acceptodds
Under review as a conference paper at ICLR 2027

MISR-GS: Test-Time Multi-Image Super-Resolution with Explicit 2D Gaussian Fields

Abstract

Multi-image super-resolution reconstructs a high-resolution image from multiple low-resolution observations captured with sub-pixel misalignments. Recent test-time optimization methods address this problem by representing the unknown high-resolution signal as a shared implicit neural field and jointly optimizing frame-wise alignment. However, coordinate-based implicit representations rely on global MLPs and frequency-based positional encodings, which may limit local capacity allocation and introduce domain-sensitive hyperparameters. We propose MISR-GS, a test-time multi-image super-resolution framework based on explicit 2D Gaussian fields. MISR-GS represents the unknown high-resolution image as a set of learnable anisotropic 2D Gaussian primitives shared across all low-resolution frames. Each frame is reconstructed through a differentiable render-and-degrade process, where the shared Gaussian field is rendered at the target resolution, warped by a frame-specific affine transformation, photometrically corrected, and downsampled to match the corresponding low-resolution observation. To adapt Gaussian representations to this inverse problem, we introduce observation-driven initialization and multi-frame residual-guided densification, both derived solely from low-resolution inputs. Experiments on synthetic and real burst datasets demonstrate the effectiveness, efficiency, and robustness of MISR-GS for test-time multi-image super-resolution.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-image Super-resolution

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_n7BEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

MISR-GS represents the latent high-resolution image as a shared explicit field of learnable anisotropic 2D Gaussian primitives. MISR-GS derives both initialization and progressive capacity allocation solely from the low-resolution burst. A pseudo-HR reference is used only to initialize Gaussian positions and colors, while multi-frame reprojection residuals guide subsequent densification. The resulting field, frame-wise affine alignment, and photometric parameters are jointly optimized through a differentiable render-and-degrade model in the low-resolution observation space, with an LR-only ringing-aware regularizer stabilizing weakly constrained high-frequency solutions. MISR-GS improves PSNR over the strongest evaluated optimization-based TTO baseline while requiring no paired HR training data.

Critical strengths

An explicit field of learnable anisotropic 2D Gaussian primitives is introduced into the multi-image super-resolution task to achieve the state-of-the-art performance.

Critical weaknesses

The novelty of proposed method is insufficient. The observation-driven Gaussian allocation strategy and LR-constrained optimization framework are not the core novelties. The core contributions should be clarified and refined.

Some recent multi-image super-resolution methods should be discussed and compared. Such as [1] Superf: Neural implicit fields for multi-image super-resolution. ICLR2026 [2] A graph attention network for real-world multi-image super-resolution. Information Fusion 2025 [3] HyDA-Net: a hybrid dense attention network for remote sensing multi-image super-resolution. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. [4] Post-Processing for Bridging the Simulated-to-Real Domain Gap in Multi-Image Super-Resolution. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing.

In quantitative comparison, some representative works should be compared in Table 1 for compensative comparisons rather than some outdated methods.

The paper writing and presentation should be improved.

Some images are displayed incompletely. Such as Figure 3 (a) (c), Figure 4. In Figure 3(b), the existing and new visuals are difficult to distinguish due to their similar colors.

Questions for authors

Please see Critical Weaknesses.

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

The paper proposes MISR-GS, a test-time optimization (TTO) framework for multi-image super-resolution (MISR) that represents the latent high-resolution (HR) image as an explicit field of learnable anisotropic 2D Gaussian primitives. Unlike recent INR-based TTO methods that rely on implicit neural fields, MISR-GS constructs the Gaussian field entirely from low-resolution (LR) burst observations: a pseudo-HR reference derived from the LR burst initializes Gaussian positions and colors, while multi-frame LR reprojection residuals guide progressive densification. The field, frame-wise affine transformations, and photometric parameters are jointly optimized in the LR observation space through a differentiable render-and-degrade model, supplemented by an uncertainty-aware reprojection loss and a ringing-aware regularizer. Experiments on SatSynthBurst and SyntheticBurst at ×2, ×4, and ×8 show consistent PSNR improvements of 0.98–2.97 dB over the strongest TTO baseline (SuperF), along with a computational cost analysis demonstrating favorable efficiency at higher upsampling factors.

