimization, leaving their effects unresolved.

Questions for authors
  1. Please clarify the implementation details of w/o Stage I and w/o Stage III ablation.
  2. Please provide an analysis of the deterioration of UniTele FLUX in in-the-wild image evaluation.
  3. Is there any evidence that can distinguish the transferable device imaging features from their influence under specific display shooting conditions?
Reviewer_rZyUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper builds an SR dataset specific to images captured by the smartphone cameras. This dataset consists of low-resolution (LR) and high-resolution (HR) pair images that captured screen-displayed images using different zoom settings with different smartphone models. In addition to this, they propose a method to learn the complex degradation model from the limited amount of LR HR pair data using LoRA on Qwen-Image-Edit-2511 model. They claim that their dataset plus the synthetic data generated from the learned degradation model makes existing SR model work better on those specific datasets.

Critical strengths
  1. They built a real-world SR dataset from recent smartphone models, which is a good direction towards the practicality of image super-resolution on actual applications.

  2. They repurpose a foundational image editing model for modeling (/generating) the different degradations from different image capturing settings.

Critical weaknesses
  1. The dataset that the authors have collected shows highly noticeable pink-ish color shift and pixelated artifacts. This is natural since the authors constructed their dataset by capturing a monitor screen that displayed their selected images. In addition to this, constructing dataset from a flat screen has its limitation on being a good dataset for real-world usage, because the distortions that are presented according to the depth of the object cannot be modeled. These problems limits the goodness of the dataset that the authors have created. The authors need to take these problems into account when creating the dataset and try to minimize the downsides, and most importantly, mention about this problem on their paper.

  2. The design of their method of generating the synthetic HR and LR pairs needs to be given a rationale. The authors have proposed a cascade model as x -> h_c -> y_c (I will denote \hat{h}_c as h_c and \hat{y}_c as y_c for simplicity). However, there are more design choices as (x -> h_c ; x -> y_c) or (x -> h_c ; x, h_c -> y_c) etc. Since the captured LR and HR are two separate captures, thus having separate stochasticity, having h_c -> y_c is not fully convincing and other designs are not tested.

  3. The SR model trained on this paper's captured/synthetic HR, LR pair dataset will not produce photorealistic high quality output images. We can see in Figure 3. that the "GT"s that they are targeting are too distorted and Flux base actually gives a more photorealistic output. This is showing us that the artifact that comes from the digital monitor screen capturing is too overwhelming that it is questionable whether the degradation that comes from different smartphone devices is present within the dataset.

Questions for authors
  1. What is your rationale behind choosing the x -> h_c -> y_c design over other designs?
  2. How do you separate the effect of the artifact that comes from the digital monitor screen capturing from the smartphone degradation that you wish to model?
  3. What difference did you expect to have from different devices? What is the difference of ISP algorithm of those different devices that would produce different degradation artifact? What were some statistical properties that made the images from different devices distinctive? Do they differ in power spectra density or any color statistics?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.