Source-linked AI summary
Training-Free Model Ensemble for Single-Image Super-Resolution via Strong-Branch Compensation
Gengjia Chang, Xining Ge, Weijun Yuan, Zhan Li, Qiurong Song, Luen Zhu, Shuhong Liu
TL;DR
The paper addresses how to combine complementary pretrained super-resolution models without additional training or architectural redesign. It uses asymmetric output-level compensation between HAT + TLC and MambaIRv2 with self-ensemble, improving over the base branch and slightly exceeding the strong branch in PSNR at the best operating point. The approach is presented as a practical upgrade path, with added inference latency and evaluation limited to unified DIV2K bicubic ×4.
Problem
The central gap is how to combine complementary pretrained super-resolution models effectively when additional training, data, or architectural scaling is undesirable.
Method
The method independently runs HAT + TLC as a stable base branch and MambaIRv2 with geometric self-ensemble as a detail-oriented compensation branch, then fuses outputs with a weighted image-space combination.
Results
The compensated result improves over the base branch and slightly surpasses the strong branch alone in PSNR at the best operating point, while the challenge submission achieves 33.47 dB PSNR and 0.9105 SSIM.
Takeaways & Limitations
Output-level compensation provides a low-overhead, deployment-friendly upgrade path for existing super-resolution systems without retraining or an additional fusion module.
Takeaways & Limitations
The method increases inference latency because both branches run at test time, and the current evaluation is limited to a unified DIV2K bicubic ×4 setting.
Abstract
from arXiv · showhide
Single-image super-resolution has progressed from deep convolutional baselines to stronger Transformer and state-space architectures, yet the corresponding performance gains typically come with higher training cost, longer engineering iteration, and heavier deployment burden. In many practical settings, multiple pretrained models with partially complementary behaviors are already available, and the binding constraint is no longer architectural capacity but how effectively their outputs can be combined without additional training. Rather than pursuing further architectural redesign, this paper proposes a training-free output-level ensemble framework. A dual-branch pipeline is constructed in which a Hybrid attention network with TLC inference provides stable main reconstruction, while a MambaIRv2 branch with geometric self-ensemble supplies strong compensation for high-frequency detail recovery. The two branches process the same low-resolution input independently and are fused in the image space via a lightweight weighted combination, without updating any model parameters or introducing an additional trainable module. As our solution to the NTIRE 2026 Image Super-Resolution ($\times 4$) Challenge, the proposed design consistently improves over the base branch and slightly exceeds the pure strong branch in PSNR at the best operating point under a unified DIV2K bicubic $\times 4$ evaluation protocol. Ablation studies confirm that output-level compensation provides a low-overhead and practically accessible upgrade path for existing super-resolution systems.
1. Introduction
The paper argues that combining complementary pretrained super-resolution models can improve reconstruction without additional training. It proposes asymmetric output-level compensation and reports gains on unified DIV2K ×4 evaluation and the NTIRE 2026 challenge.
- Existing pretrained models may have complementary strengths, making output combination without retraining a practical alternative to further architectural scaling.
- The proposed framework uses HAT + TLC for stable reconstruction and MambaIRv2 with geometric self-ensemble for high-frequency detail compensation.
- Output-level fusion uses a lightweight weighted combination without updating model parameters or adding a trainable module.
- Under unified DIV2K bicubic ×4 evaluation on 200 images, even a small strong-branch contribution produced measurable gains.
- 33.47 dB PSNR and 0.9105 SSIM placed the team second in the NTIRE 2026 Image Super-Resolution ×4 Challenge.
2. Related Work
Related work motivates combining pretrained branches because no single architectural family dominates every quality-efficiency tradeoff. The paper positions its approach among inference-time enhancement, self-ensemble, and training-free restoration aggregation methods.
- Transformer, attention, and state-space models have expanded the quality-efficiency design space for super-resolution.
- No single architectural family dominates every quality-efficiency tradeoff, supporting the use of complementary pretrained branches.