Critical strengths
  1. Novel representation for TTO MISR.​ The paper is the first to explore explicit 2D Gaussian fields as an alternative to implicit neural representations for LR-supervised test-time MISR. This is a meaningful and timely investigation given the growing interest in Gaussian-based image modeling.
  2. Well-designed LR-only pipeline.​ The framework avoids HR supervision entirely. Gaussian initialization from an LR-derived pseudo-HR reference and residual-guided densification from multi-frame LR consistency are principled and practically appealing.
  3. Technically solid components.​ The differentiable render-and-degrade optimization, uncertainty-weighted reprojection, and ringing-aware regularization are well motivated and complement each other. The ablations show that each contributes positively.
  4. Comprehensive experimental study for a TTO paper.​ The paper reports results across multiple scales, includes robustness tests under large misalignment, evaluates computational cost, and provides ablation studies of the objective and allocation strategies.
  5. Clear writing and structured presentation.​ The method is logically organized, and the relation to prior work—particularly the distinction from Image-GS and Gaussian-based SR methods—is explained clearly.
Critical weaknesses
  1. Incremental novelty relative to existing Gaussian-based SR methods.​ The representation itself (2D Gaussian primitives, rotation-scaling covariance decomposition, normalized aggregation, rasterization backend) is adopted from prior works such as Image-GS and Fast2DGS. The main contribution is thus the adaptation to LR-supervised TTO MISR and the observation-driven allocation strategy. The authors should more explicitly articulate what is fundamentally new beyond the application setting, and preferably include a comparison against a direct adaptation of an existing Gaussian SR method to MISR.
  2. Unfair or incompletely justified comparison with SuperF.​ The computational cost table shows that SuperF runs for exactly 2,000 iterations, whereas MISR-GS uses approximately 4,000 iterations. If SuperF’s iteration budget is not chosen to reach convergence, the reported PSNR gains may partially reflect an unfair optimization budget. The authors should either run SuperF to convergence with early stopping or provide sensitivity results with respect to the iteration budget.
  3. Potential bias from the pseudo-HR initialization.​ The pseudo-HR reference is built by aligning and median-fusing all LR frames. This is a strong initialization that already fuses the burst. The subsequent Gaussian optimization may merely refine this initial estimate. The ablated allocation strategies still rely on the same pseudo-HR reference for initialization. An ablation starting from simple bicubic upsampling of a single frame or from uniformly random positions is needed to isolate the contribution of the Gaussian field itself and of the proposed densification strategy.
  4. Limited scale of evaluation.​ SatSynthBurst is constructed from only 20 HR images, and SyntheticBurst is evaluated on a 50-sample subset. This raises concerns about statistical robustness and generalization. No significance tests (e.g., paired t-tests or confidence intervals) are reported. The authors should consider evaluation on more samples or at least report per-sample variance and significance.
  5. Missing comparison with learning-based MISR methods.​ While the paper is positioned as a TTO approach, state-of-the-art burst SR methods such as BSRT and Burstormer achieve much higher PSNR on SyntheticBurst. A comparison using publicly available pretrained models would help the reader understand the performance gap between test-time optimization and learned burst fusion, and would contextualize the practical value of the proposed method.
  6. Heuristic nature of the ringing-aware regularizer.​ The gating function in Eq. (10) is intuitive, but it is based on local sign consistency of first-order differences. The choice of α = 6 and λ_ring = 0.08 is not justified, and no sensitivity analysis is provided. The authors should include an analysis showing how the regularizer behaves across different α/λ values and why the selected configuration is optimal.
  7. Computational cost results are not fully analyzed.​ At ×2, MISR-GS is slower (21.47s vs. 12.64s) and uses more memory (652 MB vs. 188 MB) than SuperF. At larger scales, the trend reverses. The authors should explain this behavior. Specifically, why does the per-iteration time of MISR-GS decrease with the upsampling factor (from 5.28 ms at ×2 to 4.54 ms at ×16) in Table 5? This is counterintuitive and needs a clear explanation.
  8. Incomplete presentation of results and notation.​ Table 3 appears incomplete in the main text, and some equations (e.g., the rotation matrix decomposition around Eq. (3)) are rendered incorrectly. The reference list contains duplicates (e.g., GaussianSR appears twice). Additionally, the appendix reports ×16 results in Table 5 while the main text only describes ×2/×4/×8, which should be reconciled.
