TL;DR
RareFlow is a physics-aware super-resolution framework for remote sensing that enhances cross-sensor imagery, ensuring physical accuracy, robustness to out-of-distribution data, and uncertainty quantification, leading to high-fidelity, reliable results.
Contribution
It introduces a dual-conditioning architecture with physics-based loss functions and uncertainty estimation for robust, physically consistent super-resolution of multi-sensor satellite imagery under domain shifts.
Findings
Significantly outperforms state-of-the-art baselines in qualitative evaluations.
Achieves nearly 40% reduction in FID scores.
Provides reliable uncertainty estimates to identify unfamiliar inputs.
Abstract
Super-resolution (SR) for remote sensing imagery often fails under out-of-distribution (OOD) conditions, such as rare geomorphic features captured by diverse sensors, producing visually plausible but physically inaccurate results. We present RareFlow, a physics-aware SR framework designed for OOD robustness. RareFlow's core is a dual-conditioning architecture. A Gated ControlNet preserves fine-grained geometric fidelity from the low-resolution input, while textual prompts provide semantic guidance for synthesizing complex features. To ensure physically sound outputs, we introduce a multifaceted loss function that enforces both spectral and radiometric consistency with sensor properties. Furthermore, the framework quantifies its own predictive uncertainty by employing a stochastic forward pass approach; the resulting output variance directly identifies unfamiliar inputs, mitigating…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
1) They designed losses tied to sensor spectra which helps keeping outputs physically plausible, not just visually sharp. This also ensure that the large-scale radiometric information are preserved and the hallucination is prevented. 2) A gated ControlNet (for structure) plus text prompts (for semantics) jointly handle appearance and geometry under distribution shift. 3) They got a strong performance on benchmark compared to SOTA methods.
1) How sensitive are the results to prompt wordings? Can you do some ablation on this? 2) Do the physics losses transfer to unseen sensors and bands without retraining? Can you provide any zero-shot evaluation? 3) Do the authors have any intuition why PSNR is low comparatively compared to other metrics in Table 3? 4) Do the author have some ablations on different parts of the losses. Which components (spectral vs. radiometric) drive more gains? Provide loss term ablations.
The paper presents a well-motivated and technically innovative framework that effectively addresses the issue of generating physically inconsistent details in out-of-distribution scientific imagery. RareFlow’s dual-conditioning architecture successfully balances structural fidelity and semantic guidance, while its uncertainty-gated control mechanism adaptively suppresses hallucinated features under high uncertainty. The introduction of a physics-aware loss formulation, which integrates spectral
1.While the paper repeatedly emphasizes its “physics-aware”, the methodology relies primarily on heuristic loss formulations rather than an explicit physical modeling framework. No clear theoretical connection is established between the proposed losses and physics. As a result, the term “physics-aware” feels more empirical than principled, potentially overstating the scientific rigor of the approach. 2.Although most backbone components are frozen, the overall training pipeline still feels fragme
- The paper establishes a clear knowledge gap and highlights issues that are important to make super-resolution approaches usable in real-life satellite-image applications. Class imbalance/ lack of examples of rare phenomena are a pervasive problem that is important to address. - The contributions are relevant, as the issues highlighted in the paper are specific to satellite images and are not necessarily a problem in the natural image domain (specifically the importance of preserving the ‘styl
- Motivation/clarity can be improved: key methodological concepts (uncertainty quantification, physics-aware components, prompts, dual-conditioning, ControlNet) are introduced too late or not sufficiently motivated. ‘Uncertainty quantification’ is mentioned in the abstract, then the next occurrence is in section 3.2. It would be good to explain it also in the introduction and start of methods section. The title mentions ‘physics-aware’, but this also does not appear in the introduction. Promp
The paper proposes a pipeline for training a ContolNet for super resolution. One contribution is to modulate the control adapters using uncertainty. The uncertainty is computed using Monte Carlo Dropout. The remaining of the training pipeline is rather classical with a combined loss that assemble various desired properties (spectral faithfulness, color consistency, LPIPS, see Eq. (12)). The reported experimental results are very good. The method coined RareFlow achieves best performances on f
The main weakness of this paper is the lack of novelty in its methodology for adapting a computer vision tool for a scientific application. The use of GPT5 for the data expertise is also debatable as a scientific expertise. All in all, even though this is an interesting problem, this may be of limited interest for the ICLR community. Another main limitation is that reproducibility is not discussed. Performance in terms of computation times are not discussed. The computation of the uncertainty
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