- Efficient SR research emphasizes sparse inference, compact designs, adaptive modulation, and lightweight attention mechanisms.
- TLC and geometric self-ensemble improve restoration at inference time without retraining, but do not directly specify how heterogeneous branches should be combined.
- Unlike symmetric model averaging, the proposed approach uses a designated base branch and introduces a strong branch as output-level compensation.
3. Method
The method independently runs a stable HAT + TLC base branch and a detail-oriented MambaIRv2 self-ensemble branch, then combines their outputs with a scalar compensation weight. This asymmetric design preserves pretrained models and exposes the fusion tradeoff through an interpretable control variable.
- 3.1. Problem Formation: The problem assumes two pretrained branches with complementary behavior: one stable on main structure and one stronger on difficult details.
- 3.2. Overall Framework: HAT + TLC generates the main reconstruction, while MambaIRv2 with geometric self-ensemble supplies stronger detail recovery.
- 3.2. Overall Framework: The final image is formed by a lightweight image-space combination of the branch outputs.
- 3.2. Overall Framework: α ∈ [0, 1] controls the weight assigned to the strong branch.
- 3.5. Why Output-Level Compensation: The asymmetric formulation keeps pretrained branches unchanged and makes the strong branch's contribution directly observable through α.
- 3.6. Compensation Weight Analysis: Figure 2 evaluates PSNR and SSIM as the strong-branch weight varies.
4. Experiments
Experiments evaluate the proposed compensation under a unified DIV2K bicubic ×4 protocol, combining quantitative comparisons, weight sweeps, and qualitative analysis. The compensated output improves the base branch, slightly exceeds the strong branch in PSNR, and preserves structure while enhancing local detail, with added inference cost and scope limited to the evaluation setting.
- Experimental Setup: The evaluation uses 200 DIV2K images with bicubic ×4 degradation, reporting PSNR and SSIM for HAT + TLC, MambaIRv2 with self-ensemble, and the compensated output.No retraining, fine-tuning, or learnable fusion module is introduced.
- Main Quantitative Results: 30.3527 dB PSNR / 0.880438 SSIM is achieved by the compensated result at Hybrid = 0.11 and Mamba = 0.89, versus 29.1696 dB / 0.854802 for the base branch and 30.3451 dB / 0.880466 for the strong branch.Relative to the base branch, this corresponds to +1.1831 dB PSNR and +0.025636 SSIM.
- Weight Analysis: Small strong-branch contributions already improve over the pure base configuration, while the best PSNR occurs at MambaIRv2 weight 0.89 and the best SSIM region occurs around 0.92.The curves indicate useful compensation across a broad weight interval rather than only at a narrow optimum.
- Qualitative Results: Qualitative examples show that compensation preserves the base branch’s global structure while improving repeated patterns, local edges, and texture continuity.The result is described as detail enhancement rather than indiscriminate sharpening.
- Discussion and Limitations: The dual-branch design increases inference latency, and conclusions are restricted to the unified DIV2K bicubic ×4 evaluation setting.The current method uses a fixed scalar weight; more adaptive compensation is left for future work.
5. Conclusion
The paper presents a training-free output-level ensemble that combines HAT as the base branch with MambaIRv2 as a strong compensation path. Under unified DIV2K bicubic ×4 evaluation, it improves over the base branch and slightly outperforms the strong branch alone in PSNR, supporting output-level compensation as a practical upgrade path.
- Conclusion: The proposed framework combines HAT with TLC as the base branch and MambaIRv2 with geometric self-ensemble as the strong compensation path without retraining or an additional fusion module.The method targets existing super-resolution systems.
- Conclusion: Under the unified DIV2K bicubic ×4 protocol, the compensated result clearly improves over the base branch and slightly outperforms the strong branch alone in PSNR at the best operating point.The reported conclusion concerns the specified evaluation protocol.
- Conclusion: Output-level compensation is presented as a practical and deployment-friendly upgrade path for existing super-resolution systems.This conclusion follows from combining pretrained branches without parameter updates or a learned fusion module.