Questions for authors
  1. The comparison with SuperF uses different iteration budgets. How was SuperF’s 2,000-iteration limit chosen? Would SuperF’s performance improve if it were optimized for more iterations or with the same early-stopping criterion? Have you tested this?
  2. The pseudo-HR reference is a strong initialization that already fuses the LR burst. Have you evaluated the contribution of the Gaussian field itself by initializing from a simpler baseline, such as bicubic upsampling of a single reference frame or random spatial centers? If so, what were the results?
  3. Beyond the comparison with SuperF, have you considered other INR-based TTO baselines, such as Neural Image Representations (Nam et al.) or a re-implemented INR with comparable capacity? Would a fair budget comparison change your claim that explicit Gaussian fields are superior for LR-supervised TTO MISR?
  4. Could you provide a more formal analysis of the ringing-aware regularizer? Why does the product d_x⁻ · d_x⁺ reliably distinguish ringing from legitimate edges? Have you performed a sensitivity study over α and λ_ring?
  5. The Gaussian budget of 50,000 is fixed across datasets and scales. How was this number chosen? What is the impact of increasing or decreasing the budget on PSNR, optimization time, and memory usage?
  6. At ×2, MISR-GS has higher total runtime and memory than SuperF, while at larger scales it becomes more efficient. Could you explain the source of this scaling behavior and whether it reflects the fixed Gaussian count relative to the HR image resolution?
  7. In the allocation ablation (Table 3), how exactly is “pseudo-HR error guidance” computed? If it uses the error between the pseudo-HR reference and the rendered HR image, does that create a circular dependency on the pseudo-HR reference? How does this affect the claim that allocation is driven purely by LR observations?
  8. The uncertainty map is regularized with an ℓ2 penalty and a smoothness penalty. What is the sensitivity of the final reconstruction to these weights? Is the uncertainty map visually interpretable, and does it correlate with known noise or occlusion regions?
  9. What happens when the LR burst has limited sub-pixel diversity (e.g., all frames nearly aligned) or strong aliasing? Is there a minimum number of frames or a minimum sub-pixel shift distribution required for the method to be effective?
  10. In the large-misalignment robustness test (Table 4), the performance drop is smaller for MISR-GS than for SuperF, particularly at ×8. Can you provide an analysis of why the Gaussian representation is more robust to large initial misalignment? Does the explicit field admit a different optimization landscape than the INR?
Reviewer_Q5XEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MISR-GS is proposed, which is a TTO method for MISR representing the HR image as a shared field of learnable anisotropic 2D gaussians. Gaussian params, frame alignments, and photometric corrections are jointly optimized by rendering and degrading the reconstruction to match the observed LR frames. Gaussians are initialized using an aligned, fused estimate derived from the input burst, while subsequent is guided by multi-frame reprojection residuals. Reported experiments show improved reconstruction quality over the baselines and computational advantages at larger upsampling factors.

Critical strengths

First strength: At a fixed Gaussian budget, the residual-guided allocation improves the PSNR by approx. 0.85 dB on both datasets, supporting the value for LR-only reconstruction. Second strength: They have a promising computational efficiency. At 8x, the method reports approx. 2.7x faster optimization and a third of the memory usage, offering a nice practical benefit at larger scales.

Critical weaknesses

First weakness: Unexplained SuperF baseline gap? The reported SuperF scores are substantially lower than reported in the original paper. Key settings such as the fourier scale is not reported. Second weakness: Kinda unclear source of improvement. The method changes representation, initialization, losses, and optimization all together. There is no matched INR comparison, the benefit specifically attributable to Gaussian representation remains uncertain.

Questions for authors
  1. Which SuperF implementation, loss, and fourier scales did you use? E.g. the reported 10 and 3 scales for each of the datasets. What explains the gap to the original paper?
  2. Can you provide a controlled comparison with an INR using the comparable losses, alignment initialization, and optimization budget to clarify the contribution of the Gaussian representation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.